JP5452563B2 - Method and apparatus for extracting evaluation information - Google Patents

Description

  The present invention relates to the field of data mining, and more particularly to a method and apparatus for extracting evaluation information.

  With the development of the Internet, it is no longer satisfactory to simply passively receive information on the Internet, and more and more people want to express their opinions on the network or disclose personal information ing. These comments and opinions are generally published on various network media such as shopping websites, electronic bulletin boards, personal blogs, mini-blogs, among others, user comments on products, readers on events and policies. Many opinions are included. Content that includes such text as comments is analyzed and mined using opinion mining technology, and individuals and organizations understand the general opinions and attitudes of the public about products, events, and policies. It is useful for. Therefore, it can be inferred that opinion mining technology has extremely high application value.

  Opinion mining is mainly related to automatically acquiring useful evaluation information and related knowledge from subjective expression text (text such as words, phrases, simple sentences, sentences including sensory expressions). The main purpose of opinion mining is to identify evaluation information from text and thereby perform appropriate analysis. The method of acquiring the current evaluation information is roughly classified into the following three.

  The first method is a method of extracting evaluation information semi-automatically based on a coexistence template. In this method, the evaluation information is defined as a triple (object, attribute, evaluation), and each element of the evaluation information is regarded as a slot value of the coexistence template. And what matches three elements of evaluation information is extracted from a sensory expression text, and is analyzed by a coexistence template. For example, if the coexistence template is "<object> <attribute> is <evaluation>" and the sensory expression text to be analyzed is "this camera photo is great", then match this sentence with the coexistence template. The extracted triples are considered to be <this camera>, <photo>, and <great>, respectively. In this method, three dictionaries “object”, “attribute”, and “evaluation” are constructed, and the word that is an element in each dictionary is manually selected, and a plurality of coexistence templates having relatively high appearance frequencies are selected. There is a need. Therefore, “attributes” and “evaluations” are first generated, then the generated “attributes” and “evaluations” are manually filtered, and the appropriate “attributes” and “evaluations” after filtering are stored in the respective dictionaries. Placed in.

  The second method is a method for extracting evaluation information based on a collocation dictionary. The evaluation information is obtained by examining two sets of evaluation information (opinion word, target feature). This method requires the construction of three dictionaries. The first is an opinion word dictionary, the second is a target feature dictionary acquired by manual search, and the third is a link insertion dictionary that mainly describes syntax relationships. In this method, in order to obtain evaluation information, opinion words and target features included in the sensory expression text are marked using the opinion word dictionary and the target feature dictionary. Next, using the link insertion dictionary, the syntactic relationship between the opinion word and the target feature pair is referred to, and it is determined whether or not the opinion word and the target feature pair match correctly.

  The third method is a method of extracting evaluation information based on a grammatical path dictionary (Patent Document 1 (Chinese Patent Application No. 2000009100822.1) “Methods, Devices and Systems for Observing Evaluation Unit and Establishing Syntax Pax”. Dictionary (see Method, Apparatus, and System for Obtaining Evaluation Units and Building Syntax Path Dictionaries)). In this method, the relationship between product features and opinion words is described using a syntax path. First, you need to build a syntax path dictionary. Next, all product feature words and opinion words in the sensory representation text corpus are recognized, and a syntax path between all product features and opinion words is generated. These syntax paths are generalized and the frequency of occurrence of the generalized syntax path is calculated. Then, a syntax path whose appearance frequency exceeds a specific threshold is regarded as a reference syntax path, and is inserted into the syntax path dictionary. After obtaining the syntax path dictionary, the product features and opinion words of the input sensory expression text are recognized, a syntax analysis is performed, and a corresponding syntax tree is constructed. Based on the syntax path dictionary, querying the path where the syntax tree matches the reference syntax path allows product features and opinion words linked by this path to be considered as evaluation units (ie, evaluation information). .

Chinese Patent Application No. 200000913422.1

  The problem with the first method is that the type of coexistence template is simple and covers only a narrow range, so the call rate is low, and the selection procedure requires manual filtering.

  The problem with the second method is that the call rate is low because some types of syntactic relations are excluded from the target, the dictionary needs to be built manually, and the portability of the dictionary is not desirable. It is not.

  The problems with the third method are that it depends on a large number of resources such as dictionaries and syntax analyzers, that the system is highly complex, that the call rate is low, and that some of the generated syntax path dictionaries This is necessarily a drop in extensibility because the syntactic relationship may be excluded.

  In view of the above-mentioned problems, the present invention provides a method for extracting evaluation information, which is not as complex as before, has a low dependency on a dictionary, and is inherently efficient.

  According to the first aspect of the present invention, the step of acquiring the opinion word set and the target feature set from the corpus, the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the target feature set A step of optimizing the opinion word set and the target feature set based on the similarity of the elements, and a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set An evaluation information extraction method is provided.

  According to the second aspect of the present invention, the acquisition means configured to acquire the opinion word set and the target feature set from the corpus, the relationship between the opinion word set and the target feature set, and the elements of the opinion word set Optimization means configured to optimize the opinion word set and the target feature set based on the similarity and the similarity of the elements of the target feature set, and the optimized opinion word set and the optimized target There is provided an evaluation information extraction device comprising extraction means configured to extract evaluation information based on a feature set.

  According to the present invention, a machine learning method is provided that is less expensive and does not require management as compared with the conventional method. In the method according to the present invention, the dependence on the dictionary is low, and it is not necessary to use a plurality of dictionaries as in the prior art.

  Other features and advantages of the present invention will become apparent upon reference to the following description of the preferred embodiment and the drawings which illustrate the principles of the invention.

  The following description with reference to the drawings will provide a more comprehensive understanding of the present invention, so that other objects and advantages of the present invention will become clearer and easier to understand.

4 is a flowchart of an evaluation information extraction method according to an embodiment of the present invention. 6 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention. 6 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention. 6 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention. It is a block diagram of the evaluation information extraction apparatus by one Example of this invention.

  In all the above figures, the same reference numbers indicate the same, similar or corresponding features or functions.

  In the following, the invention will be described and illustrated in more detail with reference to the figures. It should be understood that the figures and examples of the present invention are used for illustrative purposes only and are not intended to limit the protection scope of the present invention.

For clarity, the terminology used in the present invention is first explained.
1. Corpus

In the present invention, the corpus may be composed of a plurality of text files. The text file may be preprocessed according to a predetermined processing granularity in order to obtain a plurality of text units. The text unit is a minimum language unit in which the target feature and the opinion word coexist. The text unit can be free text such as a word, phrase, sentence, paragraph, entire article, or a combination thereof.
2. Evaluation information

  In the present invention, the evaluation information can include a target feature and an opinion word.

  Opinion words are words or phrases that are included in a text unit and evaluate something to express an opinion. Opinion words include, for example, “good”, “high”, “beautiful”, “elegant”, “convenient”, and the like.

  The target feature is an evaluation target that is modified by an opinion word included in a text unit, and includes products and services. Examples of the target feature include “fuel consumption rate”, “appearance”, “price”, “safety”, “maneuverability”, and the like.

The evaluation information is a pair composed of opinion words included in text units and corresponding target features, and has a clear opinion polarity (positive, negative, neutral). The evaluation information may be a pair of target features and opinion words, that is, [target features, opinion words]. For example, from the three text units "High emission fuel consumption is very high", "The price of this model's mobile phone is high" and "This brand car has very good maneuverability" Three pieces of evaluation information are obtained: [Fuel consumption rate, high], [Price, high], and [Maneuverability, good].
3. Relationship between opinion word set and target feature set

  When an opinion word and a target feature are detected from the same text unit, the opinion word and the target feature are considered to be related. By searching for opinion words, target features, and the relationship between them based on a large number of text units obtained from the corpus, the appearance frequency of opinion words, the appearance frequency of target features, and the frequency of the relationship between them can be obtained. .

  In the present invention, the relationship between the individual opinion words included in the opinion word set and the individual target features included in the target feature set and the frequency of the relationship are referred to as the relationship between the opinion word set and the target feature set.

