JP6715758B2 - Classifier generation device, classifier generation method, and computer program - Google Patents

Description

The present invention relates to a classifier generation device, a classifier generation method, and a computer program.

For example, Patent Documents 1 and 2 are known as conventional text classification techniques. In the text classification technique of Patent Document 1, texts are classified into hierarchical categories. In the text classification technique of Patent Document 2, in the hierarchical classification of documents, a category of information is configured as a binary tree including nodes of a binary tree including information related to search.

JP, 2006-251975, A JP, 2006-18829, A

However, in the above-described conventional text classification technique, the classification accuracy may decrease as the hierarchy becomes deeper in the hierarchical categories. In addition, when the hierarchical structure of the categories becomes large, there is a possibility that an unnecessary category may be included in the actual classification and the efficiency of the classification process may be deteriorated. As a method of coping with such a problem, it is possible to manually optimize the hierarchical structure of categories, but this may be difficult due to the burden of increased processing time and increased costs.

The present invention has been made in view of such circumstances, and an object thereof is to improve the efficiency of text classification processing.

One aspect of the present invention is a classifier generation device that generates a classifier for classifying each answer to a semantically opposite question obtained from the same respondent, in which words are replaced or words are added or deleted. The sentence changing unit for changing the meaning of the sentence, and the sentence included in the first classification structure data indicating the classification item and the hierarchy of the classification item to which the sentence data is classified are changed by the sentence changing unit to perform the second classification. Structure data is generated, and a classifier is generated using the second classification structure data and the second teacher data to which the classification item of the second classification structure data and the label for identifying the hierarchy of the classification item are attached. A classifier generation unit is provided , and the sentence modification unit performs a morphological analysis and a dependency analysis on the input sentence. As a result of the morphological analysis and the dependency analysis, when the input sentence has a predicate term structure, A classifier generation device that acquires an antonym of a word of the predicate from an antonym dictionary and replaces the word of the predicate in the input sentence with the acquired antonym .
In one aspect of the present invention, when the input sentence has a predicate-argument structure, when the antonym of the word of the predicate cannot be obtained from an antonym dictionary, A classifier generation device that adds a negative word to a word of a predicate or deletes a negative word from the word of the predicate in the input sentence.
In one aspect of the present invention, the sentence modification unit, as a result of morphological analysis and dependency analysis on the input sentence, when a predicate-argument structure is not found in the input sentence, It is a classifier generation device that acquires an antonym from an antonym dictionary and replaces the word of the word stop in the input sentence with the acquired antonym.

In one aspect of the present invention, the classifier generation unit causes the sentence modification unit to include a sentence included in the first teacher data to which a label for identifying a classification item of the first classification structure data and a hierarchy of the classification item is added. It is a classifier generation device that is changed to generate the second teacher data.

According to one aspect of the present invention, a representative sentence generation unit that classifies the classification target text data using the classifier generated by the classifier generation unit and generates a representative sentence of the classifier based on the result of the classification is provided. A classifier generation device further provided.

One aspect of the present invention uses the classifier generated by the classifier generation unit to classify the classification target data related to the second classification structure data, and corrects the classifier based on a result of the classification. The classifier generation device further includes a classifier correction unit that determines whether or not it is.

One aspect of the present invention is to provide a sentence changing unit that replaces a word or adds or deletes a word to change the meaning of a sentence, and a classification structure data that is a destination of classifying text data and classification structure data indicating a hierarchy of the classification item. And a classifier generating unit that changes the sentence included in the first classifier included by the sentence changing unit to generate a second classifier.

One aspect of the present invention is a classifier generation method for generating a classifier for classifying each answer to a semantically opposite question obtained from the same respondent, wherein the classifier generation device replaces or replaces words. The sentence changing step of changing the meaning of the sentence by adding or deleting words, and the classifier generating device are included in the first classification structure data indicating the classification item and the hierarchy of the classification item to which the sentence data is classified. A sentence is changed by the sentence changing step to generate second classification structure data, and the second classification structure data, a classification item of the second classification structure data, and a label for identifying a hierarchy of classification items are attached. a classifier generation step of generating a classifier using a 2 teacher data, only contains the sentence changing step performs a morphological analysis and dependency analysis on the input sentence, the morphological analysis and dependency analysis As a result, when there is a predicate-argument structure in the input sentence, an antonym of the word of the predicate is acquired from an antonym dictionary, and the word of the predicate in the input sentence is replaced with the acquired antonym, classification It is a container generation method.

One aspect of the present invention is a sentence changing step in which a classifier generating device replaces a word or adds or deletes a word to change the meaning of a sentence, and the classifier generating device classifies sentence data. A classifier generation step of generating a second classifier by changing the sentence included in the first classifier having the classification structure data indicating the hierarchy of the classification items and the classification items by the sentence changing step. Is the way.

One aspect of the present invention is to replace a word or add or delete a word in a computer of a classifier generation device that generates a classifier for classifying each answer to a semantically opposite question obtained from the same respondent. Sentence change function for changing the meaning of the sentence, and the sentence included in the first classification structure data indicating the classification item and the hierarchy of the classification item from which the sentence data is classified by the sentence modification function to change the second classification Structure data is generated, and a classifier is generated using the second classification structure data and the second teacher data to which the classification item of the second classification structure data and the label for identifying the hierarchy of the classification item are attached. A computer program for realizing a classifier generation function , wherein the sentence modification function performs a morphological analysis and a dependency analysis on an input sentence, and a result of the morphological analysis and the dependency analysis results in the input sentence. A computer program that acquires an antonym of a word of the predicate from an antonym dictionary when there is a predicate structure and replaces the word of the predicate in the input sentence with the acquired antonym .

According to one embodiment of the present invention, a sentence changing function of changing a meaning of a sentence by replacing a word or adding or deleting a word to a computer, a classification item that is a destination of classifying text data, and a classification indicating a hierarchy of the classification item. It is a computer program for realizing a classifier generation function of generating a second classifier by changing a sentence included in a first classifier having structural data by the sentence changing function.

According to the present invention, it is possible to obtain the effect that the efficiency of text classification processing can be improved.

