JP5146979B2 - Ambiguity resolution device and computer program in natural language - Google Patents

Description

  The present invention relates to natural language processing, and more particularly to eliminate ambiguity seen in word reading (kana notation in Japanese), full spelling of acronyms (acronyms), and correspondence between translated words between two languages. Related to natural language processing.

  Natural language is always ambiguous. For example, there is a problem of isomorphic allophones. An isomorphic allophone is a word with multiple readings. For example, the word “bow” in English has two readings: “bow” (butterfly ribbon) and “bow” (bow). There are many examples in Japanese. For example, the word “Ohira” can be read as “Ohira”, “Taihei”, and “Odaira”.

  Such ambiguity also exists in acronyms. For example, an acronym “ACL” can be interpreted as any of “The Association for Computational Linguistics”, “Anterior Cruciate Ligament”, and “Access Control List”. Similar ambiguity can also be found in the way of deciding the translation word for translation.

  In the case of humans, such ambiguity is best suited to the situation according to the situation in which the word occurs, as appropriate, or by using one of the means to find candidates. It is solved by selecting what seems to be. However, it is difficult to realize such processing by natural language processing.

  Such ambiguity can be a serious problem in natural language processing. For example, in the case of reading a Japanese text, when an isomorphic allophone is encountered, it is necessary to determine the pronunciation (kana notation) in order to read out with proper pronunciation. Otherwise, inappropriate reading will occur.

A proposal for solving such a problem is made in Non-Patent Document 1. In Non-Patent Document 1, learning data in which a certain word W and its corresponding meaning Si are described in advance are manually prepared, and when a certain word W is given using the learning data, Create a classifier that selects the appropriate one.
Umemura Yoshiyuki, Shimizu Tsukasa, “Different Homophones for Speech Synthesis System”, Toyota Central R & D Review, 2000, Vol. 35, No. 1, pp. 67-74

  However, in the method disclosed in Non-Patent Document 1, it is necessary to prepare learning data manually, and there is a problem that time and cost are increased. In addition, since learning data is created by a limited number of people, there is a possibility that the learning data may be biased, and there is a problem that reliability is low.

  Therefore, an object of the present invention is to provide an ambiguity resolution device in natural language that can easily and reliably resolve the ambiguity associated with natural language.

  According to the first aspect of the present invention, an ambiguity resolution device in a natural language represents a certain word, a context in which a certain word is placed in the input sentence, and the meaning of a certain word in an input sentence composed of natural language sentences. Given a set of semantic candidates including a plurality of possible semantic candidates, an ambiguity resolution device in natural language that selects the most appropriate meaning of a word in context from the set of semantic candidates A document collection means for collecting a set of documents in which a word constituting the combination co-occurs from a predetermined corpus for each combination of a word and a meaning candidate in the meaning candidate set; Using a set of documents collected for each combination by the document collection means as learning data, given a word and the context of the word in the document, A classifier creating means for automatically creating a classifier that selects an optimal meaning candidate from the semantic candidate set as the meaning of the meaning of the word, and the meaning of the word based on the context in which the word is placed in the input sentence Classification execution means for selecting an optimal one from a set of semantic candidates using a classifier.

  Given a word in an input sentence, the context in which the word is placed, and a set of semantic candidates that include multiple semantic candidates that might represent the meaning of the word, the combination of the word and the semantic candidate For each of the above, the document collection means collects a set of documents in which words constituting the combination co-occur from a predetermined corpus. The classifier creating means creates a classifier using a set of documents collected for each combination as learning data. This classifier has a function of selecting a semantic candidate most suitable as a meaning of a certain word from the semantic candidate set given a certain word and the context of the word in the document. The classification means gives the word in the input sentence and the context in which the word is placed to the classifier created in this way, and based on the result, the most appropriate meaning of the word in the input sentence, Select from a set of semantic candidates.

  That is, in this apparatus, given a certain word, the context of the word, and a plurality of semantic candidates that may correspond to the word, it is possible to automatically select a semantic candidate that seems appropriate from the context. This operation does not require manual intervention. Therefore, it is possible to provide an ambiguity resolution device that can easily select an appropriate meaning candidate and eliminate the ambiguity of an input word.

  Preferably, the classifier creating unit selects a document set collected by the document collecting unit for each of the combinations according to a predetermined criterion, and selects a document having a large number of documents included in the set, and corresponds to the set of documents. Semantic candidate selection means for performing processing for selecting only the meaning candidates to be selected as elements of the semantic candidate set, and using the document set selected by the semantic candidate selection means as learning data, a word and a document in the word And a machine learning means for automatically creating a classifier for selecting a semantic candidate most suitable as the meaning of a certain word from the semantic candidate set by machine learning. .

  Among the collected document sets, those with a small number of documents included in the set are rejected by the semantic candidate selection means. The fact that the number of documents included in the set is small means that the possibility that the word and the semantic candidate corresponding to the set co-occur is smaller than others. Therefore, it is possible to exclude inappropriate words as meanings of a word in a given context. As a result, the reliability of classification can be improved.

  More preferably, for each of the documents included in the document set selected by the document set selection unit, the machine learning unit calculates the document from a word string existing in a predetermined range before and after a certain word position in the document. Each of the documents included in the document set selected by the document set selection means and the feature quantity vector calculation means for calculating the feature quantity vector for learning having a predetermined configuration representing the context feature quantity of a certain word The learning feature data calculated by the feature value vector calculating means and the semantic candidates used at the time of searching the document are created as learning data, and the machine using the learning data is used. When learning provides a feature vector for classification having the same configuration as the feature vector for learning, a single unit in the context corresponding to the feature vector for classification is provided. The optimum as meaning, and means for automatically generating a predetermined classifier selected from means candidate set in.