For example, “The fuel consumption rate of high-emission vehicles is extremely high”, “The price of this model's mobile phone is high”, “The driving performance of this brand car is very good”, “The price of the purchased printer is too high If there are four text units “high”, the opinion word set contains two opinion words “good” and “high”, the appearance frequency of “good” is 1, and the appearance of “high” The frequency is 3. The target feature set includes three target features “fuel consumption rate”, “price”, and “maneuverability”. The appearance frequency of “fuel consumption rate” is 1, and the appearance frequency of “price” is 2. , And the appearance frequency of “maneuverability” is 1. The opinion word “good” and the target feature “fuel consumption rate” do not appear in the same text unit (hereinafter referred to as “co-occurrence”). The relationship frequency is recorded as “0”. Similarly, the opinion word “good” is related to both of the target characteristics “price” and “maneuverability”, and the frequency of the relationship between the two and “good” is “1”. Similarly, for the opinion word “high”, it is possible to obtain the presence / absence of the relationship with the target characteristics “fuel consumption rate”, “price”, and “maneuverability” and the frequency of the relationship. The following example shows an opinion word set (indicated by “O”) that includes two opinion words of “good” and “high”, and three target features of “fuel consumption rate”, “price”, and “maneuverability”. A relationship with a target feature set (indicated by “F”) including

  In the present invention, the relationship may be expressed in a matrix format for easy calculation.

Based on the relationship between the opinion word set and the target feature set, a conversion relationship from the opinion word set to the target feature set ("shown as TO _-F ") can be obtained. Since the appearance frequency of “good” in the opinion word set is 1 and the appearance frequency of “high” is 3, the conversion relationship T _O-F from the opinion word set to the target feature set can be expressed as follows. .

  In the above table, “/” indicates a proportional relationship. For example, “1/3” in the column “O (“ high ”)” indicates that the co-occurrence frequency of the opinion word “high” and the target feature “fuel consumption rate” is “high” in all text units. “2/3” indicates that the co-occurrence frequency of the opinion word “high” and the target feature “price” is “high” in all text units. “0/3” indicates that the appearance frequency of the opinion word “high” is 3 in all text units, but the opinion word “high” and the target feature “maneuverability” Indicates that they do not co-occur.

Furthermore, based on the relationship between the opinion word set and the target feature set, a conversion relationship from the target feature set to the opinion word set (referred to as “T _F−O ”) can be obtained. Since the appearance frequency of “fuel consumption rate”, “price”, and “maneuverability” of the target feature set is 1, 2, and 1, respectively, the conversion relationship T _F-O from the target feature set to the opinion word set is Can be shown as:

  In the above table, “/” indicates a proportional relationship. For example, in the column “F“ price ””, “0/2” indicates that the target feature “price” and the opinion word “good” are present when the target feature “price” appears twice in all text units. ”Does not co-occur, that is, the appearance frequency of the target feature“ price ”in all text units is 2, whereas the target feature“ price ”and the opinion word“ good ”do not co-occur. It is shown that. “2/2” indicates that the co-occurrence frequency of the target feature “price” and the opinion word “high” is 2 when the target feature “price” appears twice in all text units. Show. The above reflects that, in a situation where the target feature “price” appears in the text unit, the possibility that the opinion word “high” appears is large and the possibility that the opinion word “good” appears is small.

In the present invention, in order to facilitate calculation, both the conversion relation T _O-F and the conversion relation T _F-O can be expressed in a matrix form.
4). Similarity of elements in the target feature set and similarity of elements in the opinion word set

The similarity of elements of the target feature set is a combination of similarities between all target features included in the target feature set. For example, the target feature set includes three target features, “fuel consumption rate”, “price”, and “maneuverability”. By calculating similarity, the similarity between “fuel consumption rate” and “price” If the result that the similarity between “fuel consumption rate” and “maneuverability” is 0.2 and the similarity between “price” and “maneuverability” is 0.01 is obtained, The similarity of the elements of the set can be shown as follows.

Similarly, the similarity of the elements of the opinion word set obtained by the above method is shown below.

  In the present invention, in order to facilitate the calculation, the similarity between the elements of both sets can be expressed in a matrix format.

  The similarity between any two elements in the set can be calculated based on various existing methods. For example, a method based on a semantic dictionary (semantic dictionary) may be used. In this way, the similarity between any two elements in the set can be calculated by examining the path length of two words in the semantic structure tree based on an external dictionary such as Synonym Dictionary or HowNet. . As another example, a machine-based learning method may be used. In this method, the similarity between any two elements in the set can be calculated based on algorithms such as principal component analysis, latent semantic analysis, context vector similarity.

  The present invention relates to a method for extracting evaluation information. This method is based on the step of obtaining the opinion word set and the target feature set from the corpus, the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. The method may include a step of optimizing the opinion word set and the target feature set, and a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set.

  The present invention adjusts the rank of the target feature set and opinion word set based on the principle that “information co-occurrence exists in the relationship and similarity between objects of the same category”, and the target is obtained by continuous iteration. Gradually optimize the feature set and opinion word set, extract high-ranked target features and opinion words from the final ranking results obtained, and use the relationship between the target features and opinion words. Establish evaluation information including target features and opinion words.

  The present invention provides a machine learning method that is low in cost and does not require management as compared with a conventional method. In this method, it is not necessary to use a plurality of dictionaries as in the prior art. On the contrary, the present invention can achieve the object without using a dictionary at all, and can be implemented only by using a degree adverb. Furthermore, the present invention allows dynamic filtering and expansion of the target feature set and opinion word set, and dynamic construction of the modification relationship between the target feature and opinion word, so that evaluation information can be quickly and effectively received from the corpus. Can be extracted.

  FIG. 1 is a flowchart of an evaluation information extraction method according to an embodiment of the present invention.

  In step S101, an opinion word set and a target feature set are acquired from the corpus.

  In this step, first, corpus preprocessing is performed to obtain a text unit, then, based on the obtained text unit, an opinion word set is obtained according to the opinion word extraction rule, and according to the target feature extraction rule. A target feature set may be acquired.

  As the corpus preprocessing, for example, text files in the corpus can be processed by simple sentence division, word division, part-of-speech tagging, traditional / simplified character conversion, and the like. In one embodiment, one sentence is set as a text unit, and the sentence is handled as a minimum language unit in which the target feature and the opinion word co-occur. For example, a text file in a text corpus is encoded for seven punctuation marks (punctuation “.”, Comma “,”, semicolon “;”, exclamation mark “!”, Question mark “?”, Punctuation mark “,”). By substituting and breaking the line, it is divided into simple sentences. Next, processing such as word division, part-of-speech tagging, and traditional / simplified character conversion is performed on the acquired sentence as necessary. Then, based on the preprocessed sentence (ie, text unit), the opinion word set is acquired according to the opinion word extraction rule, and the target feature set is acquired according to the target feature extraction rule.

  Opinion word extraction rules include, for example, a phrase that immediately follows a degree adverb, an adjective, a phrase that does not include a function word, a phrase that is shorter or equal in length to the longest opinion word, and an opinion word that has the lowest frequency of appearance A method may be used in which one or more phrases having high frequency are extracted from the text unit as opinion words.

  In one embodiment, all possible phrases of opinion words are detected by searching individual text units obtained from the corpus according to opinion word extraction rules. For example, if there is a text unit that “the fuel consumption rate of high-emission vehicles is very high”, and the opinion word extraction rule is “extract from the text unit the phrase immediately following the degree adverb as an opinion word” Since the text unit has only one degree adverb “Taien”, “High” immediately after “Taien” is extracted as an opinion word. In this way, a set of opinion word candidates is acquired by performing the above processing for each text unit acquired from the corpus.

  The target feature extraction rules are, for example, “a basic noun phrase, a combination of basic noun phrases, a combination of a basic noun phrase and a noun / verbal noun, a combination of a basic noun phrase and a noun, and a combination of a noun phrase and a noun / noun. , Extract one or more phrases from the text unit that contain no functional words, phrases that are shorter or equal to the longest target feature, and phrases that appear more frequently than the target feature with the lowest occurrence frequency. May be used.

  In one embodiment, by searching individual text units taken from the corpus according to opinion word extraction rules, all possible phrases of opinion words are detected, and then the detected opinion words are Among the existing text units, the target feature is detected from the context according to the target feature extraction rule. When a new target feature is detected, the target feature is added to the target feature set. This procedure creates statistics on the frequency with which target features and opinion words appear in pairs, and which opinion words in the opinion word set and which target features in the target feature set appeared in the same text unit. , And the appearance frequency thereof, the relationship between the opinion word set and the target feature set can be obtained.

  It should be noted here that a person skilled in the art can define opinion word extraction rules and target feature extraction rules using an appropriate method of the prior art without being limited to the method disclosed herein.

  In step S102, the opinion word set and the target feature set are optimized based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set.