It is a block diagram showing an example of composition of classification system 1 concerning one embodiment. It is an explanatory view of the outline of the classifier generation method concerning one embodiment. It is explanatory drawing of the classifier which concerns on one Embodiment. It is explanatory drawing about the classifier generation part 11 which concerns on one Embodiment. It is explanatory drawing about the representative sentence production|generation part 13 which concerns on one Embodiment. It is explanatory drawing about the classifier correction|amendment part 14 which concerns on one Embodiment. It is explanatory drawing which shows the procedure of the example 1 of the classifier generation method which concerns on one Embodiment. It is explanatory drawing which shows the procedure of the example 2 of the classifier generation method which concerns on one Embodiment. It is explanatory drawing which shows the procedure of the example 3 of the classifier generation method which concerns on one Embodiment. It is explanatory drawing of the modification 1 which concerns on one Embodiment. It is a figure which shows the structural example of classification structure data D1 which concerns on one Embodiment. It is a flow chart which shows the procedure of the example of the classifier amendment method concerning one embodiment. It is a figure showing an example of amendment to classification structure data D1 concerning one embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration example of a classification system 1 according to an embodiment. 1, the classification system 1 includes a classifier generation device 10, a content classification device 20, a conversion dictionary 31, and a database 32. The classifier generation device 10 generates a classifier. Text data (input text data) is input to the content classification device 20 as input data. The content classification device 20 classifies the input text data using the classifier generated by the classifier generation device 10. The content classification device 20 outputs classification result data indicating a result of classification of input text data.

The conversion dictionary 31 is an electronic dictionary. The conversion dictionary 31 has word pair data. The conversion dictionary 31 outputs a pair of words (converted words) for an input word (input word). In this embodiment, an antonym dictionary is provided as an example of the conversion dictionary 31. The conversion dictionary 31 as an antonym dictionary outputs an antonym of the input word as a converted word.

The database 32 stores classification structure data and teacher data. The classification structure data is data indicating a classification item and a hierarchy of classification items which are the destinations of classification of text data (sentence data). The classification structure data is created in advance and stored in the database 32. To create the classification structure data, for example, the Grounded Theory Approach technology can be used.

The teacher data is text data to which a label for identifying the classification item of the classification structure data and the hierarchy of the classification item is added. The original text data of the teacher data is, for example, text data of a sentence (free description sentence) freely described as a response to a questionnaire conducted on the general public. The data to which the corresponding classification item of the classification structure data and the label for identifying the hierarchy of the classification item are added to the text data is used as teacher data. For example, a person analyzes the text data to determine the label to be attached to the text data, and attaches the label of the determination result to the text data.

The classifier generating device 10 includes a classifier generating unit 11, a sentence changing unit 12, a representative sentence generating unit 13, and a classifier correcting unit 14. The classifier generation unit 11 has a classifier generation function of generating a classifier. The classifier generation unit 11 generates a classifier using the classification structure data and the teacher data. The sentence changing unit 12 has a sentence changing function of changing the meaning of a sentence. The sentence changing unit 12 changes the meaning of a sentence by replacing words or adding or deleting words. The representative sentence generation unit 13 has a representative sentence generation function of generating a representative sentence of the classifier. The representative sentence generation unit 13 uses the classifier generated by the classifier generation unit 11 to classify the classification target text data, and generates a representative sentence of the classifier based on the result of the classification. The classifier correction unit 14 has a classifier correction function for correcting the classifier. The classifier correction unit 14 classifies the classification target data using the classifier generated by the classifier generation unit 11, and determines whether to correct the classifier based on the result of the classification.

The classifier generation device 10 may be realized by dedicated hardware, or may be a general-purpose computer device such as a personal computer, and is a computer program for realizing the functions of the classifier generation device 10. The function may be realized by executing. The content classification device 20 may be realized by dedicated hardware, or may be configured by a general-purpose computer device such as a personal computer and executes a computer program for realizing the function of the content classification device 20. The function may be realized by doing so.

The classifier generation device 10 and the content classification device 20 may be configured as separate devices, or may be configured as one device. For example, in the classifier generation device 10 and the content classification device 20, the functions of the respective devices 10 and 20 may be realized by computer programs by separate computer devices, or both devices 10 and 10 may be realized by one computer device. The 20 functions may be implemented by a computer program.

The classifier generation device 10 accesses the conversion dictionary 31. The conversion dictionary 31 may be provided inside the classifier generation device 10, or may be provided outside the classifier generation device 10 and accessed from the classifier generation device 10 by communication.

The classifier generation device 10 accesses the database 32. The database 32 may be provided inside the classifier generation device 10 or may be provided outside the classifier generation device 10 and accessed from the classifier generation device 10 by communication.

Next, the outline of the classifier generation method according to this embodiment will be described with reference to FIG. Generally, in the questionnaire survey, there are 2 points of good and bad points such as "Question a: Please tell us the good points of XX." and "Question b: Please tell us the bad points of XX." I may ask you two questions. When the answers to these two questions a and b are classified using machine learning, it is necessary to generate a classifier for the question a and a classifier for the question b. Therefore, the classification structure data and the teacher data used to generate the classifier for the question a and the classification structure data and the teacher data used to generate the classifier for the question b are prepared.

FIG. 2 is an explanatory diagram of the outline of the classifier generation method according to the present embodiment. In the example of FIG. 2, in a questionnaire survey on a communication service, a question A to ask unsatisfactory points about the service and a question B to ask unsatisfactory points about the service are performed. Then, in order to classify each answer to these two questions A and B using machine learning, a classifier A3 for question A and a classifier B3 for question B are generated respectively. For this classifier generation, the classification structure data A1 and the teacher data A2 used for generating the classifier of the question A, and the classification structure data B1 and the teacher data B2 used for generating the classifier of the question B are generated. prepare. When the classification structure data A1 and the teacher data A2, and the classification structure data B1 and the teacher data B2 are manually generated, a burden such as an increase in processing time and an increase in cost becomes a problem.

In view of such a problem, in the present embodiment, the processing time is increased and the cost is increased by automatically generating the other classification structure data and the teacher data from the one classification structure data and the teacher data prepared in advance. Reduce the burden of increase. In the example of FIG. 2, one of the classification structure data A1 and the teacher data A2 is prepared in advance by, for example, manually generating it. However, the other classification structure data B1 and teacher data B2 are automatically generated from the classification structure data A1 and teacher data A2. This is because, for example, when asking questions of the same respondent such as question A and question B, which are semantically opposite, similar classification structure data can be considered.