  The context of a word is represented by a learning feature value vector created from a word string existing in a predetermined range before and after the word. A classifier can be automatically created by machine learning using such learning feature vectors. As a result, it is possible to provide an ambiguity resolution device that can automatically select an appropriate meaning candidate without human intervention and eliminate the ambiguity of an input word.

  More preferably, the semantic candidate selection unit selects a predetermined number of sets having a large number of documents included in the set from the set of documents collected for each combination by the document collection unit, and sets the set of documents as the set of documents. Means for selecting only the corresponding semantic candidate as the semantic candidate is included.

  A large number of documents included in a set of documents collected for a combination of an input word and a semantic candidate means that there is a high possibility that the words constituting the combination will co-occur. is there. Therefore, there is a high possibility that such semantic candidates are appropriate semantic candidates for the input word. Further, since the upper limit number of semantic candidates is set at this time, the subsequent processing can be completed in a stable time. As a result, it is possible to provide an ambiguity resolution device that can automatically select a proper meaning candidate with high reliability and stable time without manual intervention, and eliminate the ambiguity of the input word.

  The semantic candidate selection unit selects a set in which the number of documents included in the set is larger than a predetermined threshold from the set of documents collected for each combination by the document collection unit, and corresponds to the set of documents. Means for performing processing for selecting only meaning candidates to be selected as meaning candidates may be included.

  If the number of documents included in a set of documents collected for a combination of an input word and a semantic candidate is greater than a certain threshold value, the word constituting the combination may co-occur. Is high. Therefore, there is a high possibility that such semantic candidates are appropriate semantic candidates for the input word. As a result, it is possible to provide an ambiguity resolution device capable of automatically and reliably selecting an appropriate meaning candidate without human intervention and eliminating the ambiguity of an input word.

  For each combination of a word and a meaning candidate in a meaning candidate set, the document collection means uses a virtual corpus consisting of a web page existing on the Internet to generate a web page in which the words constituting the combination co-occur. Includes search means for searching and collecting sets.

  Web pages on the Internet are created and maintained by many people. Therefore, the word usage used there covers a very large number of use cases. Therefore, if a classifier is created based on such a document, the bias of the classification result can be eliminated and the reliability can be improved.

  Preferably, the means for collecting includes, for each combination of a word and a meaning candidate in the meaning candidate set, a word constituting the combination from a virtual corpus including web pages existing on the Internet. Means for searching for a set of web pages to occur and collecting them as a set of elements up to a predetermined constant.

  There is an upper limit on the number of web pages collected for a set. Therefore, there is little possibility that the learning of the classifier becomes an excessive load. As a result, it is possible to provide an ambiguity resolution device that can automatically and reliably select an appropriate meaning candidate without human intervention and eliminate the ambiguity of an input word.

  The computer program according to the second aspect of the present invention, when executed by a computer, causes the computer to function as an ambiguity resolution device in any of the natural languages described above.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. There are three embodiments. The first embodiment relates to an apparatus for determining kana notation of a word having a plurality of kana notations (readings) in speech synthesis for Japanese input sentences. The second embodiment relates to an apparatus for giving an English definition (full spelling) to an English acronym. The third embodiment relates to an apparatus for selecting one of a plurality of English translations for a Japanese word in translation from Japanese to English. That is, in the present invention, the “meaning” of a word can be evaluated not only as “meaning” in the Japanese language dictionary in the case of Japanese, but also as equivalent to the word as seen from a certain standard. Such a word or a set of words or a character string.

  In the drawings used for the description of the following embodiments, the same reference numerals are assigned to the same components. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated. As will be described later, each embodiment can be realized by computer hardware and a computer program executed on the computer hardware. Therefore, about a part of functional block in the block diagram shown below, the function and structure are shown in the flowchart format of the computer program for implement | achieving it.

<First Embodiment>
[Constitution]
FIG. 1 shows a block diagram of a speech synthesis system 30 according to the first embodiment of the present invention. Referring to FIG. 1, the speech synthesis system 30 sequentially extracts and stores a predetermined length from an input sentence storage unit 40 for storing a Japanese input sentence that is a target of speech synthesis, and the input sentence storage unit 40. Input sentence buffer 42, a dictionary group 46 composed of a plurality of dictionaries that store Japanese words and their kana notations in association with each other, and morphological analysis of sentences included in the input sentence buffer 42 If there is a word to be included, a kana notation is searched with reference to the dictionary group 46, and a kana conversion unit 44 for outputting a morpheme string to which information such as kana notation is attached is included.

  As already mentioned, some words including kanji may have a plurality of kana notations. For speech synthesis, it is necessary to select an appropriate one of the plural kana notations. For this purpose, the speech synthesis system 30 is connected to the kana conversion unit 44 and the so-called Internet 52, and the kana conversion unit 44 has a plurality of kana notation candidates Rk (k = 1 to K: K is the number of kana notation candidates). ) Is detected for each combination of the word W and the kana notation candidate Rk on the Internet 52 for a web page on which the word W and the kana notation candidate Rk co-occur. And an isomorphic allophone elimination processing unit 50 for determining a kana notation suitable for the word W and giving it to the kana conversion unit 44 by classification by machine learning using the text of the obtained web page as learning data. In other words, in this system, an example document is collected by regarding a set of web pages on the Internet 52 as one virtual corpus.

  The speech synthesis system 30 further includes a kana notation input sentence storage unit 54 for storing an input sentence with kana notation output from the kana conversion unit 44, and a speech storing speech corresponding to kana notation for speech synthesis. A speech synthesizer 56 for reading an input sentence with kana notation from the database 48 and the kana notation input sentence storage unit 54, performing speech synthesis with reference to the speech database 48, and outputting an analog speech signal, and a speech synthesizer 56 And a speaker 58 for converting an analog audio signal output from the audio into audio.