  In one embodiment, step S102 is performed in the following manner. 1) According to the score of the first set of the opinion word set and the target feature set, the score of the second set of the opinion word set and the target feature set is based on the relationship between the opinion word set and the target feature set. And 2) the score of the second set is adjusted using the similarity of the elements of the second set, and 3) the relationship between the opinion word set and the target feature set according to the adjusted score of the second set. The score of the first set is calculated, and 4) the score of the first set is adjusted using the similarity of the elements of the first set, and the opinion word according to the adjusted score of the first set Based on the relationship between the set and the target feature set, the score of the second set is calculated. 5) The difference between the scores before and after the adjustment of the first set satisfies a predetermined requirement, or the score of the second set If the difference between the scores before and after adjustment meets the predetermined requirements, The elements of the first set are ranked according to the adjusted score of the second set, and the elements of the second set are ranked according to the adjusted score of the second set. Next, this embodiment will be described with reference to FIG.

  In another embodiment, step S102 is performed in the following manner. 1) The score of the first set of the opinion word set and the target feature set is adjusted using the similarity of the elements of the first set, and 2) the opinion word set according to the adjusted score of the first set The score of the second set of the opinion word set and the target feature set is calculated based on the relationship between the target set and the target feature set, and 3) the score of the second set using the similarity of the elements of the second set And the score of the first set is calculated based on the relation between the opinion word set and the target feature set according to the adjusted score of the second set. 4) Scores before and after adjustment of the first set If the difference between the two sets satisfies a predetermined requirement, or if the difference between the scores before and after adjustment of the second set satisfies the predetermined requirement, the first set is adjusted according to the adjusted score of the first set. The elements of the set are ranked and the adjusted score of the second set It, elements of the second set is ranked. Next, this embodiment will be described with reference to FIG.

  In the present invention, for convenience of explanation, one of the opinion word set and the target feature set is a first set, and another set different from this set is a second set. The first set is either an opinion word set or a target set, and the second set is either an opinion word set or a target feature set, but the first set is not the same as the second set. That is, when the first set is an opinion word set, the second set is a target feature set, and when the first set is a target feature set, the second set is an opinion word set.

  In step S103, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  In one embodiment, step S103 is performed in the following manner.

  First, high rank opinion words are extracted from the optimized opinion word set according to the predetermined threshold value of the opinion word, and high rank target features are extracted from the optimized target feature set according to the predetermined threshold value of the target feature. Next, based on the relationship between the opinion word set and the target feature set, evaluation information is acquired from the high rank opinion word and the high rank target feature.

  The predetermined threshold value of the opinion word and the predetermined threshold value of the target feature can be obtained by various methods. For example, it may be an appropriate preset value specified according to experience values, determined according to the prior art, calculated according to a mathematical model, or a predetermined appropriate value set by a person skilled in the art.

  In the present embodiment, opinion words ranked before a specific ranking threshold set in advance in the opinion word set are extracted as high ranking opinion words. For example, if 10,000 opinion words are included in the opinion word set and the ranking threshold is set to 5,000, the top 5,000 opinion words are extracted. Similarly, by the method as described above, a target feature having a higher rank than a specific ranking threshold (for example, 4,000) set in advance is extracted from the target feature set.

  Since the relationship between the opinion word set and the target feature set is acquired according to the text unit in the corpus, the top 5,000 opinion words and the top 4,000 target features are paired according to this relationship. In this way, evaluation information is acquired by making a pair of opinion words and target features having a relationship within the same text unit.

  It should be noted that a person skilled in the art is not limited to the method disclosed here, but the evaluation information based on the optimized opinion word set and the optimized target feature set by an appropriate method of the prior art. It can be extracted.

  This is the end of the flow of FIG.

  FIG. 2 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention.

  In step 201, an opinion word set and a target feature set are acquired from the corpus.

  Since this step is similar to step S101, the details are omitted.

  In step S202, according to the score of the first set of the opinion word set and the target feature set, the second set of the opinion word set and the target feature set is determined based on the relationship between the opinion word set and the target feature set. A score is calculated.

  The initial value of the score of the first set can be obtained by a number of methods. In one embodiment, statistics are first generated for frequency information for each element of the first set of corpora, and then a score for the first set is obtained based on a pre-defined policy. For example, the frequency information of each element of the first set of corpus may be used as the score of the first set as it is, but the first frequency information is weighted using a weighting factor set in advance. It can also be a set score. Specifically, if the first set is an opinion word set, it is determined whether or not the opinion word is used together with the degree adverb in the degree adverb dictionary and the frequency of such use. Then, by correcting the frequency information of each element of the first set of corpus, the score of the first set is obtained.

As described above, the conversion relationship T _O-F from the opinion word set to the target feature set and the conversion relationship T _F-O from the target feature set to the opinion word set are in accordance with the relationship between the opinion word set and the target feature set. Desired.

When the first set is an opinion word set, the score of the target feature set is calculated according to the score of the opinion word set using the conversion relation _TO-F . If the first set is a target feature set, the score of the opinion word set, using the transformation relationship T _F-O, is calculated according to the score of the target feature set. The two calculation procedures are symmetric and both may perform vector matrix multiplication as is, or may perform vector matrix multiplication after weighting with a weighting factor, as well as other possible implementations by those skilled in the art. It can also be performed by any suitable method.

  In step S203, the score of the second set is adjusted using the similarity of the elements of the second set.

  In one embodiment, the second set of score adjustments made by using the second set of element similarity is based on the second set of element similarity and a priori confidence, An adjusted score is obtained by adjusting the score of the second set, and if the difference between the adjusted score and the score of the second set satisfies a predetermined requirement, the adjusted score is the adjusted score of the second set. If the difference between the adjusted score and the score of the second set does not satisfy the predetermined requirement, the score of the second set is set as the adjusted score. This is done by updating.

  In step S204, according to the adjusted score of the second set, the score of the first set is calculated based on the relationship between the opinion word set and the target feature set.

As described above, the conversion relationship T _O-F from the opinion word set to the target feature set and the conversion relationship T _F-O from the target feature set to the opinion word set are acquired according to the relationship between the opinion word set and the target feature set. Is done.

If the second set is an opinion word set, the score of the target feature set is calculated using the conversion relation T _O-F according to the score of the opinion word set. If the second set is subject feature set in accordance with the score of the target feature set, using the conversion relation T _F-O, score opinion word set is calculated. The two calculation procedures are symmetric and both may perform vector matrix multiplication as is, or may perform vector matrix multiplication after weighting with a weighting factor, as well as other possible implementations by those skilled in the art. It can also be performed by any suitable method.

  In step S205, the score of the first set is adjusted using the similarity of the elements of the first set.

  In one embodiment, the adjustment of the score of the first set using the similarity of the elements of the first set is based on the similarity of the elements of the first set and a priori confidence. An adjusted score is obtained by adjusting the score of one set, and if the difference between the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is the adjusted score of the first set. And the adjustment of the score of the first set is stopped, and if the difference between the adjusted score and the score of the first set does not satisfy the predetermined requirement, the score of the first set is This is done by updating.

  In this embodiment, whether or not the difference between the adjustment score and the score of the first set satisfies a predetermined requirement can be determined by various methods. For example, the number of adjustments performed may be counted. If the count result exceeds a preset number of iterations, the difference between the adjusted score and the score of the first set is considered to meet a predetermined requirement. As another example, if the difference between the adjusted score and the score of the first set is obtained and the difference is smaller than the predetermined threshold, the difference between the two satisfies the predetermined requirement. It is regarded. As yet another example, if the angle cosine value of the adjusted score and the first set of scores is calculated and the angle cosine value is less than a predetermined threshold, the difference between the two is a predetermined requirement. Is considered to be satisfied. Further, those skilled in the art are not limited to the methods disclosed herein, and can make this determination using any suitable method of the prior art.

  In step S206, it is determined whether the difference between the scores before and after adjustment of the first set satisfies a predetermined requirement.

  Step S206 can be performed in a number of ways. For example, the number of times the score of the first set is adjusted may be counted. If the count result exceeds a preset number of iterations, the difference in score before and after adjustment of the first set is considered to satisfy the predetermined requirement. As another example, if the difference between the scores before and after the first set of adjustments is obtained and the score difference is smaller than a predetermined threshold, the difference between the two satisfies the predetermined requirement. Is considered. As yet another example, if the angle cosine value of the score before and after the first set of adjustments is calculated and the angle cosine value is smaller than a predetermined threshold, the difference between the two will satisfy the predetermined requirement. It is considered to satisfy. Further, those skilled in the art are not limited to the methods disclosed herein, and can make this determination using any suitable method of the prior art.

  In another embodiment of the present invention, the step described in step S206 is replaced with a determination as to whether or not the difference between the scores before and after the adjustment of the second set satisfies a predetermined requirement.