In FIG. 2, the classification structure data A1 has three classification items “communication quality dissatisfaction”, “terminal dissatisfaction”, and “after-satisfaction” as classification items of the first layer. In addition, the classification structure data A1 includes three classification items “a battery is poor”, “a screen is small”, and “a storage capacity is low as a classification item of a second layer which is a lower layer of the classification item “terminal dissatisfaction” of the first layer. Have a "small". The sentence of each classification item is a representative sentence. The representative sentence is a sentence provided for a person to understand the classification item. The classification structure data A1 has a representative sentence of each classification item. The classification structure data A1 has link information for associating corresponding classification items in each hierarchy. For example, the classification structure data A1 is link information associating the classification items of the first hierarchy "terminal dissatisfaction" and the classification items of the second hierarchy "bad battery life", "small screen" and "small storage capacity" with each other. Have.

In the present embodiment, the classification structure data A1 is converted into the classification structure data B1 by changing the sentence so that the meaning of each representative sentence of the classification structure data A1 prepared in advance is changed. In the example of FIG. 2, the word of the representative sentence is replaced with the opposite word to change the meaning of the sentence. For example, the word "dissatisfaction" in the representative sentence of the classification structure data A1 is changed to the opposite word "satisfaction" to make the representative sentence of the classification structure data B1. Further, the word "bad" in the representative sentence of the classification structure data A1 is changed to the opposite word "good" to make it the representative sentence of the classification structure data B1. As described above, in the present embodiment, the classification structure data A1 is converted into the classification structure data B1 by semantically changing each representative sentence of the classification structure data A1.

The classification item of the classification structure data A1 and the hierarchy of the classification item correspond to the classification item of the classification structure data B1 and the hierarchy of the classification item, respectively. The sentence in which the representative sentence of the m-th classification item in the n-th hierarchy of the classification structure data A1 is semantically changed becomes the representative sentence of the m-th classification item in the same n-th hierarchy of the classification structure data B1.

In FIG. 2, the teacher data A2 has a set of a free description sentence and a classification result of the free description sentence by the classification structure data A1. For example, the teacher data A2 has a set of free description sentences "the battery does not last long" and the classification result "the battery is poor". This indicates that the free description “Battery does not last long” is classified into the classification item “Battery lasts poor” in the classification structure data A1. The teacher data A2 is prepared in advance. For example, it is manually determined to which classification item of the classification structure data A1 each free descriptive sentence of the teaching data A2 applies, and the classification item of the judgment result and the label for identifying the hierarchy of the classification item are assigned to the corresponding freedom item of the teaching data A2. Add to the description.

In the present embodiment, the teacher data A2 is converted into the teacher data B2 by changing the sentence so that the meaning of each free description sentence of the teacher data A2 prepared in advance is changed. In the example of FIG. 2, the meaning of the sentence is changed by replacing the word in the free description sentence with the opposite word. For example, the free description sentence "battery lasts" of the teacher data B2 is changed by changing the word "not" in the free description sentence "battery does not last long" of the teacher data A2 to the opposite word "yes". As described above, in this embodiment, the teacher data A2 is converted into the teacher data B2 by semantically changing each free description sentence of the teacher data A2. In the example of FIG. 2, the semantic change of the free description sentence of the teacher data A2 is replaced with the same opposite word as the semantic change of the representative sentence of the classification structure data A1.

The label given to the free description sentence of the teacher data A2 and the label given to the free description sentence of the teacher data B2 correspond to each other. The sentence in which the nth free description sentence of the teacher data A2 is semantically changed becomes a free description sentence to which the same label as the nth free description sentence of the teacher data A2 is given in the teacher data B2.

In FIG. 2, a classifier A3 of the question A that asks a dissatisfaction point for a service (may be referred to as a dissatisfaction classifier A3 for convenience of description) is generated using the classification structure data A1 and the teacher data A2. To be done. The classifier B3 of the question B that asks satisfaction of the service (may be referred to as a satisfaction classifier B3 for convenience of description) is generated using the classification structure data B1 and the teacher data B2.

FIG. 3 is an explanatory diagram of the classifier according to this embodiment. In the example of FIG. 3, it is an explanatory diagram of the satisfaction classifier B3. It can be said that the classifier is a set of label assigning rules for assigning a label to text data to which no label is attached (text data without label). The classifier attaches a label to the input unlabeled text data according to its own labeling rule, and outputs the text data with the label attached to the unlabeled text data (labeled text data).

In FIG. 3, unsatisfied classifier B3 is input with unlabeled text data "The smile of the clerk was nice". The satisfaction classifier B3 attaches a label corresponding to the representative sentence "the store staff has a good customer service attitude" to the unlabeled text data "The smile of the store clerk was good" according to the self-labeling rule. The satisfaction classifier B3 outputs the text data “the smile of the clerk was good” with the label as the classification result data. This classification result data indicates that the text data "The smile of the store clerk was good" was classified into the category item of the representative sentence "The store staff has a good customer service".

Next, the classifier generation device 10 shown in FIG. 1 will be described.

[Sentence change part]
The sentence changing unit 12 changes the meaning of a sentence by replacing words or adding or deleting words. The sentence changing unit 12 performs morphological analysis and dependency analysis on the input sentence. For the morphological analysis, for example, software called "Mecab" may be used. For the dependency analysis, for example, software called "CaboCha" may be used.

(Example 1 of sentence change method)
An example 1 of the sentence changing method will be described. Example 1 of the sentence modification method is a case where a predicate term structure is found in an input sentence as a result of morphological analysis and dependency analysis. As a result of the morphological analysis and the dependency analysis, the sentence modification unit 12 modifies the word of the predicate when the input sentence has a predicate term structure. In changing the word of the predicate, the sentence changing unit 12 passes the word of the predicate to the conversion dictionary 31. The conversion dictionary 31 returns a pair of words (converted word) for the word of the predicate (input word) to the sentence changing unit 12. In this embodiment, the conversion dictionary 31 is an antonym dictionary. Therefore, the sentence changing unit 12 receives from the conversion dictionary 31 an antonym of the word of the predicate in the input sentence. The sentence changing unit 12 replaces the word of the predicate in the input sentence with the opposite word received from the conversion dictionary 31. The sentence changing unit 12 outputs a sentence (change sentence) in which the word in the predicate of the input sentence is replaced with the opposite word. As an example of Example 1 of the sentence changing method, the sentence changing unit 12 replaces the word “slow” in the predicate with the opposite word “fast” for the input sentence “slow”, and changes the sentence “fast”. Is output.