  In the present embodiment, a part called “snippet” is used for machine learning in the text of the web page searched by the homomorphic abnormal word elimination processing unit 50 from the Internet 52. A “snippet” refers to a short sentence for explaining the content of a searched web page in a search result by a so-called search engine on the Internet. In many cases, the snippet is composed of a portion of text including a word that is regarded as a search keyword. In addition, although the original search program may be used for the search of the web page by the isomorphic abnormal word elimination processing unit 50, in this embodiment, the existing search service site is used, and the word W and the kana notation candidate. A query for performing an AND search on Rk is issued to the search service site, and the result is obtained. In the present embodiment, in order to stabilize the processing time, a criterion of 1000 searches per search is provided as the upper limit of the number of searches.

  The portion of speech synthesis performed by the speech synthesizer 56 is not directly related to the present invention, and the details thereof will not be described here.

  FIG. 2 shows a detailed block diagram of the homomorphic abnormal word elimination processing unit 50. Referring to FIG. 2, when a word W is given, the homomorphic abnormal word elimination processing unit 50 is included in a window having a predetermined length centered on the word W among the input sentences stored in the input sentence buffer 42. If a predetermined feature vector related to the word W is given by learning based on the word, a decision tree 82 that can be learned to output an appropriate kana notation corresponding to the word W, and the word W from the kana conversion unit 44 A decision tree creation unit for receiving the combination (W, Rk) with the kana notation candidate Rk, collecting web page snippets co-occurring from the Internet 52, and learning the decision tree 82 using the result 80, connected to the kana conversion unit 44, and given from the kana conversion unit 44 a word W in the combination (W, Rk) and a word string 85 in a predetermined range centered on the word W in the input sentence, So A classification feature vector creation unit 84 for creating and outputting a classification feature vector suitable for classification by the decision tree 82, and a classification feature vector output from the classification feature vector creation unit 84. And a classification execution unit 86 for providing the kana notation which is the classification result obtained from the decision tree 82 to the kana conversion unit 44 as a result.

  When a combination (W, Rk) of the word W and its kana notation Rk is given, the decision tree creating unit 80 searches the web page on which they co-occur on the Internet 52, and the search unit A search result storage unit 102 for storing a set of snippets of web pages searched by 100 (hereinafter simply referred to as “set of web pages”) for each combination (W, Rk), and a combination (W, Rk) Among them, the set of web pages is sorted in descending order of the number of acquired web pages, and only N items (N is a natural number) with the highest number of items are selected, whereby learning data for the decision tree 82 is obtained. And a sort and selection unit 104 for creation. In the present embodiment, the N kana notation candidates Rk included in (W, Rk) selected by the sorting and selecting unit 104 are left as kana notation candidates and are used for later learning of the decision tree.

  In this process, other document sets are rejected, and kana notation candidates used for searching these document sets are also rejected. This is because a kana notation candidate with a low frequency of co-occurring with the word W is generally considered to be inappropriate as a candidate. However, depending on the application, it may be better not to reject even such a low frequency kana notation candidate.

  The decision tree creation unit 80 further stores a learning data storage unit 106 for storing the learning data created by the sorting and selection unit 104, and the learning data storage unit 106 for the search target word W. The learning feature vector creation unit 108 for creating a predetermined learning feature vector and the learning feature vector creation unit 108 created from each of the web pages of the kana notation candidates Rk searched for the word W. A feature vector storage unit 110 for storing the feature vector and a decision tree learning unit 112 for learning the decision tree 82 using the feature vector stored in the feature vector storage unit 110 are included.

  FIG. 3 is a flowchart of a computer program for realizing the search unit 100 shown in FIG. Referring to FIG. 3, this program includes three steps of steps 130 to 134 that are repeated for each of kana notation candidates Rk (k = 1 to K) for a certain word W.

  In step 130, a request for searching the web page with the maximum number MAX = 1000 is transmitted to a search engine on the Internet by the query “word W and word Rk”.

  In step 132, as a search result, a set of snippets {Sn (W, Rk)} (n = 1 to Lk, k = 1 to K) including the word W and the kana notation candidates Rk (k = 1 to N). To get. Here, Lk is the number of search results obtained for the combination of the word W and the kana notation candidate Rk.

  In step 134, the kana notation candidate Rk is deleted from each set Sn, thereby generating a set of search result snippets {(Tn (W), Rk) | n = 1 to Lk}.

  The above three steps are repeated for all the kana notation candidates Rk for the word W.

  The function of the search unit 100 shown in FIG. 2 is realized by such a program.

  The sort and selection unit 104 extracts a set of snippets with the highest number of searches {(Tn (W), Rk) | n = 1 to Lk} and stores them in the learning data storage unit 106 as learning data. Shall be.

  FIG. 4 shows the configuration of the learning feature vector creation unit 108 shown in FIG. 2 in a block diagram format. Referring to FIG. 4, learning feature vector creation unit 108 is included in a set of learning data snippets {(Tn (W), Rk) | n = 1 to Lk} stored in learning data storage unit 106. From each snippet, M word groups (total 2M words) before and after the word W existing in the snippet (the total 2M word groups are called “windows”) are extracted. In order to determine the configuration of the classification feature vector used for learning of the decision tree 82 (see FIG. 2) by the vocabulary other than the word W that appears in the learning data 106 stored in the extraction unit 150 and the learning data storage unit 106. Each of the snippets based on the word group extracted for each snippet by the extraction unit 150 according to the configuration of the feature vector determined by the vector configuration determination unit 152 and the vector configuration determination unit 152 To calculate the elements of the feature vector, to create a feature vector of each snippet, and a component calculation unit 154 for storing the feature vector storage unit 110.