  If it is determined in S206 that the difference between the scores before and after the adjustment of the first set satisfies the predetermined requirement, the flow further proceeds to step S207, but the difference in score satisfies the predetermined requirement. If it is determined that there is no, the process returns to step S202.

  In step S207, the elements of the first set are ranked according to the adjusted score of the first set, and the elements of the second set are ranked according to the adjusted score of the second set.

  The adjusted score of the first set and the adjusted score of the second set are both obtained by repeating the steps from S202 to S206 many times.

  The adjusted score of the first set is a vector whose value is a score corresponding to each element of the first set. Accordingly, the elements of the first set are re-ranked so that the elements of the first set are ranked in order of their corresponding score values. The first set thus obtained is the optimized first set.

  Similarly, the adjusted score of the second set is a vector whose value is a score corresponding to each element of the second set. Thus, the elements of the second set are re-ranked so that the elements of the second set are ranked in order of their corresponding score values according to the adjusted score of the second set. The second set thus obtained is the optimized second set.

  In step S208, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  Since this step is similar to step S103, details are omitted here.

  This is the end of the flow of FIG.

  It should be noted here that in one embodiment of the present invention, the embodiment shown in FIG. 2 is based on the frequency information of each element of the first set of corpora, and in accordance with a predefined policy, the first And further comprising the step of initializing the score of the set. Initialization performed in accordance with a pre-defined policy uses a method of initializing the score of the first set using the frequency information of each element of the first set of corpus as it is, and a weighting factor set in advance. Thus, it is possible to implement the method by initializing the score of the first set by weighting the frequency information. Specifically, if the first set is an opinion word set, it is determined whether or not the opinion word is used together with the degree adverb in the degree adverb dictionary and the frequency of such use. Then, the frequency information of each element of the first set of corpus is corrected, and the score of the first set is initialized using the corrected frequency information.

  In the present invention, as described above, the first set is either the opinion word set or the target feature set, and the second set is either the opinion word set or the target feature set. The set is not the same as the second set.

  In one embodiment of the present invention for performing the steps from S201 to S208, the first set is a target feature set and the second set is an opinion word set. After the opinion word set and the target feature set are acquired from the corpus in step S201, the score of the opinion word set is calculated in step S202 based on the relationship between the opinion word set and the target feature set according to the score of the target feature set. The Next, in step S203, the score of the opinion word set acquired from step S202 is adjusted using the similarity of the elements of the opinion word set. Thereafter, in step S204, the score of the target feature set is calculated based on the relationship between the opinion word set and the target feature set according to the score of the opinion word set adjusted in step S203. Next, in step S205, the score of the target feature set is adjusted using the similarity of the elements of the target feature set. Thereafter, in step S206, it is determined whether or not the difference between the score of the target feature set adjusted in step S205 and the score before adjustment satisfies a predetermined requirement. If the difference in score does not satisfy the predetermined requirement, the flow returns to step S202, but if the difference in score satisfies the predetermined requirement, the flow further proceeds to step S207, and in step S205. By ranking the elements of the target feature set according to the adjusted score of the target feature set, an optimized target feature set is obtained, and the opinion word set is scored according to the score of the opinion word set adjusted in step S203. By ranking the elements, an optimized opinion word set is obtained. Finally, in step S208, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  In another embodiment of the present invention for carrying out the steps from S201 to S208, the first set is an opinion word set and the second set is a target feature set. After the opinion word set and the target feature set are acquired from the corpus in step S201, in step S202, the score of the target feature set is calculated based on the relationship between the opinion word set and the target feature set according to the score of the opinion word set. The Next, in step S203, the score of the target feature set acquired from step S202 is adjusted using the similarity of the elements of the target feature set. Thereafter, in step S204, the score of the opinion word set is calculated based on the relationship between the opinion word set and the target feature set according to the score of the target feature set adjusted in step S203. Next, in step S205, the score of the opinion word set is adjusted using the similarity of the elements of the opinion word set. Thereafter, in step S206, it is determined whether or not the difference between the score of the opinion word set adjusted in step S205 and the score before adjustment satisfies a predetermined requirement. If the difference in score does not satisfy the predetermined requirement, the flow returns to step S202, but if the difference in score satisfies the predetermined requirement, the flow further proceeds to step S207, and in step S205. By ranking the elements of the opinion word set according to the score of the adjusted opinion word set, an optimized opinion word set is obtained, and the target feature set according to the score of the target feature set adjusted in step S203 is obtained. By ranking the elements, an optimized target feature set is obtained. Finally, in step S208, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  FIG. 3 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention.

  In step S301, an opinion word set and a target feature set are acquired from the corpus.

  Since this step is similar to step S101, the details are omitted.

  In step S302, the score of the first set of the opinion word set and the target feature set is adjusted using the similarity of the elements of the first set.

  The initial value of the score of the first set can be obtained by a number of methods. In one embodiment, statistics are first generated for frequency information for each element of the first set of corpora, and then a score for the first set is obtained based on a pre-defined policy. For example, the frequency information of each element of the first set of corpus may be used as the score of the first set as it is, but the first frequency information is weighted using a weighting factor set in advance. It can also be a set score. Specifically, if the first set is an opinion word set, it is determined whether or not the opinion word is used together with the degree adverb in the degree adverb dictionary and the frequency of such use. Then, by correcting the frequency information of each element of the first set of corpus, the score of the first set is obtained.

  Since the other part of this step is similar to step S205, details are omitted.

  In step S303, according to the adjusted score of the first set, the score of the second set of the opinion word set and the target feature set is calculated based on the relationship between the opinion word set and the target feature set. This step is similar to step S202.

  In step S304, the score of the second set is adjusted using the similarity of the elements of the second set. This step is similar to step S203.

  In step S305, it is determined whether or not the difference between the scores before and after adjustment of the second set satisfies a predetermined requirement. This step is similar to step S206.

  In another embodiment of the present invention, the step described in step S305 is replaced with a determination as to whether or not the difference between the scores before and after the adjustment of the first set satisfies a predetermined requirement.

  If the determination result in step S305 is “Yes”, the flow further proceeds to step S306, and if the determination result is “No”, the flow returns to step S302.

  In step S306, the elements of the first set are ranked according to the adjusted score of the first set, and the elements of the second set are ranked according to the adjusted score of the second set. This step is similar to step S207.

  In step S307, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  Since this step is similar to step S103, details are omitted here.

  The flow in FIG. 3 ends here.

  It should be noted here that in one embodiment of the present invention, the embodiment shown in FIG. 3 is further based on the frequency information of each element of the first set of corpora according to a pre-defined policy. A step of initializing a score of the set of. The initialization performed in accordance with the predefined policy is based on the method of initializing the score of the first set by using the frequency information of each element of the first set of corpus as it is, and a preset weighting factor. It can be implemented in such a way that the first set of scores is initialized by using and weighting the frequency information. Specifically, if the first set is an opinion word set, it is determined whether or not the opinion word is used together with the degree adverb in the degree adverb dictionary and the frequency of such use. By doing so, the frequency information of each element of the first set of corpus is corrected, and the corrected frequency information is used to initialize the score of the first set.

  In the present invention, as described above, the first set is either the opinion word set or the target set, and the second set is either the opinion word set or the target feature set. Is not identical to the second set.

  In one embodiment of the present invention for performing the steps from S301 to S307, the first set is a target feature set and the second set is an opinion word set. In another embodiment of the present invention for carrying out the steps from S301 to S307, the first set is an opinion word set and the second set is a target feature set.

  FIG. 4 is a flowchart of an evaluation information extraction method according to another embodiment of the present invention. The embodiment shown in FIG. 4 assumes the situation in the embodiment shown in FIG. 2 where the first set is an opinion word set and the second set is a target feature set.

  In step 401, an opinion word set and a target feature set are acquired from the corpus.

  In this example, after pre-processing the corpus, "the fuel consumption rate of the high emission car is very high", "the price of the mobile phone of this model is high", "the driving performance of this brand car is very good Suppose that four text units are obtained, "Purchased printer is too expensive."

  The opinion word set is acquired according to the opinion word extraction rule, and the target feature set is acquired according to the target feature extraction rule.

  Based on the above four text units of the corpus, an opinion word set including two opinion words “good” and “high” is acquired by following the opinion word extraction rule, and the appearance frequency of “good” is 1, The appearance frequency of “high” is determined to be 3.

  By following the target feature extraction rules based on the above four text units of the corpus, the target feature set includes three target features of “fuel consumption rate”, “price”, and “maneuverability”, and “fuel consumption” It is determined that the appearance frequency of “rate” is 1, the appearance frequency of “price” is 2, and the appearance frequency of “maneuverability” is 1.