(Example 2 of sentence change method)
An example 2 of the sentence changing method will be described. Example 2 of the sentence changing method is a case where the conversion dictionary 31 does not have a word (converted word) paired with the input word in the example 1 of the sentence changing method described above. As a result of the morphological analysis and the dependency analysis, the sentence modification unit 12 modifies the word of the predicate when the input sentence has a predicate term structure. In changing the word of the predicate, the sentence changing unit 12 passes the word of the predicate to the conversion dictionary (antonym dictionary) 31. The conversion dictionary 31 replies to the sentence changing unit 12 that there is no word (conversion word) paired with the word of the predicate (input word) (no conversion word). The sentence changing unit 12 adds a negative word to the word of the predicate in the input sentence or deletes the negative word from the word of the predicate in the input sentence in response to a response from the conversion dictionary 31 that there is no converted word. To do. The sentence changing unit 12 adds the negative word to the word of the predicate in the input sentence, while adding the negative word to the word of the predicate in the input sentence. Delete negative words. The sentence changing unit 12 outputs a sentence (change sentence) in which a negative word is added to the word of the predicate of the input sentence or the negative word is deleted from the word of the predicate of the input sentence. As an example of Example 2 of the sentence changing method, the sentence changing unit 12 deletes the negative word “no” attached to the word “phone is not connected” in the predicate from the input sentence “phone is not connected”. , The change sentence "A phone is connected" prepared as an affirmative sentence is output.

(Example 3 of sentence change method)
Example 3 of the sentence changing method will be described. Example 3 of the sentence modification method is a case where the predicate-argument structure is not found in the input sentence as a result of the morpheme analysis and the dependency analysis (when it is a word stop). As a result of the morpheme analysis and the dependency analysis, the sentence changing unit 12 changes the word of the word stop (noun of the final phrase) in the input sentence when the predicate term structure is not found in the input sentence. In changing the word of the word stop, the sentence changing unit 12 passes the word of the word stop to the conversion dictionary (opposite word dictionary) 31. The conversion dictionary 31 returns the opposite word, which is a word (converted word) paired with the word (input word) of the predicate, to the sentence changing unit 12. The sentence changing unit 12 replaces the word of word stop in the input sentence with the opposite word received from the conversion dictionary 31. The sentence changing unit 12 outputs a sentence (change sentence) in which the word of word stop of the input sentence is replaced with the opposite word. As an example of Example 3 of the sentence changing method, the sentence changing unit 12 replaces the input sentence “Communication quality dissatisfaction” with the word “Dissatisfaction”, which is a word stop, by the opposite word “Satisfaction”, and changes sentence “Communication quality satisfaction”. Is output. As another example of Example 3 of the sentence changing method, the sentence changing unit 12 replaces the word “increase” of the wording stop with the opposite word “decrease” for the input sentence “increase in price”, and changes the sentence “price”. Is output.

[Classifier generator]
The classifier generation unit 11 generates a classifier using the classification structure data and the teacher data. Examples of the classifier include a classifier such as SVM (Support Vector Machine) or "Bayesian Network". The classifier generation unit 11 will be described with reference to FIG. In the database 32, the classification structure data A1 and the teacher data A2 used for generating the classifier for the question A are stored in advance. The classifier generation unit 11 acquires the classification structure data A1 and the teacher data A2 from the database 32.

The classifier generation unit 11 uses the classification structure data A1 and the teacher data A2 to generate a classifier (dissatisfied classifier) A3 of the question A.

A method of generating the classifier (satisfaction classifier) B3 of the question B will be described. The classifier generation unit 11 changes each representative sentence included in the classification structure data A1 by the sentence changing unit 12 to generate the classification structure data B1. In FIG. 4A, the classifier generation unit 11 passes each representative sentence included in the classification structure data A1 to the sentence changing unit 12, and receives the changed sentence for each representative sentence from the sentence changing unit 12. The classifier generation unit 11 replaces each representative sentence of the classification structure data A1 with each modified sentence received from the sentence modification unit 12. The classification structure data after this replacement is used as the classification structure data B1. As the link information of the classification structure data B1, the link information of the classification structure data A1 is used as it is.

The classifier generation unit 11 changes each free description sentence included in the teacher data A2 by the sentence change unit 12 to generate teacher data B2. In FIG. 4(2), the classifier generation unit 11 passes each free description sentence included in the teacher data A2 to the sentence modification unit 12, and receives the modified sentence for each free description sentence from the sentence modification unit 12. The classifier generation unit 11 gives each modified sentence received from the sentence modification unit 12 the same label as each original free description sentence of the teacher data A2. The modified sentence group after the label is assigned is used as the teacher data B2.

The classifier generation unit 11 uses the classification structure data B1 and the teacher data B2 to generate a classifier (satisfaction classifier) B3 for the question B.

In the present embodiment, the classification structure data A1 corresponds to the first classification structure data, and the teacher data A2 corresponds to the first teacher data. The classification structure data B1 corresponds to the second classification structure data, and the teacher data B2 corresponds to the second teacher data.

[Representative sentence generator]
The representative sentence generation unit 13 uses the classifier generated by the classifier generation unit 11 to classify the classification target text data, and generates a representative sentence of the classifier based on the result of the classification. The representative sentence generator 13 will be described with reference to FIG. The database 32 stores the satisfaction classification target data B4 in advance. The satisfaction classification target data B4 is text data of an answer to the question B that asks about satisfaction with the service. No label is given to the satisfaction classification target data B4. The representative sentence generation unit 13 acquires the satisfaction classification target data B4 from the database 32. The representative sentence generation unit 13 uses the satisfaction classifier B3 generated by the classifier generation unit 11 to classify the satisfaction classification target data B4. The satisfaction classifier B3 outputs satisfaction classification result data B5 in which a label is added to the input satisfaction classification target data B4. Using the satisfaction classifier B3, the representative sentence generation unit 13 acquires a plurality of satisfaction classification result data B5 (satisfaction classification result data group) from a plurality of satisfaction classification target data B4 (satisfaction classification target data group).