FIG. 5 schematically shows the configuration of the “window” centered on the word W. Referring to FIG. 5, among word strings of learning snippet 170, word strings existing around and around word W include words W and “W _−m , W _{− (m−1).} , W _{− (m−2)} ,..., W ₋₂ , W ₋₁ , W, W ₁ , W ₂ ,..., W _m−2 , W _m−1 , W _m ”. A window 172 having a window length of 2 m is formed by a word string including a word string 174 consisting of m words preceding the word W and a word string 176 consisting of m words after the word W. The In this embodiment, the window length is 2M.

  The vector configuration determination unit 152 determines the configuration of the feature vector as follows. That is, the vector configuration determination unit 152 calculates the frequency of words appearing in the learning data existing in the learning data storage unit 106 for each word. The vector configuration determination unit 152 further selects only H words having the highest frequency. Further, the vector configuration determining unit 152 sets the dimension of the feature vector to H dimension, and associates the 1st to Hth elements with the words having the 1st to Hth frequencies, respectively. Thereby, the structure of the feature vector is determined. The number of elements of this feature vector is H. Each element takes a value of 0 or 1. Each element has a value of 1 when a word corresponding to the element appears in a window having a window length of 2M centered on the word W in the snippet, and a value of 0 if it does not appear.

  Therefore, the process performed by the element calculation unit 154 for a certain learning snippet Ti is as follows. That is, for each element of the H-dimensional feature vector corresponding to the snippet Ti, the element calculation unit 154 determines whether the corresponding word appears in the window with a window length 2M centered on the word W in the snippet Ti. Check for no. The value of the element is 1 if the word appears and 0 if it does not appear. By performing this process for all H elements, the feature vector Vi of the snippet Ti is calculated. This feature vector Vi and the kana notation candidate Rk of the combination (W, Rk) from which the feature vector was obtained are associated with each other (assuming that the correct answer to the feature vector is the kana notation candidate Rk) and used for learning of the decision tree 82. .

  The creation of classification feature vectors performed by the classification feature vector creation unit 84 shown in FIG. 2 is basically the same as this. That is, the classification feature vector creation unit 84 receives information on the word corresponding to each element of the feature vector from the vector configuration determination unit 152 (see FIG. 4) of the learning feature vector creation unit 108, and the word to be processed For W, a classification feature vector for the word W is created depending on whether or not a predetermined word appears in a window having a window length 2M centered on the word W. That is, this feature vector has the same configuration as the feature vector created by the learning feature vector creation unit 108.

  The decision tree learning unit 112 learns the decision tree 82 by machine learning. Since this learning method is a matter commonly used in the field of machine learning, a detailed description thereof will not be given here.

  FIG. 6 shows a decision tree 200 related to the word “Sahara”, which is an example of a decision tree created by the element calculation unit 154 according to the present embodiment. Referring to FIG. 6, this decision tree includes four intermediate nodes 210, 212, 214, and 216 and five terminal nodes 230, 232, 234, 236, and 238, and each node 210, 212, 214 And at 216, a question is made as to whether each word in the window meets a certain condition.

  The question of node 210 is whether there is a keyword “Chiba Prefecture” in a window with a window length of 2M centering on the word “Sahara”. If there is, the process proceeds to node 230, and "Sawara" is selected as the kana notation corresponding to "Sahara". If not, go to node 212. In FIG. 6, whether a specific word such as “Chiba Prefecture” is in the window is asked, but in actual processing, the word “ This determination is made by checking whether the value of the element (bit) corresponding to “Chiba Prefecture” is 1 or 0.

  The question at node 212 is whether there is a keyword “Kanagawa Prefecture”. If there is, the process proceeds to node 232, and “Sahara” is selected as the kana notation corresponding to “Sahara”. If not, go to node 214.

  The question of node 214 is whether there is a keyword “Narita”. If there is, the process proceeds to node 234, and "Sawara" 234 is selected as the reading corresponding to "Sahara". If not, go to node 216.

  The question at node 216 is whether there is a keyword “Yokosuka”. If there is, the process proceeds to node 236, and “Sahara” is selected as the kana notation corresponding to “Sahara”. If not, the process proceeds to node 238, where “Sawara” is selected as the kana notation corresponding to “Sahara”.

  In the present embodiment, a decision tree 200 is basically created for each word. Given a feature vector corresponding to a word, the kana notation of that word is selected by following the decision tree corresponding to that word according to the value of each element of the feature vector.

[Operation]
1 to 6, the above-described speech synthesis system 30 operates as follows. In the input sentence storage unit 40 shown in FIG. 1, a Japanese sentence to be subjected to speech synthesis is stored in advance. A predetermined length portion is read out and stored in the input sentence buffer 42.

The kana conversion unit 44 performs morphological analysis on the sentence stored in the input sentence buffer 42 with reference to the dictionary group 46. As a result, the part of speech, kana notation (in the case of kanji), utilization type, utilization form, etc. are determined for each word. If a plurality of kana notations are obtained for one word (that is, if there is an isomorphic allophone), the kana conversion unit 44 uses the same combination of the word (word W) and the kana notation. This is given to the abnormal word elimination processing unit 50. In the following description, the notation used in the configuration is used. That is, K kana notation candidates obtained for a certain word W are set as kana notation candidates R _{1 to} R _K.

  Referring to FIG. 2, when a combination (W, Rk) of word W and kana notation candidate Rk (k = 1 to K) is given, search unit 100 uses (word W and word Rk) as a query. A search request is transmitted to the search engine on the Internet 52 under the condition that the search number upper limit = 1000 (step 130 in FIG. 3). Then, a set of web page snippets {Sn (W, Rk)} (n = 1 to Lk) obtained from the search engine in response to the search request is acquired (step 132 in FIG. 3). Here, Lk is the number of search results (web pages) obtained for the query (word W and word Rk). A set of search result snippets obtained by deleting the word Rk from each set of snippets is stored in the search result storage unit 102 (step 134 in FIG. 3). A set of these snippets is obtained for each combination of (word W, kana notation candidate Rk). In the present embodiment, the upper limit MAX of the number Lk of each element of the set of snippets is 1000 as described above.