  In step S402, according to the opinion word set score, the score of the target feature set is calculated based on the relationship between the opinion word set and the target feature set.

As shown in Table 1, the relationship between the two sets can be obtained according to the opinion word set and the target feature set acquired in step S401. The relationship (referred to as “C”) is expressed in the following matrix format.

  As described in step S202, the initial value of the opinion word score can be obtained by a number of methods. For example, in one embodiment of step S402, a pre-defined degree adverb dictionary is used. When an opinion word is used once together with an adverb to some extent in the degree adverb dictionary, the opinion word score is 1 (the initial value of the opinion word score is set to 0). Thus, the score of an opinion word can be calculated | required by producing the statistics of the frequency with which an opinion word is used with a degree adverb. The score of each opinion word is obtained by executing this for each opinion word in the opinion word set. In the present invention, scores of individual opinion words included in the opinion word set are collectively referred to as “opinion word set score”.

  The degree adverb dictionary includes, for example, “Taien (へ)”, “Most (most)”, “Extremely (pole)”, “Too much (thick)”, “Very (very)”, “Sufficiently” (Sufficient), very (extra), very much (extra), more (excess), excessive (excess), more (exit), extremely (extreme) ) "," Exceptionally (extraordinary) "," especially (extraordinary) "," little (pointed 儿) "," special (biased) "," slightly (稍) "," just a little (minor) " Can include one or more adverbs that indicate degrees such as “almost (几乎)”, “slightly (substantially)”, “too much”, and “especially”.

In the text unit acquired from the corpus, the appearance frequency of the opinion word “good” is 1, and the appearance frequency of “high” is 3. “Tahen” and “too” are degree adverbs specified in the degree adverb dictionary, so the frequency of occurrence of the opinion word “good” following the degree adverb is 1, and the appearance of the opinion word “high” following the degree adverb Frequency 2 is determined. As a result, the score of the opinion word set (referred to as “O _score ”) is as follows.

O _score = (1 2) (2)

  In the present invention, the higher the score of the opinion word in the opinion word set, the higher the frequency of use of the opinion word in the corpus, and the higher the usefulness of the opinion word for evaluation information extraction. Similarly, the larger the score of the target feature in the target feature set, the higher the usage frequency of the target feature in the corpus, and the higher the usefulness of the target feature for extracting the evaluation information.

  In the subsequent step S412, if a certain condition is satisfied, the flow further proceeds to step S413, where the opinion word set score is updated with the adjusted opinion word set score. Thereafter, the flow returns from step S413 to step S402, but in step S402, instead of the score of the opinion word set obtained by creating the statistics of the frequency information of the opinion words in the corpus, the flow is obtained in step S411. The score of the opinion word set after adjustment may be used.

The score of the target feature set (referred to as “F _score ”) can be calculated in several ways based on the relationship C between the opinion word set and the target feature set according to the score O _score of the opinion word set.

For example, the conversion relationship TO _-F from the opinion word set to the target feature set is obtained from the relationship C between the opinion word set and the target feature set as shown in Table 2. This conversion relationship T _O-F can be expressed in the following matrix format.

As an example, the product of the conversion relationship T _O-F and the score O _score of the opinion word set can be regarded as the score F _score of the target feature set.

In another example, the result of the above equation (4) can be weighted by a weighting coefficient (numerical value, vector, or matrix), and the weighted result is regarded as the score F _{score of the} target feature set.

  In step S403, the score of the target feature set is adjusted based on the similarity and a priori reliability of the elements of the target feature set to obtain an adjusted score.

Table 4 shows the similarity of the elements of the target feature sets in, which is expressed in the form: may be S _F.

Step S403 can be performed in a number of ways as indicated in step S203. For example, the score of the target feature set is adjusted based on the similarity S _F and a priori reliability of the elements of the target feature set (and alpha), adjusted score X is obtained.

X = S _F · F _score + α · F ¹ _score (6)

In Equation (6), F ¹ _score is an initial value of the score of the target feature set (for example, the score of the target feature set first obtained in accordance with the opinion word set score in step S402). The a priori confidence α may be specified in advance according to empirical values, determined according to the prior art, calculated according to a mathematical model, or set as a suitable preset value by those skilled in the art.

  In step S404, it is determined whether or not the difference between the adjustment score and the target feature set score satisfies a predetermined requirement.

  Similar to step S203, it can be determined by a number of methods whether or not the difference between the adjustment score and the score of the target feature set satisfies a predetermined requirement. For example, whether the difference between the adjusted score and the target feature set score satisfies a predetermined requirement is determined by counting the number of iterations, comparing the adjusted score and the target feature set score with a threshold, Alternatively, the determination may be made by evaluating the angle cosine value between the adjustment score and the score of the target feature set. Further, those skilled in the art are not limited to the methods disclosed herein, and can make this determination by any suitable method of the prior art.

  If it is determined in step S404 that the difference between the adjustment score and the score of the target feature set satisfies a predetermined requirement, the flow further proceeds to step S406, where it is determined that the predetermined requirement is not satisfied. In step S405, the flow proceeds to step S405.

  In step S405, the score of the target feature set is updated with the adjustment score.

  By replacing the adjustment score with the score of the target feature set, the score of the target feature set adjusted in step S403 becomes the adjustment score instead of the original score of the target feature set, so that the iterative procedure can be easily corrected.

  In step S406, the adjustment score is regarded as the adjusted score of the target feature set.

In this embodiment, the adjusted score of the target feature set is as follows.

  In step S407, the score of the opinion word set is calculated based on the relationship between the opinion word set and the target feature set according to the adjusted score of the target feature set.

The score O _score of the opinion word set can be calculated by a number of methods based on the relationship C between the opinion word set and the target feature set according to the score F _score of the target feature set.

For example, conversion relationship T _F-O from the target feature set to opinion word set, as shown in Table 3, obtained from the relationship C. This conversion relationship _TFO is expressed in the following matrix format.

As an example, the product of the conversion relationship T _O-F and the score F _score of the target feature set obtained from step S406 is regarded as the score O _score of the opinion word set.

  As another example, the result of the above equation (8) can be weighted by a weighting coefficient (numerical value, vector, or matrix), and the weighted result is regarded as the score of the opinion word set.

  In step S408, the score of the opinion word set is adjusted based on the similarity of the elements of the opinion word set and the a priori reliability, and an adjusted score is obtained.

Although the degree of similarity of the elements of the opinion word set is shown in Table 5, it may be expressed in the following format and may be So.

Step S408 can be performed in a number of ways as indicated in step S205. For example, the score of the opinion word set is adjusted based on the similarity So of the elements of the opinion word set and the a priori confidence level (assumed to be β), and the adjusted score Y is obtained.

Y = S _O · O _score + β · O ¹ _score (11)

In Expression (11), O ¹ _score indicates the initial _score of the opinion word set. The a priori confidence β may be specified in advance according to empirical values, determined according to the prior art, calculated according to a mathematical model, or may be set as a suitable preset value by those skilled in the art.

  In step S409, it is determined whether or not the difference between the adjustment score and the opinion word set score satisfies a predetermined requirement.

  As in step S205, it can be determined in a number of ways whether the difference between the adjustment score and the opinion word set score satisfies a predetermined requirement. For example, when the number of adjustments performed is counted and the count result exceeds a preset number of iterations, the difference between the adjustment score and the opinion word set score is considered to satisfy a predetermined requirement. In another example, if the difference between the adjusted score and the opinion set score is found and the difference is less than the predefined threshold, the difference is considered to meet the predefined requirements. It is. As another example, if the angle cosine value of the adjustment score and the opinion set score is calculated and the angle cosine value is less than the default threshold, the difference between the two satisfies the default requirement. Is considered to be. Further, those skilled in the art are not limited to the methods disclosed herein, and can make this determination by any suitable method of the prior art.

  If it is determined in step S409 that the difference between the adjustment score and the score of the target feature set satisfies a predetermined requirement, the flow further proceeds to step S411, where it is determined that this difference does not satisfy the predetermined requirement. If so, flow proceeds to step S410.

  In step S410, the score of the opinion word set is updated with the adjusted score.

  In step S411, the adjustment score is regarded as the score of the adjusted opinion word set.

  In step S412, it is determined whether the difference between the scores before and after adjustment of the opinion word set satisfies a predetermined requirement.