The representative sentence generation unit 13 generates a representative sentence of the satisfaction classifier B3 based on the satisfaction classification result data group. In the generation of this representative sentence, the representative sentence generation unit 13 extracts a phrase (important phrase) that appears prominently in the satisfaction classification result data group. An information criterion called AIC (Akaike's Information Criterion), for example, may be used to extract the important terms. The representative sentence generation unit 13 extracts a phrase having a high co-occurrence frequency with the important phrase in the satisfaction classification result data group. In the satisfaction classification result data group, the representative sentence generation unit 13 selects, from among the sentences classified into one classification item, a sentence including both an important phrase and a phrase having a high co-occurrence frequency with the important phrase. .. When there are a plurality of selected sentences, the representative sentence generation unit 13 further selects the shortest sentence from the selected sentences. The representative sentence generation unit 13 determines the sentence resulting from the selection as the representative sentence of the classification item.

The representative sentence generation unit 13 passes the representative sentence of the determined classification item to the classifier generation unit 11. The classifier generation unit 11 replaces the representative sentence of the classification item received from the representative sentence generation unit 13 with the representative sentence of the classification item of the satisfaction classifier B3. As a result, the representative sentence of the satisfaction classifier B3 is replaced with the representative sentence generated by the representative sentence generation unit 13.

According to this embodiment, the representative sentence of the classification item of the satisfaction classifier B3 can be generated from the sentence included in the satisfaction classification target data B4. Since the sentence included in the satisfaction classification target data B4 is the sentence of the person who answers the question B, it is considered that the expression of the sentence is natural. Therefore, by generating the representative sentences of the classification items of the satisfaction classifier B3 from the sentences included in the satisfaction classification target data B4, the representative sentence corresponding to the classification result data of the satisfaction classifier B3 is easy for a person to understand. Can be a natural sentence. This is because, in the present embodiment, the representative sentence of the classification structure data B1 used for generating the satisfaction classifier B3 is changed from the representative sentence of the classification structure data A1 used for generating the dissatisfaction classifier A3 by the sentence changing unit 12. The special effect of solving the unnaturalness of the representation of the representative sentence due to the change and generation.

[Classifier corrector]
The classifier correction unit 14 classifies the classification target data using the classifier generated by the classifier generation unit 11, and determines whether to correct the classifier based on the result of the classification. The classifier correction unit 14 will be described with reference to FIG. If the appearance tendency of the answer differs between question A that asks unsatisfied points for the service and question B that asks unsatisfied points about the service, the classification structure of the classification structure data A1 related to question A is As it is, there is a possibility that the classification structure of question B may not be met.

In the example of FIG. 6, the amount and type of answers are large for the second-level classification item “store correspondence”, which is a lower layer of the first-level classification item “after-satisfaction” of the classification structure data A1 related to question A. On the other hand, with respect to the classification item "corresponding to over-the-counter" of the second layer, which is a lower layer of the classification item "after-sales dissatisfaction" of the same first layer of the classification structure data B1 related to the question B, the amount and type of answers are small. Therefore, if the classification structure data B1 is used as it is, although the amount and type of the reply of the classification item “corresponding to store” are small, the classification of the subordinate subdivision of the classification item “corresponding to store” is small. By using the structure, the classification accuracy may decrease. In order to solve such a problem, in the present embodiment, the classifier correction unit 14 corrects the classifier generated by the classifier generation unit 11.

The classifier corrector 14 uses the satisfaction classifier B3 generated by the classifier generator 11 to calculate a plurality of satisfaction classification result data B5 (satisfaction classification) from a plurality of satisfaction classification target data B4 (satisfaction classification target data group). (Result data group) is acquired. The classifier correction unit 14 calculates the appearance frequency of each classification item (label) in the satisfaction classification result data group. When the difference between the appearance frequencies of the respective classification items is equal to or more than a predetermined value, the classifier correction unit 14 determines to correct the satisfaction classifier B3, and the difference between the appearance frequencies of the respective classification items is If it is less than the predetermined value, it is determined that the satisfaction classifier B3 is not corrected. When determining that the satisfaction classifier B3 is to be corrected, the classifier correction unit 14 corrects the satisfaction classifier B3 by a predetermined classifier correction method.

Next, an example of the classifier generation method according to this embodiment will be described.

[Example 1 of classifier generation method]
An example 1 of the classifier generation method according to this embodiment will be described with reference to FIG. 7. FIG. 7 is an explanatory diagram showing the procedure of Example 1 of the classifier generation method according to the present embodiment.

(Step S11) The classifier generation unit 11 changes each representative sentence included in the dissatisfied classification structure data A1 by the sentence modification unit 12 to generate satisfaction classification structure data B1. The classification structure data related to the question A that asks unsatisfied points about the service is referred to as dissatisfaction classification structure data A1. The classification structure data related to the question B for asking satisfaction about the service is referred to as satisfaction classification structure data B1.

The classifier generation unit 11 changes each free description sentence included in the dissatisfied teacher data A2 by the sentence changing unit 12 to generate the satisfied teacher data B2. The teacher data related to the question A that asks unsatisfied points about the service is referred to as dissatisfied teacher data A2. The teacher data relating to the question B for asking satisfaction about the service is referred to as satisfaction teacher data B2. A label is attached to each free description sentence in the dissatisfied teacher data A2 and the satisfied teacher data B2.

(Step S12) The classifier generation unit 11 uses the satisfaction classification structure data B1 and the satisfaction teacher data B2 to generate a satisfaction classifier B3 for the question B.

[Example 2 of classifier generation method]
Example 2 of the classifier generation method according to this embodiment will be described with reference to FIG. 8. FIG. 8 is an explanatory diagram showing the procedure of Example 2 of the classifier generation method according to this embodiment. 8, parts corresponding to the respective steps in FIG. 7 are designated by the same reference numerals, and the description thereof will be omitted. Steps S11 and S12 are executed to generate the satisfaction classifier B3. Then, step S21 is executed.

(Step S21) The representative sentence generation unit 13 uses the satisfaction classifier B3 to extract a plurality of satisfaction classification result data B5 (satisfaction classification result data group) from the plurality of satisfaction classification target data B4 (satisfaction classification target data group). To get

(Step S22) The representative sentence generation unit 13 generates a representative sentence of the satisfaction classifier B3 based on the satisfaction classification result data group. The classifier generation unit 11 replaces the representative sentence of the satisfaction classifier B3 with the representative sentence generated by the representative sentence generation unit 13 to generate the satisfaction classifier B3a with the changed representative sentence.