  The search unit 100 executes the above-described processing for each combination of the word W and the kana notation candidate Rk. That is, the processing of steps 130 to 134 in FIG. 3 is executed for each combination. As a result, the search result storage unit 102 stores a set {Sn (W, Rk)} of search result snippets for each of these combinations.

  The sort and selection unit 104 sorts the set of snippets {Sn (W, Rk)} stored in the search result storage unit 102 in descending order using the number of elements Lk as a key. The sort and selection unit 104 further selects a set of top N snippets {(Tn (W), Rk) | n = 1 to Lk} from the sort results, and the kana notation element from which these snippets are obtained In association with Rk, it is stored as learning data in the learning data storage unit 106. That is, the learning data storage unit 106 stores N pieces of snippets in order from the search result having the largest number.

  Referring to FIG. 4, when learning data is stored in learning data storage unit 106, vector configuration determining unit 152 of learning feature vector creation unit 108 calculates the frequency of words appearing in these learning data for each word. To do. The vector configuration determination unit 152 further selects words up to the highest H frequency of appearance. The structure of the feature vector is determined by associating the 1st to Hth elements of the feature vector with the words having the appearance frequencies of 1st to Hth. The vector configuration determining unit 152 gives the configuration of the feature vector (that is, information relating to the word corresponding to each element of the feature vector) to the classification feature vector creating unit 84 shown in FIG. 2 and the element calculating unit 154 shown in FIG.

  On the other hand, for each snippet stored in the learning data storage unit 106, the extraction unit 150 extracts a window having a window length 2M centered on the word W and gives the extracted value to the element calculation unit 154.

  The element calculation unit 154 calculates the value of each element of the feature vector of each snippet based on the word included in the window provided from the extraction unit 150 according to the vector configuration provided from the vector configuration determination unit 152. As a result, a feature vector for each snippet is obtained. The element calculation unit 154 stores each snippet as learning data in the feature vector storage unit 110 in association with the kana notation candidate Rk when the snippet is searched.

  Referring to FIG. 2, the decision tree learning unit 112 learns the decision tree 82 by machine learning using the feature vectors stored in the feature vector storage unit 110 and kana notation candidates associated with the feature vectors. To do.

  With the above processing, when the decision tree 82 is given a feature vector representing a word string in a window having a window length of 2M centered on a certain word W, that is, the context of the word W, the kana notation of the word W in that context is given. Will function to output the most appropriate one.

  On the other hand, the kana conversion unit 44 requests the classification feature vector creation unit 84 to cancel the homomorphic abnormal word, and a word string included in a window having a window length of 2M centering on the word W in the input sentence. 85 is given. For the word W, the classification feature vector creation unit 84 supplies the word string 85 included in the window having a window length 2M centered on the word W in the input sentence, given from the kana conversion unit 44, and the vector shown in FIG. A feature vector of the word W is created by the same processing as the element calculation unit 154 based on the vector configuration given from the configuration determination unit 152 and given to the classification execution unit 86.

  The classification execution unit 86 gives this feature vector to the decision tree 82. The decision tree 82 has been learned to output an appropriate kana notation of the word W when the feature vector created by the above method is given from the window length M centering on the word W. The classification execution unit 86 obtains this kana notation from the decision tree 82 and gives it to the kana conversion unit 44.

  The kana conversion unit 44 adds the kana notation thus obtained from the homomorphic abnormal word elimination processing unit 50 to the problematic word W in the same manner as the result of the morphological analysis. Further, the kana conversion unit 44 finishes the morphological analysis, and gives the synthesizer 56 a morpheme string to which information such as part of speech, kana notation (in the case of kanji), a utilization type, a utilization form, and the like is attached. In this case, since the homomorphic allophone word has already been eliminated by the isomorphic allophone word elimination processing unit 50, only one kana notation is attached to one word.

  The speech synthesizer 56 extracts an appropriate speech waveform from the speech database 48 using a kana notation attached to the morpheme based on the given morpheme sequence, creates synthesized speech waveform data by waveform connection processing, The synthesized speech waveform data is converted to analog and supplied to the speaker 58. The speaker 58 converts this sound signal into sound.

  As described above, according to the speech synthesis system 30, even if the input sentence stored in the input sentence storage unit 40 includes an isomorphic allophone word, the isomorphic allophone word elimination unit 50 eliminates the isomorphic allophone word. And only one kana expression is assigned to the word. A web page on the Internet 52 is used as a virtual corpus, and a decision tree is automatically learned to eliminate this homomorphic allophone. There is no need to create learning data by hand, and much less time is required to eliminate homomorphic abnormal words than in the past. Further, since the web page searched on the Internet 52 is created by a large number of people, the learning data is less biased compared to the case where a small number of people create learning data, and the range covered by the web page is searched. Also become wider. Therefore, there is an effect that the reliability of eliminating homomorphic abnormal words is higher than in the prior art.

[Realization by computer]
The speech synthesis system 30 according to the first embodiment described above is realized by computer hardware and computer software executed on the computer hardware as described above. FIG. 7 shows an external appearance of a general computer system 250 for realizing the speech synthesis system 30, and FIG. 8 shows an internal configuration of the computer system 250 in a block diagram form.