  This determination can be performed, for example, by counting the number of adjustments of the score of the opinion word set. If the count result exceeds the preset number of iterations, the difference in score before and after adjustment of the opinion word set is considered to satisfy the predetermined requirement. As another example, if the difference in scores before and after adjustment of the opinion word set is determined and the difference in scores is smaller than the predetermined threshold, the difference between the two is considered to satisfy the predetermined requirement. It is. As another example, if the angle cosine value of the score before and after adjustment of the opinion word set is calculated and the angle cosine value is smaller than the predetermined threshold value, the difference between the two satisfies the predetermined requirement. Is considered to be. Further, those skilled in the art are not limited to the methods disclosed herein, and can make this determination by any suitable method of the prior art.

  In another embodiment of the present invention, the step shown in step S412 is replaced with a determination as to whether or not the difference between the scores before and after the target feature set adjustment satisfies a predetermined requirement. Even in this case, the present invention can be implemented, and the effects of the present invention are achieved.

  If it is determined in step S412 that the difference between the scores before and after adjustment of the opinion word set satisfies the predetermined requirement, the flow further proceeds to step S414, and this difference does not satisfy the predetermined requirement. If it is determined, the flow proceeds to step S413.

  In step S413, the score of the opinion word set is updated with the adjusted score of the opinion word set.

  In step S414, the elements of the opinion word set are ranked according to the adjusted score of the opinion word set, and the elements of the target feature set are ranked according to the adjusted score of the target feature set.

In the present embodiment, the latest score of the opinion word set obtained from step S412 is O _score = ( ₂₄ ), and “high” is “good” in the opinion word set including the opinion words “good” and “high”. Will be ranked higher. That is, opinion words are ranked in the order of “high” and “good”. Here, the opinion word set for which ranking has been completed is referred to as an optimized opinion word set.

３つの対象特徴「燃料消費率」、「価格」、および「操縦性」を含む対象特徴集合において、対象特徴は、「価格」、「燃料消費率」、「操縦性」の順にランク付けされる。ここでは、ランク付けが完了した対象特徴集合は、最適化された対象特徴集合と呼ばれる。 Assume that the adjusted score of the target feature set is as follows.

In the target feature set including three target features “fuel consumption rate”, “price”, and “maneuverability”, the target features are ranked in the order of “price”, “fuel consumption rate”, “maneuverability”. . Here, the target feature set for which ranking has been completed is referred to as an optimized target feature set.

  In step S415, evaluation information is extracted based on the optimized opinion word set and the optimized target feature set.

  This step is similar to step S103.

  In this embodiment, the default threshold value of opinion words is set to 2, and a high rank opinion word that is higher than the second opinion word in the opinion word set, that is, “high” is extracted.

  Further, assuming that the predetermined threshold value of the target feature is 3, target features higher than the third highest target feature in the target feature set, that is, “price” and “fuel consumption rate” are extracted.

  After that, whether the extracted opinion word “high” is related to the extracted target characteristics “price”, “fuel consumption rate” (for example, whether these words co-occur in the same text unit) Is determined according to the relationship between the opinion word set and the target feature set. In the present embodiment, the evaluation information is constituted by a combination of the extracted opinion word and the extracted target feature that co-occurs at least once within the same text unit. Since “price ... high” and “fuel consumption rate ... high” appear in the four text units of this embodiment, [fuel consumption rate, high] and [price, high] are acquired as evaluation information. .

  The flow in FIG. 4 ends here.

  FIG. 5 is a block diagram of an evaluation information extraction apparatus 500 according to one embodiment of the present invention. The evaluation information extraction apparatus 500 includes an acquisition unit 501, an optimization unit 502, and an extraction unit 503.

  The acquisition unit 501 is configured to acquire the opinion word set and the target feature set from the corpus. The optimization unit 502 optimizes the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. It is configured. The extraction unit 503 is configured to extract evaluation information based on the optimized opinion word set and the optimized target feature set.

  In one embodiment, the acquisition means 501 generates a set of opinion words according to opinion word extraction rules based on the means configured to perform corpus preprocessing to acquire text units and the acquired text units. Means configured to acquire and means configured to acquire a target feature set according to a target feature extraction rule based on the acquired text unit.

  In one embodiment, the opinion word extraction rules are: “a clause immediately following a degree adverb, an adjective, a clause that does not include a function word, a clause that is shorter or equal in length to the longest opinion word, and a minimum occurrence frequency. It is also possible to extract one or more phrases from the text unit that have a higher appearance frequency than the opinion word and make it an opinion word.

  In one embodiment, the target feature extraction rules are “basic noun phrases, combinations of basic noun phrases, basic noun phrases and noun / dynamic noun combinations, basic noun phrases and qualifier combinations, and qualifiers and nouns”. One or more text units that contain combinations of verbal nouns, phrases that do not contain function words, phrases that are shorter or equal to the longest target feature, and phrases that appear more frequently than the target feature with the lowest occurrence frequency It may be “extracted and set as a target feature”.

In one embodiment, the optimization means 502 relates the scores of the second set of the opinion word set and the target feature set according to the scores of the first set of the opinion word set and the target feature set. First transform calculating means configured to calculate based on the second set of means, and second adjusting means configured to adjust the score of the second set using the similarity of the elements of the second set; The second transformation calculating means configured to calculate the score of the first set based on the relationship according to the adjusted score of the second set, and the first using the similarity of the elements of the first set Adjusting the score of the set of the first set, and adjusting the score of the second set according to the adjusted score of the first set based on the relationship; and before and after adjusting the first set The difference in scores of the two meets the predetermined requirements or the second Rank the elements of the first set according to the adjusted score of the first set and the adjusted score of the second set when the difference between the scores before and after adjustment of the set satisfies a predetermined requirement And ranking means configured to rank the elements of the second set according to
In one example, the optimizer 502 is configured to initialize the score of the first set according to a predefined policy based on the frequency information of each element of the first set of corpus. Means.
In another example, the first adjuster of the optimizer 502 adjusts the score of the first set by adjusting the score of the first set based on the similarity and a priori confidence of the elements of the first set. And if the difference between the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the first set. Means configured to stop adjusting the score of the first set, and if the difference between the adjusted score and the score of the first set does not satisfy the predetermined requirement, the score of the first set is Means configured to update with an adjusted score.
In another example, the second adjusting means of the optimizing means 502 determines the adjusted score by adjusting the score of the second set based on the similarity and a priori confidence of the elements of the second set. If the difference between the means configured to determine and the adjusted score and the score of the second set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the second set and Means configured to stop adjusting the score of the second set and adjust the score of the second set if the difference between the adjusted score and the score of the second set does not meet the predetermined requirement Means configured to update with a score.

In one embodiment, the optimization means 502 is configured to adjust the score of the first set of the opinion word set and the target feature set using the similarity of the elements of the first set. First conversion means, and conversion calculation means configured to calculate the score of the second set of the opinion word set and the target feature set according to the relationship according to the adjusted score of the first set, Adjusting the score of the second set using the similarity of the elements of the second set to calculate the score of the first set based on the relationship according to the adjusted score of the second set The difference between the second adjustment means configured in the above and the score before and after adjustment of the first set satisfies a predetermined requirement, or the difference between the scores before and after adjustment of the second set satisfies a predetermined requirement If the first set according to the adjusted score of the first set It ranked elements of the set, and includes a configured ranked unit to rank the elements of the second set according to the score after adjustment of the second set.
In one example, the optimizer 502 is configured to initialize the score of the first set according to a predefined policy based on the frequency information of each element of the first set of corpus. Means.
In another example, the first adjuster of the optimizer 502 adjusts the score of the first set by adjusting the score of the first set based on the similarity and a priori confidence of the elements of the first set. If the difference between the means configured to determine the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the first set; And the means configured to stop adjusting the score of the first set, and the difference between the adjusted score and the score of the first set does not satisfy the predetermined requirement, the score of the first set Means configured to update with an adjusted score.
In another example, the second adjustment means of the optimization means 502 may adjust the adjustment score by adjusting the score of the second set based on the similarity of elements of the second set and the a priori confidence. And if the difference between the adjusted score and the score of the second set satisfies the predetermined requirement, the adjusted score is determined as the adjusted score of the second set. Means configured to stop adjusting the score of the second set, and if the difference between the adjusted score and the score of the second set does not satisfy the predetermined requirement, the score of the second set is Means configured to update with an adjusted score.

  In one embodiment, the extracting means 503 includes means adapted to extract high ranking opinion words from the optimized opinion word set according to a predetermined threshold of opinion words, and an optimized object. Based on the means configured to extract high-order target features from the feature set according to a predetermined threshold of the target features, and based on the relationship between the opinion word set and the target feature set, Means for obtaining evaluation information from the target feature.