[Example 3 of classifier generation method]
Example 3 of the classifier generation method according to this embodiment will be described with reference to FIG. 9. FIG. 9 is an explanatory diagram showing the procedure of Example 3 of the classifier generation method according to the present embodiment. In FIG. 9, parts corresponding to the steps in FIGS. 7 and 8 are given the same reference numerals, and description thereof will be omitted. Steps S11 and S12 are executed to generate the satisfaction classifier B3. Then, step S31 is executed.

(Step S31) The classifier correcting unit 14 uses the satisfaction classifier B3 to convert a plurality of satisfaction classification target data B4 (satisfaction classification target data group) into a plurality of satisfaction classification result data B5 (satisfaction classification result data group). To get

(Step S32) The classifier correction unit 14 calculates the appearance frequency of each classification item (label) in the satisfaction classification result data group. When the difference between the appearance frequencies of the respective classification items is equal to or more than a predetermined value, the classifier correction unit 14 determines to correct the satisfaction classifier B3, and the difference between the appearance frequencies of the respective classification items is If it is less than the predetermined value, it is determined that the satisfaction classifier B3 is not corrected. When determining that the satisfaction classifier B3 is to be corrected, the classifier correction unit 14 corrects the satisfaction classifier B3 by a predetermined correction method to generate the satisfaction classifier B3b.

Next, a modified example according to the present embodiment will be described.

[Modification 1]
FIG. 10 is an explanatory diagram of Modification 1 according to the present embodiment. In the first modification, the classifier generating unit 11 changes the sentence included in the classifier by the sentence changing unit 12 to generate another classifier. In FIG. 10, the classifier generation unit 11 changes each sentence included in the dissatisfied classifier A3 by the sentence changing unit 12. The classifier generation unit 11 sets the dissatisfied classifier A3 after the change of the sentence as a satisfaction classifier B3. In the first modification, the dissatisfaction classifier A3 corresponds to the first classifier, and the satisfaction classifier B3 corresponds to the second classifier.

[Modification 2]
Modification 2 is a modification of the sentence changing method. When a sentence has a subject and a predicate, it may be preferable to select a predicate to change to depending on the subject. Therefore, in the second modification, the predicate to be changed is selected according to the subject of the sentence. In the second modification, the conversion dictionary 31 has word pair data associated with a word that can be a subject. When the predicate term structure and the subject are found in the input sentence as a result of the morphological analysis and the dependency analysis, the sentence changing unit 12 passes the word of the subject and the word of the predicate to the conversion dictionary 31. The conversion dictionary 31 uses the data of the pair of words associated with the word of the subject to return the pair of words (converted word) for the word (input word) of the predicate to the sentence changing unit 12. .. The sentence changing unit 12 replaces the word of the predicate in the input sentence with the converted word received from the conversion dictionary 31. As a result, the sentence changing unit 12 can appropriately change the predicate according to the subject in the input sentence.
An application example of the modification 2 will be described. Even if different services X and Y are used, if the users have similar interests in the services, the structure of the classification items of the classification structure data and the hierarchy of the classification items can be the same. However, since the service X and the service Y are different services, the representative sentence used in the classification structure data may be expressed differently between the service X and the service Y. For example, when the subject of the representative sentence is different between the service X and the service Y, it may be more natural that the words used in the predicate are different even if the expressions of the same satisfaction and dissatisfaction relationships are expressed. In such a case, according to the second modification, an appropriate predicate can be changed according to the subject, and therefore the subject of the representative sentence of the service X is changed to the subject suitable for the service Y, and then the sentence changing unit 12 is used. By changing the predicate, the representative sentence of the service X can be changed to the representative sentence suitable for the service Y. As a result, it is possible to generate the classification structure data and the teacher data regarding the service Y from the classification structure data and the teacher data regarding the service X for different services X and Y.

Next, an example of the classifier correction method according to this embodiment will be described.
[Example of classifier correction method]
An example of the classifier correction method according to the present embodiment will be described with reference to FIGS. 11, 12, and 13. FIG. 11 is a diagram showing a configuration example of the classification structure data D1 according to this embodiment. The classification structure data D1 is classification structure data of the classifier D. In FIG. 11, the classification structure data D1 has three layers from the first layer to the third layer, and has a classification item for each layer.

In the classification structure data D1, the classification item of the first layer is “music”, and the identification number of the classification item “music” of the first layer is “1”. The category items of the second layer, which is a lower layer belonging to the category item "Music" of the first layer, are "genre" and "timbre", and the identification number of the category item "genre" of the second layer is "1". The identification number of the classification item "timbre" of the second layer is "2". The classification items of the third layer, which is a lower layer belonging to the classification item "genre" of the second layer, are "classic" and "pops", and the identification number of the classification item "classic" of the third layer is "1". The identification number of the classification item "pops" in the third hierarchy is "2". The classification items of the third layer, which is a lower layer belonging to the classification item of the second layer "timbre", are "about", "bright", and "warm", and the identification number of the classification item "about" of the third layer is "about". 1", the identification number of the classification item "bright" of the third layer is "2", and the identification number of the classification item "warm" of the third layer is "3".

The classification structure data D1 is used, for example, when classifying text data by music genre or impression. In the classification structure data D1, the first layer is a destination where the text data whose topic is “music” is classified from the text data to be classified. The second layer is a destination for classifying text data relating to "genre" or "timbre" among the text data whose topic is "music". The third layer is a text data whose topic is “music” and is a destination for classifying text data related to a specific classification item of “genre” or “timbre” among text data related to “genre” or “timbre”. Become.

The text data classified by the classifier D based on the classification structure data D1 is provided with a label including the identification number of the classified classification item. FIG. 11 shows labels “111”, “112”, “121”, “122”, and “123” that are given to the text data classified up to the third layer. For example, the classification result label “111” is added to the text data “My favorite music is classical music”. The classification result label “123” is given to the text data “I like music with a warm tone”. Further, as the text data classified up to the second layer, for example, “I like any genre of music”, the classification result label “11” is given. As the text data classified up to the first layer, for example, "I like music" is given a classification result label "1".

The classifier correction unit 14 corrects the classification structure data D1 of the classifier D. The operation related to the classifier correction of the classifier correction unit 14 will be described with reference to FIG. FIG. 12 is a flowchart showing the procedure of an example of the classifier correction method according to this embodiment. The procedure (steps S111 to S114) of the classifier correction method illustrated in FIG. 12 is sequentially executed from the upper hierarchy to the lower hierarchy in the classification structure data. With respect to the classification structure data D1, steps S111 to S114 are first performed on the first layer, steps S111 to S114 are then performed on the second layer, and finally steps S111 to S111 on the third layer. Execute S114.