  Referring to FIG. 7, a computer system 250 includes a computer 260, a monitor 262, a keyboard 266, a mouse 268, a microphone 290, and a pair of speakers 58, all connected to the computer 260. The computer 260 includes a DVD drive 270 capable of reproducing and recording a DVD (Digital Versatile Disc), and a memory port 272 into which a semiconductor memory storage device conforming to a predetermined standard can be mounted. The internal configuration of the computer 260 will be described later with reference to FIG.

  Referring to FIG. 8, in addition to DVD drive 270 and memory port 272 shown in FIG. 7, computer 260 is connected to CPU (Central Processing Unit) 276 and bus 286 connected to CPU 276, both connected to bus 286. ROM (Read Only Memory) 278, RAM (Random Access Memory) 280, hard disk 274, network interface 296, and sound board 288.

  A DVD 282 is attached to the DVD drive 270. A semiconductor memory storage device 284 is attached to the memory port 272. CPU 276 can access DVD 282 and memory 284 via bus 286 and DVD drive 270 and memory port 272, respectively.

  The keyboard 266, the mouse 268, and the monitor 262 are all connected to the bus 286 of the computer 260 via an interface (not shown). The speaker 58 and the microphone 290 are connected to the sound board 288. In this computer system 250, the speech synthesis program executed by the CPU 276 finally generates digital waveform waveform data. When the sound board 288 receives the sound waveform data from the CPU 276, the sound board 288 converts the sound waveform data into an analog signal and generates sound through the speaker 58.

  The input sentence storage unit 40, dictionary group 46, kana notation input sentence storage unit 54, speech database 48, search result storage unit 102, learning data storage unit 106, feature vector storage unit 110, etc. in the above embodiment are the RAM 280, hard disk 274, DVD disk 282, and semiconductor memory storage device 284 can be realized. Actually, the most efficient storage device is selected to realize each storage unit depending on the capacity of data to be stored, the speed required for reading and writing, and the like.

  The computer program for realizing the speech synthesis system 30 according to the first embodiment described above may be a single program or a combination of a plurality of programs. In particular, among the functions of each unit described above, morpheme analysis processing performed in the kana conversion unit 44 shown in FIG. 1, speech synthesis processing performed in the speech synthesis unit 56, and snippet search processing executed by the search unit 100 shown in FIG. As for the sorting and selection processing executed by the sorting and selection unit 104, the learning processing of the decision tree 82 executed by the decision tree learning unit 112, etc., programs that have already been widely distributed can be used as they are. Of course, since these programs are created for general use, it is necessary to make appropriate adjustments, but for those who have ordinary knowledge in this technical field, they are within the range that can be easily realized according to the purpose. stay.

  Furthermore, the processing in the learning feature vector creation unit 108 and the classification feature vector creation unit 84 is appropriately realized according to the specifications as long as it has ordinary knowledge in the technical field based on the above description. Is possible.

  These programs are stored in a storage medium such as a DVD disk 282 or distributed through a network such as the Internet 52, and are usually stored in a nonvolatile external storage device such as a hard disk 274. At the time of execution, the CPU 276 executes an instruction that is copied from the hard disk 274 to the RAM 280 and read from an address indicated by a program counter (not shown) in the CPU 276, thereby realizing the above-described expected function. Since the operation mode of the computer hardware itself is well known, no further detailed description will be given here.

<Second Embodiment>
FIG. 9 is a block diagram showing a configuration of a multiple definition acronym resolution system 330 that gives an appropriate definition to an English acronym having a plurality of definitions according to the second embodiment of the present invention. This ambiguous acronym resolution system 330 utilizes the fact that there are many documents that give the definition of the acronym in the vicinity of the acronym, and that the words existing in the vicinity of the acronym have some tendency depending on the field of the document. Thus, an appropriate definition is given to the acronym based on the same principle as the elimination of the homomorphic allophone in the first embodiment.

  Referring to FIG. 9, this ambiguous acronym resolution system 330 includes an input sentence storage unit 340 for storing an input sentence that may contain an acronym, and a predetermined portion of the input sentence stored in the input sentence storage unit 340. Morphological analysis of the input sentence buffer 342 for reading, a dictionary group 346 composed of data including acronyms and a list of definitions thereof, and the input sentence stored in the input sentence buffer 342, and finding an acronym with no definition And an acronym interpretation unit 344 for determining the definition of the acronym by the dictionary group 346, attaching the definition to the acronym in the input sentence, and outputting the input sentence.

  The ambiguous acronym elimination system 330 further includes an acronym definition-added input sentence storage unit 354 for storing an input sentence in which an acronym is defined and an acronym definition-added input sentence storage unit 354, which are output from the acronym interpretation unit 344. A sentence understanding device 356 for understanding the meaning of the stored input sentence.

  As already mentioned, some acronyms can have multiple definitions. In such a case, the acronym interpretation unit 344 cannot output the acronym with a plurality of definitions. This is because the text comprehension device 356 becomes an obstacle to text comprehension. Therefore, when a plurality of definitions are found from the dictionary group 346 for an acronym that is not defined in the input sentence, it is necessary to be able to automatically select an appropriate one of them by some means. .

  In order to solve such a problem, the ambiguous acronym resolution system 330 according to the present embodiment is connected to the acronym interpretation unit 344 and the Internet 52, and the acronym from the acronym interpretation unit 344 and a plurality of acronyms obtained for the acronym are obtained. And a word string existing in a predetermined window before and after the acronym, a given word string among a plurality of given definition candidates by a learning process using the Internet 52 as a corpus The ambiguous acronym elimination processing unit 350 is selected to select the most appropriate one for the above and perform the processing given to the acronym interpretation unit 344.

  Although the details of the configuration of the ambiguity acronym elimination processing unit 350 will not be described here, the configuration and operation of the ambiguity acronym elimination processing unit 350 are the same as those of the isomorphic abnormal word elimination processing unit 50 in the first embodiment. That is, the ambiguous acronym elimination processing unit 350 determines an appropriate definition for an acronym in the following procedure.