  It should be noted here that the present invention is not limited to the processing of Chinese, but can be used for processing of many kinds of languages such as English, French, and German.

  The method according to the invention can also be implemented as software, hardware or a combination of software and hardware. The hardware portion can be implemented using dedicated logic circuitry, and the software portion can be stored in memory and executed by a suitable instruction execution system such as a microprocessor, personal computer (PC), mainframe or the like.

  The present invention has been described above with reference to preferred embodiments. However, the present invention is not necessarily limited to the above embodiments, and various modifications can be made within the scope of the technical idea. it can.

  Further, a part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.

(Appendix 1)
Obtaining an opinion word set and a target feature set from a corpus;
Optimizing the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set;
Extracting the evaluation information based on the optimized opinion word set and the optimized target feature set. An evaluation information extraction method comprising:

(Appendix 2)
Obtaining the opinion word set and the target feature set from the corpus,
Performing corpus preprocessing to obtain text units;
Obtaining a set of opinion words according to opinion word extraction rules based on the obtained text units;
The evaluation information extracting method according to claim 1, further comprising: acquiring a target feature set according to the target feature extraction rule based on the acquired text unit.

(Appendix 3)
The opinion word extraction rule has a phrase that immediately follows a degree adverb, an adjective, a phrase that does not include a function word, a phrase that is shorter or equal in length to the longest opinion word, and an appearance word that has a minimum frequency of appearance The evaluation information extracting method according to appendix 2, wherein one or more texts are extracted from the text unit to be an opinion word.

(Appendix 4)
The target feature extraction rule is a basic noun phrase, a combination of basic noun phrases, a combination of a basic noun phrase and a noun / verbal noun, a combination of a basic noun phrase and a qualifier, and a combination of a qualifier and a noun / dynamic noun, a function Extract from the text unit one or more phrases that do not contain words, phrases that are shorter or equal to the longest target feature, and those that appear more frequently than the target feature that has the lowest occurrence frequency, and use them as target features The evaluation information extracting method according to supplementary note 2, wherein the evaluation information is extracted.

(Appendix 5)
The step of optimizing the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set,
Calculating a score of the second set of the opinion word set and the target feature set based on the relationship according to the score of the first set of the opinion word set and the target feature set;
Adjusting the score of the second set using the similarity of the elements of the second set;
Calculating a score of the first set based on the relationship according to the adjusted score of the second set;
Adjusting the score of the first set using the similarity of the elements of the first set and calculating the score of the second set based on the relationship according to the adjusted score of the first set;
After adjustment of the first set, if the difference in score before and after adjustment of the first set meets a predetermined requirement, or the difference in score before and after adjustment of the second set satisfies a predetermined requirement Ranking the elements of the first set according to the score of and ranking the elements of the second set according to the adjusted score of the second set. Information extraction method.

(Appendix 6)
The step of optimizing the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set,
Adjusting the score of the first set of the opinion word set and the target feature set using the similarity of the elements of the first set;
Calculating a score of the second set of the opinion word set and the target feature set according to the adjusted score of the first set based on the relationship;
Adjusting the score of the second set using the similarity of the elements of the second set to calculate the score of the first set based on the relationship according to the adjusted score of the second set; ,
If the difference between the scores before and after adjustment of the first set meets a predetermined requirement, or if the difference between the scores before and after adjustment of the second set satisfies a predetermined requirement, the adjustment of the first set Ranking the elements of the first set according to a later score and ranking the elements of the second set according to the adjusted score of the second set. Evaluation information extraction method.

(Appendix 7)
Appendix 5 or appendix 6 further comprising the step of initializing the score of the first set according to a predefined policy based on the frequency information of each element of the first set of corpus Evaluation information extraction method.

(Appendix 8)
Adjusting the score of the first set using the similarity of the elements of the first set;
Determining an adjusted score by adjusting the score of the first set based on the similarity and a priori confidence of the elements of the first set;
When the difference between the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the first set, and the adjustment of the score of the first set is adjusted. A step to stop,
Updating the first set of scores with the adjusted score if the difference between the adjusted score and the score of the first set does not meet the predetermined requirement;
Adjusting the score of the second set using the similarity of the elements of the second set;
Determining an adjusted score by adjusting the score of the second set based on the similarity and a priori confidence of the elements of the second set;
If the difference between the adjusted score and the score of the second set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the second set, and the adjustment of the score of the second set is stopped. And steps to
Supplementary note 5 or Supplementary note 6, comprising the step of updating the score of the second set with the adjusted score when the difference between the adjusted score and the score of the second set does not satisfy the predetermined requirement The evaluation information extraction method described in 1.

(Appendix 9)
Extracting the evaluation information based on the optimized opinion word set and the optimized target feature set,
Extracting high ranking opinion words from the optimized opinion word set according to a predetermined threshold of opinion words;
Extracting a high-order target feature from the optimized target feature set according to a predetermined threshold of the target feature;
The evaluation information extraction method according to appendix 1, further comprising the step of acquiring evaluation information from a high ranking opinion word and a high ranking target feature based on a relationship between the opinion word set and the target feature set.

(Appendix 10)
An acquisition means configured to acquire an opinion word set and a target feature set from a corpus;
Configured to optimize opinion word set and target feature set based on relationship between opinion word set and target feature set, similarity of elements of opinion word set, and similarity of elements of target feature set Optimization means;
An evaluation information extraction apparatus comprising: extraction means configured to extract evaluation information based on an optimized opinion word set and an optimized target feature set.

(Appendix 11)
The acquisition means includes
Means configured to pre-process the corpus to obtain a text unit;
Means configured to obtain a set of opinion words according to the opinion word extraction rules based on the obtained text units;
The evaluation information extraction device according to appendix 10, further comprising means configured to acquire a target feature set according to a target feature extraction rule based on the acquired text unit.

(Appendix 12)
The opinion word extraction rule has a phrase that immediately follows a degree adverb, an adjective, a phrase that does not include a function word, a phrase that is shorter or equal in length to the longest opinion word, and an appearance word that has a minimum frequency of appearance The evaluation information extraction apparatus according to appendix 11, wherein one or more texts are extracted from the text unit to form an opinion word.

(Appendix 13)
The target feature extraction rule is a basic noun phrase, a combination of basic noun phrases, a combination of a basic noun phrase and a noun / verbal noun, a combination of a basic noun phrase and a qualifier, and a combination of a qualifier and a noun / dynamic noun, a function Extract from the text unit one or more phrases that do not contain words, phrases that are shorter or equal to the longest target feature, and those that appear more frequently than the target feature that has the lowest occurrence frequency, and use them as target features The evaluation information extraction device according to attachment 11, wherein the evaluation information extraction device is a device.

(Appendix 14)
The optimization means includes
A first transformation configured to calculate a score of a second set of the opinion word set and the target feature set according to a relationship according to a score of the first set of the opinion word set and the target feature set A calculation means;
Second adjustment means configured to adjust the score of the second set using the similarity of the elements of the second set;
Second transformation calculating means configured to calculate a score of the first set based on the relationship according to the adjusted score of the second set;
The score of the first set is adjusted using the similarity of the elements of the first set, and the score of the second set is calculated based on the relationship according to the adjusted score of the first set. First adjusting means,
After adjustment of the first set, if the difference in score before and after adjustment of the first set meets a predetermined requirement, or the difference in score before and after adjustment of the second set satisfies a predetermined requirement And ranking means configured to rank the elements of the first set according to the score of and rank the elements of the second set according to the adjusted score of the second set, The evaluation information extraction apparatus according to Supplementary Note 10.

(Appendix 15)
The optimization means includes
First adjusting means configured to adjust the score of the first set of the opinion word set and the target feature set using the similarity of the elements of the first set;
Transformation calculation means configured to calculate the score of the second set of the opinion word set and the target feature set according to the adjusted score of the first set based on the relationship;
Adjusting the score of the second set using the similarity of the elements of the second set to calculate the score of the first set according to the relationship according to the adjusted score of the second set Configured second adjusting means;
If the difference between the scores before and after adjustment of the first set meets a predetermined requirement, or if the difference between the scores before and after adjustment of the second set satisfies a predetermined requirement, the adjustment of the first set Ranking means configured to rank the elements of the first set according to the later score and rank the elements of the second set according to the adjusted score of the second set; The evaluation information extraction apparatus according to Supplementary Note 10.

(Appendix 16)
The optimization means includes
Appendix 14 or Appendix comprising means configured to initialize a score of the first set according to a predefined policy based on frequency information of each element of the first set of corpus 15. The evaluation information extraction device according to 15.