The verification teacher data D2 is used to correct the classification structure data D1. The verification teacher data D2 is, for example, the correct answer of the labels belonging to the classification structure data D1 with respect to the text data of the sentence freely described as the answer to the questionnaire conducted in the most recent fixed period (for example, the most recent half year). It is text data with a label. For example, a person analyzes the text data to determine the correct answer label of the text data, and gives the correct answer label of the determination result to the text data. The verification teacher data is stored in the database 32 in advance.

(Step S111) The classifier correction unit 14 sorts the verification teacher data D2 with respect to the verification target classification item among the verification target hierarchy classification items in the classification structure data D1. In the classification of the verification teacher data D2, the verification teacher data D2a to which the correct answer label including the identification number of the verification target classification item is assigned to the verification target classification item, and the correct answer label that does not include the identification number of the verification target classification item. It is sorted into the assigned verification teacher data D2b. For example, when the classification item “genre” of the second hierarchy of the classification structure data D1 is the verification target classification item, the label “11”, “111” or “112” including the identification number “1” of the classification item “genre”. ", and the label "1", "12", "121", "122" or "123" which does not include the identification number "1" of the classification item "genre" is added. It is sorted into the verification teacher data D2b. The verification teacher data D2a to which the correct answer label including the identification number of the verification target classification item is added is grouped into the verification positive example data group of the verification target classification item. The verification teacher data D2b to which the correct label that does not include the identification number of the verification target classification item is added is grouped into the verification negative example data group of the verification target classification item.

The classifier correction unit 14 sorts the verification teacher data D2a in the verification positive example data group of the verification target classification item into verification learning data and verification test data at a predetermined ratio. The classifier correction unit 14 also classifies the verification teacher data D2b in the verification negative example data group of the verification target classification item into verification learning data and verification test data at the same ratio. For example, 90% of the verification teacher data D2a of all verification teacher data D2a in the verification positive example data group of the verification target classification item is grouped into the verification positive example learning data group of the verification target classification item, and the remaining 10% The verification teacher data D2a is grouped into the verification positive example test data group of the verification target classification item. Similarly, of all verification teacher data D2b in the verification negative example data group of the verification target classification item, 90% of the verification teacher data D2b is grouped into the verification negative example learning data group of the verification target classification item, and the remaining 10%. The verification teacher data D2b is grouped into the verification negative example test data group of the verification target classification item.

(Step S112) The classifier correction unit 14 uses the verification positive example learning data group and the verification negative example learning data group of the verification target classification item to generate a classifier for the verification target classification item. The type of classifier generated by the classifier correction unit 14 is the same type as the classifier generated by the classifier generation unit 11. For example, the classifier generation unit 11 and the classifier correction unit 14 generate an SVM classifier. The classifier for the verification target classification item has a function of determining whether or not the text data of the classification target corresponds to the verification target classification item.

(Step S113) The classifier correction unit 14 uses the verification positive example test data group or the verification negative example test data group of the verification target classification item to calculate the suitability of the classifier for the verification target classification item. Examples 1 to 4 of the suitability of the classifier are given below.

(Example 1 of goodness of fit of classifier)
Example 1 of the goodness of fit of the classifier is the accuracy rate (Accuracy). The accuracy rate is expressed by the following equation.
Correct rate = (TP+TN)/(TP+FP+TN+FN)

(Example 2 of goodness of fit of classifier)
Example 2 of the goodness of fit of the classifier is the true positive rate. The true positive rate is expressed by the following equation.
True positive rate = TP/(TP+FN)

(Example 3 of goodness of fit of classifier)
Example 3 of the goodness of fit of the classifier is a false negative rate. The false negative rate is expressed by the following equation.
False negative rate = FN/(TP+FN)

(Example 4 of the goodness of fit of the classifier)
Example 4 of the goodness of fit of the classifier is precision. The accuracy is expressed by the following equation.
Accuracy = TP/(TP+FP)

In Examples 1 to 4 of the goodness of fit of the classifier described above, TP (True positives) is the result of applying the verification teacher data of the verification positive example test data group to the classifier It is the number. TN (True negatives) is the number of pieces of verification teacher data correctly determined as a negative example as a result of applying the verification teacher data of the verification negative example test data group to the classifier. FP (False positives) is the number of pieces of verification teacher data erroneously determined to be positive as a result of applying the verification teacher data of the verification negative example test data group to the classifier. FN (False negatives) is the number of pieces of verification teacher data erroneously determined to be negative as a result of applying the verification teacher data of the verification positive example test data group to the classifier.

Of the examples 1 to 4 of the goodness of fit of the classifier, only one of them may be used as the goodness of fit of the classifier, or a plurality may be used as the goodness of fit of the classifier. ..

(Step S114) The classifier correction unit 14 determines whether the classifier is acceptable or not based on the suitability of the classifier for the verification target classification item. For example, one or more of the examples 1 to 4 of the goodness of fit of the classifier are evaluated, and if the goodness of fit is good as a result of the evaluation, it is determined as pass. For example, if at least one of the examples 1 to 4 of the goodness of fit of the classifier is good, it may be determined as a pass. Alternatively, it may be determined as a pass only when all of the examples 1 to 4 of the goodness of fit of the classifier are good. As an evaluation method of the goodness of fit, for example, comparison of the goodness of fit with a predetermined threshold value can be mentioned.

The classifier correcting unit 14 adopts the verification target classification item when the classifier for the verification target classification item is acceptable. On the other hand, when the classifier for the verification target classification item fails, the classifier correction unit 14 rejects the verification target classification item.

Note that the steps S111 to S114 may not be executed for the lower classification items that belong to the verification target classification item that has been rejected, and may be rejected.

The classifier correction unit 14 deletes the verification target classification item that is not adopted from the classification structure data D1 or sets it as an undeterminable item of the classification structure data D1.