  (1) Given an acronym A and a definition candidate Dk (k = 1 to K: K is the number of definition candidates), for each definition candidate Dk, a snippet of a web page in which acronym A and the definition candidate Dk co-occur To the search engine on the Internet 52.

  (2) As a search result, a set of snippets {Sn (A, Dk)} (n = 1 to Lk) including acronym A and definition candidate Dk (where Lj (j = 1 to k) is acronym A and definition candidate) Represents the number of snippets retrieved for the combination with Dk).

  (3) By deleting the definition candidate Dk from each of the snippet sets {Sn (A, Dk)}, a set of search result snippets {(Tn (A), Dk) | n = 1 to Lk} Create

  (4) The above three processes are repeated for all definition candidates Dk.

  (5) The set Sn of the searched web page snippets is sorted in descending order of the number of web pages (number of search results) included in them, and only the top N are selected among them. A set of learning snippets {(Tn (A), Dk) | n = 1 to Lk} is extracted and stored as learning data.

  (6) Using this learning data, a plurality of feature vectors for learning are created in exactly the same manner as the learning feature vector creation unit 108 shown in FIG. The method of creating the feature vector is exactly the same as in the first embodiment. Similarly to the first embodiment, the window length when creating the feature vector is represented by 2M.

  (7) These feature vectors are associated with the definition candidates used for the search when the snippet giving the feature vectors is searched for as learning data.

  (8) Learning of the decision tree is performed using the learning data. As a result of this learning, when a decision vector is given a feature vector created by a word included in the window length 2M centered on the acronym A to be resolved, the appropriate definition for the acronym is given. Will be output.

  (9) A feature vector for a decision tree is created from a window having a window length of 2M, centering on the acronym A that is the object of ambiguity resolution in the input sentence.

  (10) By giving this feature vector to the decision tree, an output for selecting only one definition of acronym A is obtained from the decision tree. By giving this output from the ambiguous acronym elimination processing unit 350 to the acronym interpretation unit 344, the acronym interpretation unit 344 attaches only one definition given from the ambiguous acronym elimination processing unit 350 to the acronym, and defines the acronym definition. The data can be output to the additional input sentence storage unit 354.

<Third Embodiment>
FIG. 10 is a block diagram of a Japanese-English automatic translation system 430 according to the third embodiment. Referring to FIG. 10, automatic translation system 430 stores a Japanese sentence storage unit 440 for storing Japanese input sentences, and a predetermined amount of Japanese sentences stored in Japanese sentence storage unit 440. An input sentence buffer 442, a dictionary group 446 composed of one or more dictionaries from Japanese to English, and a Japanese sentence stored in the input sentence buffer 442 as a preprocessing for automatic translation, morphological analysis is performed for each word. A translation determining unit 444 for assigning and outputting English translations with reference to the group 446, a Japanese sentence storage unit 454 with a translation for storing the pre-translated Japanese sentence, and a translation-added Japanese sentence And an automatic translation device 456 for translating Japanese sentences with translations stored in the sentence storage unit 454 into English using the translations.

  However, as already described, there may be a case where there are a plurality of English translation candidates for one input Japanese word. If such a plurality of English translation candidates are directly attached to the Japanese word and output from the translation determination unit 444 at that time, the automatic translation apparatus 456 has trouble in translation. Therefore, it is necessary to select an appropriate one from the plurality of translated word candidates by some means.

  For this purpose, the automatic translation system 430 according to the present embodiment is connected to the translation determination unit 444 and the Internet 52, and the translation determination unit 444 provides a Japanese word and a plurality of translation candidates obtained for the word. And a word string existing in a predetermined window before and after the Japanese word in the input sentence, given by a learning process using the Internet 52 as a corpus, It includes a multiple meaning word elimination processing unit 450 for selecting the most likely word sequence for the given word string and giving it to the translated word determination unit 444.

  Although details of the configuration of the multiple meaning word elimination processing unit 450 will not be described here, the processing in the multiple meaning word elimination processing unit 450 is the same as the processing in the isomorphic abnormal word elimination processing unit 50 of the first embodiment and the second It will be understood that the processing is the same as the processing in the ambiguous acronym elimination processing unit 350 of the embodiment, and therefore the configuration thereof is the same as the configuration of the isomorphic abnormal word elimination processing unit 50.

  The translation determination unit 444 reads the sentence in the input sentence buffer 442, performs morphological analysis, assigns an English translation by referring to the dictionary group 446 for each word, and outputs the translated word to the Japanese sentence storage unit 454 with translation. . When a plurality of translation candidates appear for one Japanese word, the translation determination unit 444 delivers the Japanese word and the plurality of translation candidates to the ambiguity translation cancellation processing unit 450 and requests ambiguity resolution. The multiple meaning word elimination processing unit 450 creates a decision tree by exactly the same operation as the homomorphic abnormal word elimination processing unit 50 in the first embodiment, and the windows before and after the given Japanese word in the input sentence A feature vector is created using the word string in the list, and given to the decision tree, an appropriate translation word candidate is obtained and returned to the translation word determination unit 444. The translated word determining unit 444 attaches only one translated word given from the multiple meaning translated word elimination processing unit 450 to the Japanese word in question, and outputs it to the translated Japanese sentence storage unit 454. Accordingly, there is no problem in automatic translation processing in the automatic translation apparatus 456.

  As is apparent from the description of the first to third embodiments, the ambiguity elimination or the ambiguity elimination according to the present invention can be easily applied over a wide range in the field of natural language processing. In addition, the mechanism of the part for eliminating ambiguity may be basically the same. Of course, there may be various design items in the details of the cancellation processing, but a method effective in a certain field can be applied to other fields basically as it is.