(Appendix 17)
The first adjusting means includes
Means configured to determine an adjusted score by adjusting the score of the first set based on the similarity and a priori confidence of the elements of the first set;
When the difference between the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the first set, and the adjustment of the score of the first set is adjusted. Means configured to stop;
Means configured to update the score of the first set with the adjusted score if the difference between the adjusted score and the score of the first set does not meet the predetermined requirement;
The second adjusting means includes
Means configured to determine an adjusted score by adjusting the score of the second set based on the similarity and a priori confidence of the elements of the second set;
If the difference between the adjusted score and the score of the second set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the second set, and the adjustment of the score of the second set is stopped. Means configured to:
Means for updating the score of the second set with the adjusted score if the difference between the score of the adjusted score and the score of the second set does not satisfy the predetermined requirement. The evaluation information extraction device according to Supplementary Note 14 or Supplementary Note 15.

(Appendix 18)
The extraction means includes
Means configured to extract high ranking opinion words from the optimized opinion word set according to a predetermined threshold of opinion words;
Means configured to extract high-order target features from the optimized target feature set according to a predetermined threshold of the target features;
The evaluation according to claim 10, further comprising means configured to acquire evaluation information from a high ranking opinion word and a high ranking target feature based on a relationship between the opinion word set and the target feature set. Information extraction device.

501: Acquisition means 502: Optimization means 503: Extraction means

Claims (8)

An evaluation information extraction method by an evaluation information extraction device,
The obtaining unit of the evaluation information extracting device executes a step of obtaining an opinion word set and a target feature set from a corpus,
The optimization means of the evaluation information extracting device is configured to use the opinion word set and the target feature based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. Perform steps to optimize the set,
The extraction means of the evaluation information extraction device performs a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set,
Obtaining the opinion word set and the target feature set from the corpus,
Performing corpus preprocessing to obtain text units;
Obtaining a set of opinion words according to opinion word extraction rules based on the obtained text units;
Obtaining a target feature set according to a target feature extraction rule based on the acquired text unit,
The opinion word extraction rule has a phrase that immediately follows a degree adverb, an adjective, a phrase that does not include a function word, a phrase that is shorter or equal in length to the longest opinion word, and an appearance word that has a minimum frequency of appearance An evaluation information extraction method characterized in that one or more texts are extracted from a text unit as opinion words . An evaluation information extraction method by an evaluation information extraction device,
The obtaining unit of the evaluation information extracting device executes a step of obtaining an opinion word set and a target feature set from a corpus,
The optimization means of the evaluation information extracting device is configured to use the opinion word set and the target feature based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. Perform steps to optimize the set,
The extraction means of the evaluation information extraction device performs a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set,
Obtaining the opinion word set and the target feature set from the corpus,
Performing corpus preprocessing to obtain text units;
Obtaining a set of opinion words according to opinion word extraction rules based on the obtained text units;
Obtaining a target feature set according to a target feature extraction rule based on the acquired text unit,
The target feature extraction rule is a basic noun phrase, a combination of basic noun phrases, a combination of a basic noun phrase and a noun / verbal noun, a combination of a basic noun phrase and a qualifier, and a combination of a qualifier and a noun / dynamic noun, a function Extract from the text unit one or more phrases that do not contain words, phrases that are shorter or equal to the longest target feature, and those that appear more frequently than the target feature that has the lowest occurrence frequency, and use them as target features An evaluation information extraction method characterized by being a thing . An evaluation information extraction method by an evaluation information extraction device,
The obtaining unit of the evaluation information extracting device executes a step of obtaining an opinion word set and a target feature set from a corpus,
The optimization means of the evaluation information extracting device is configured to use the opinion word set and the target feature based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. Perform steps to optimize the set,
The extraction means of the evaluation information extraction device performs a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set,
The step of optimizing the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set,
Calculating a score of the second set of the opinion word set and the target feature set based on the relationship according to the score of the first set of the opinion word set and the target feature set;
Adjusting the score of the second set using the similarity of the elements of the second set;
Calculating a score of the first set based on the relationship according to the adjusted score of the second set;
Adjusting the score of the first set using the similarity of the elements of the first set and calculating the score of the second set based on the relationship according to the adjusted score of the first set;
After adjustment of the first set, if the difference in score before and after adjustment of the first set meets a predetermined requirement, or the difference in score before and after adjustment of the second set satisfies a predetermined requirement Ranking the elements of the first set according to the score of and ranking the elements of the second set according to the adjusted score of the second set;
The evaluation information extraction method characterized by including . An evaluation information extraction method by an evaluation information extraction device,
The obtaining unit of the evaluation information extracting device executes a step of obtaining an opinion word set and a target feature set from a corpus,
The optimization means of the evaluation information extracting device is configured to use the opinion word set and the target feature based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. Perform steps to optimize the set,
The extraction means of the evaluation information extraction device performs a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set,
The step of optimizing the opinion word set and the target feature set based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set,
Adjusting the score of the first set of the opinion word set and the target feature set using the similarity of the elements of the first set;
Calculating a score of the second set of the opinion word set and the target feature set according to the adjusted score of the first set based on the relationship;
Adjusting the score of the second set using the similarity of the elements of the second set to calculate the score of the first set based on the relationship according to the adjusted score of the second set; ,
If the difference between the scores before and after adjustment of the first set meets a predetermined requirement, or if the difference between the scores before and after adjustment of the second set satisfies a predetermined requirement, the adjustment of the first set Ranking the elements of the first set according to a later score, and ranking the elements of the second set according to the adjusted score of the second set;
The evaluation information extraction method characterized by including . The optimization means of the evaluation information extracting device executes a step of initializing a score of the first set according to a predefined policy based on frequency information of each element of the first set of corpus The evaluation information extraction method according to claim 3 or 4, wherein the evaluation information is extracted . Adjusting the score of the first set using the similarity of the elements of the first set;
Determining an adjusted score by adjusting the score of the first set based on the similarity and a priori confidence of the elements of the first set;
When the difference between the adjusted score and the score of the first set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the first set, and the adjustment of the score of the first set is adjusted. A step to stop,
Updating the first set of scores with the adjusted score if the difference between the adjusted score and the score of the first set does not meet the predetermined requirement;
Adjusting the score of the second set using the similarity of the elements of the second set;
Determining an adjusted score by adjusting the score of the second set based on the similarity and a priori confidence of the elements of the second set;
If the difference between the adjusted score and the score of the second set satisfies a predetermined requirement, the adjusted score is determined as the adjusted score of the second set, and the adjustment of the score of the second set is stopped. And steps to
Updating the second set of scores with the adjusted score if the difference between the adjusted score and the score of the second set does not satisfy a predetermined requirement. Item 5. The evaluation information extraction method according to Item 4 . An evaluation information extraction method by an evaluation information extraction device,
The obtaining unit of the evaluation information extracting device executes a step of obtaining an opinion word set and a target feature set from a corpus,
The optimization means of the evaluation information extracting device is configured to use the opinion word set and the target feature based on the relationship between the opinion word set and the target feature set, the similarity of the elements of the opinion word set, and the similarity of the elements of the target feature set. Perform steps to optimize the set,
The extraction means of the evaluation information extraction device performs a step of extracting evaluation information based on the optimized opinion word set and the optimized target feature set,
Extracting the evaluation information based on the optimized opinion word set and the optimized target feature set,
Extracting high ranking opinion words from the optimized opinion word set according to a predetermined threshold of opinion words;
Extracting a high-order target feature from the optimized target feature set according to a predetermined threshold of the target feature;
Obtaining evaluation information from high ranking opinion words and high ranking target features based on the relationship between the opinion word set and the target feature set;
The evaluation information extraction method characterized by including . An acquisition means configured to acquire an opinion word set and a target feature set from a corpus;
Configured to optimize opinion word set and target feature set based on relationship between opinion word set and target feature set, similarity of elements of opinion word set, and similarity of elements of target feature set Optimization means;
Extraction means configured to extract evaluation information based on the optimized opinion word set and the optimized target feature set, and
The acquisition means includes
Means configured to pre-process the corpus to obtain a text unit;
Means configured to obtain a set of opinion words according to the opinion word extraction rules based on the obtained text units;
Comprising means configured to acquire a target feature set according to a target feature extraction rule based on the acquired text unit;
The opinion word extraction rule has a phrase that immediately follows a degree adverb, an adjective, a phrase that does not include a function word, a phrase that is shorter or equal in length to the longest opinion word, and an appearance word that has a minimum frequency of appearance Extract one or more sentences from a text unit to make an opinion word
An evaluation information extraction apparatus characterized by that .

Images