FIG. 13 is a diagram showing a correction example for the classification structure data D1 according to the present embodiment. FIG. 13 shows an example of a result obtained by correcting the classification structure data D1 shown in FIG. 11 by the classifier correction unit 14. In the example of FIG. 13, the classifier correction unit 14 determines that the classification item “timbre” of the second layer is not adopted. Further, since the classification item "timbre" of the second layer is determined not to be adopted, the classification items "about", "bright", and "warm" of the third layer which is a lower layer belonging to the classification item "timbre" of the second layer. Was decided not to be adopted. As a result, in the classification structure data D1 of the classifier D, the classification item “timbre” of the second hierarchy and the classification items “about”, “bright” and “warm” of the third hierarchy are deleted, or the second hierarchy. The classification item "tone" and the classification items "about", "bright", and "warm" of the third hierarchy are set as undecidable items.
The above is the description of the example of the classifier correction method according to the present embodiment.

According to the above-described embodiment, other classification structure data can be automatically generated from the classification structure data prepared in advance. This has the effect of improving the efficiency of the text classification process.

According to the above-described embodiment, other teacher data can be automatically generated from the teacher data prepared in advance. This has the effect of improving the efficiency of the text classification process.

Although both the classification structure data and the teacher data are automatically generated in the above-described embodiment, only one of the classification structure data and the teacher data may be automatically generated.

Although the embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and includes design changes and the like without departing from the scope of the present invention.
For example, in the above-described embodiment, the opposite dictionary is provided as the conversion dictionary 31, but the conversion dictionary 31 may be appropriately changed.

Further, a computer program for realizing the functions of the above-described devices may be recorded in a computer-readable recording medium, and the program recorded in this recording medium may be read by a computer system and executed. The “computer system” may include an OS and hardware such as peripheral devices.
The "computer-readable recording medium" is a writable non-volatile memory such as a flexible disk, a magneto-optical disk, a ROM, a flash memory, a portable medium such as a DVD (Digital Versatile Disc), or a built-in computer system. A storage device such as a hard disk.

Further, the "computer-readable recording medium" means a volatile memory (for example, a DRAM (Dynamic) in a computer system that serves as a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. Random Access Memory)), which also holds the program for a certain period of time.
Further, the program may be transmitted from a computer system that stores the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the "transmission medium" for transmitting the program means a medium having a function of transmitting information such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line.
Further, the program may be a program for realizing some of the functions described above. Further, it may be a so-called difference file (difference program) that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

DESCRIPTION OF SYMBOLS 1... Classification system, 10... Classifier generation device, 11... Classifier generation unit, 12... Sentence change unit, 13... Representative sentence generation unit, 14... Classifier correction unit, 20... Content classification device, 31... Conversion dictionary, 32... Database

Claims (8)

A classifier generation device for generating a classifier for classifying each answer to a semantically opposite question obtained from the same respondent,
A sentence change section that changes the meaning of a sentence by replacing words or adding or deleting words,
The sentence included in the first classification structure data indicating the classification item and the hierarchy of the classification item that is the destination of classifying the text data is changed by the sentence changing unit to generate the second classification structure data, and the second classification structure data is generated. And a classifier generation unit that generates a classifier using the classification item of the second classification structure data and the second teacher data to which the label for identifying the hierarchy of the classification item is added ,
The sentence modification unit performs a morphological analysis and a dependency analysis on the input sentence. As a result of the morphological analysis and the dependency analysis, when the input sentence has a predicate term structure, the word opposite to the predicate term is detected. Acquiring from an antonym dictionary and replacing the word of the predicate in the input sentence with the acquired antonym.
Classifier generator. When the input sentence has a predicate-argument structure in the input sentence, when the opposite word of the word of the predicate cannot be obtained from the opposite word dictionary, the word of the predicate in the input sentence is a negative word. , Or delete the negative word from the word of the predicate in the input sentence,
The classifier generation device according to claim 1. As a result of the morpheme analysis and the dependency analysis on the input sentence, the sentence changing unit, when a predicate-argument structure is not found in the input sentence, selects an antonym of the word of word stop in the input sentence from an antonym dictionary. Obtain, and replace the word of the word stop in the input sentence with the obtained opposite word,
The classifier generation device according to claim 1. The classifier generation unit changes, by the sentence change unit, a sentence included in the first teacher data to which a label for identifying a classification item and a hierarchy of the classification item of the first classification structure data is changed by the second teacher. Generate data,
The classifier generation device according to any one of claims 1 to 3 . The classifier generator further includes a representative sentence generator that classifies the text data to be classified by using the classifier generated by the classifier and generates a representative sentence of the classifier based on a result of the classification.
The classifier generation device according to any one of claims 1 to 4 . A classification for classifying the classification target data related to the second classification structure data using the classifier generated by the classifier generation unit, and determining whether to correct the classifier based on the result of the classification. Further comprises a vessel correction unit,
The classifier generation device according to any one of claims 1 to 5 . A classifier generation method for generating a classifier for classifying each answer to a semantically opposite question obtained from the same respondent,
A classifier generating device, a sentence changing step of changing the meaning of the sentence by replacing words or adding or deleting words,
The classifier generation device changes the sentence included in the first classification structure data indicating the classification item and the hierarchy of the classification item, which is the destination to classify the sentence data, by the sentence changing step to generate the second classification structure data. A classifier generating step of generating a classifier using the second classification structure data and the second teacher data provided with a label for identifying a classification item of the second classification structure data and a hierarchy of the classification item; , only including,
The sentence changing step performs a morpheme analysis and a dependency analysis on the input sentence, and as a result of the morpheme analysis and the dependency analysis, if there is a predicate-argument structure in the input sentence, an antonym of the word of the predicate is detected. Acquiring from an antonym dictionary and replacing the word of the predicate in the input sentence with the acquired antonym.
Classifier generation method. A computer of a classifier generation device for generating a classifier for classifying each answer to a semantically opposite question obtained from the same respondent ,
A sentence change function that changes the meaning of a sentence by replacing words or adding or deleting words
The sentence included in the first classification structure data indicating the classification item and the hierarchy of the classification item which is the classification destination of the text data is changed by the sentence changing function to generate the second classification structure data, and the second classification structure data is generated. When a computer for realizing the a classifier generation function of generating a classifier using a second training data label identifying the hierarchy of the classification items and the classification items of the second classification configuration data has been granted A program,
The sentence modification function performs a morphological analysis and a dependency analysis on the input sentence. As a result of the morphological analysis and the dependency analysis, when the input sentence has a predicate-argument structure, an opposite word of the word of the predicate is detected. Acquiring from an antonym dictionary and replacing the word of the predicate in the input sentence with the acquired antonym.
Computer program.

Images