  For example, the device for eliminating ambiguity or ambiguity according to the present invention can be applied not only to translation between Japanese and English but also to translation of words between all languages regardless of language differences. The mechanism for eliminating homomorphic abnormal words in the first embodiment can be applied almost as it is regardless of the language. Of course, language-specific adjustments may be necessary (for example, morphological analysis in Japanese), but if that part is always done in advance as a premise in natural language processing, it may be ambiguous or ambiguous. The mechanism of the part may be the same regardless of language.

  Therefore, the present invention can be applied to a wide area in the field of natural language processing, and transplantation from one area to another can be realized very easily.

<Possible modification>
In the embodiment described above, a decision tree is used to determine an appropriate kana notation, an acronym definition, and a translation. However, the present invention is not limited to the one using a decision tree, and any one of a plurality of objects is selected as a target object according to the context (environment) where the target word or word string is placed by learning data collected from the Internet. Any classification method can be used as long as it can be learned by machine learning. For example, naive Bayes, decision lists, k-nearest neighbor methods, online algorithms, maximum entropy methods, support vector machines, boosting, etc. can be used.

  In the above-described embodiment, web page snippets are collected as learning data. However, the present invention is not limited to such an embodiment. For example, the entire web page may be processed. Further, for example, the upper limit MAX of the number of web pages collected for a combination of one word W and kana notation Rk is limited to 1000, but it goes without saying that this number can be freely changed. It is also possible not to use such a limitation.

  Further, in the above-described embodiment, when a word in question and several candidates to be paired with the word are given, the Internet is accessed at that time, and a decision tree is created. However, the present invention is not limited to such an embodiment. For example, by performing the above-described processing on a certain test sentence in advance, a classification device is prepared in advance to eliminate some ambiguity for some words included in the test sentence. You may keep it. If such a classifier is prepared in advance for each of many words, it is not necessary to learn the classifier after the word is given, and an appropriate answer can be given immediately. If the ambiguity cannot be resolved by the plurality of classifiers, then one classifier may be created again as described in the above embodiment to obtain an appropriate answer.

  Further, in the first embodiment described above, the kana notation candidates selected by the sort and selection unit 104 are the top N items (N is a plurality) having the largest number of web pages hit by the search unit 100. The same applies to the case of the second embodiment and the third embodiment. However, the present invention is not limited to such an embodiment. For example, only the topmost kana notation candidate having a large number of hits in the processing of the sort and selection unit 104 may be adopted as the kana notation of the word W. In this case, the decision tree functions as performing a 1: 1 classification. However, since the context of the word W is not considered at all in this method, the reliability of the result is low and it cannot be said that the ambiguity is resolved.

  Also, in the sort and selection unit 104 processing of the first embodiment, not all of the top N web pages with a large number of hit web pages, but all web page hits of a predetermined threshold or higher are all kana notation candidates You may choose as Alternatively, among the total number of hits, the ratios of the candidates from the top may be integrated and all of the hits exceeding the predetermined ratio may be adopted as kana notation candidates regardless of the number.

  Further, in the above embodiment, the meaning candidate is determined for each word. However, the present invention is not limited to such an embodiment. As a dictionary heading for creating a set of meaning candidates, for example, by providing a phrase consisting of a plurality of words, the meaning of the phrase can be selected from a plurality of meaning sets. Become.

  Then, it is preferable to accumulate the classification devices thus obtained as needed, because the number of words that can be immediately resolved is increased.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

1 is a block diagram of a speech synthesis system 30 according to a first embodiment of the present invention. It is a block diagram of the isomorphic abnormal word elimination process part 50 shown in FIG. It is a flowchart which shows the control structure of the computer program for implement | achieving the search part 100 of FIG. FIG. 3 is a block diagram of a learning feature vector creation unit 108 in FIG. 2. It is a figure for demonstrating the concept of a "window." It is a figure which shows an example of a decision tree typically. It is a figure which shows the external appearance of the computer system 250 which implement | achieves the speech synthesis system 30 which concerns on 1st Embodiment. It is a block diagram which shows the internal structure of the computer system 250 shown in FIG. It is a block diagram of the ambiguous acronym elimination system 330 which concerns on the 2nd Embodiment of this invention. It is a block diagram of the automatic translation system 430 which concerns on the 3rd Embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 30 Speech synthesis system 40,340 Input sentence memory | storage part 42,342,442 Input sentence buffer 44 Kana conversion part 46,346,446 Dictionaries 48 Speech database 50 Isomorphic allophone cancellation processing part 52 Internet 54 Kana notation input sentence memory part DESCRIPTION OF SYMBOLS 56 Speech synthesizer 58 Speaker 80 Decision tree creation part 82 Decision tree 84 Classification feature vector creation part 86 Classification execution part 100 Search part 102 Search result storage part 104 Sort and selection part 106 Learning data storage part 108 Learning feature vector creation part DESCRIPTION OF SYMBOLS 110 Feature vector memory | storage part 112 Decision tree learning part 344 Acronym interpretation part 350 Ambiguous acronym resolution processing part 354 Input sentence memory part with an acronym definition 440 Japanese sentence memory | storage part 444 Translation word determination part 450 Ambiguous translation word resolution processing part 454 Japanese sentence memory part with a translation 456 Automatic translation Equipment

Claims (6)

In an input sentence composed of a natural language sentence, a certain word, a context in which the certain word is placed in the input sentence, and a set of semantic candidates including a plurality of semantic candidates that may represent the meaning of the certain word Given a ambiguity resolver in natural language that selects the most appropriate meaning of the word in the context from the set of semantic candidates,
Wherein and words, for each combination of the meanings candidates for the meaning candidate set in, from a given corpus, a document collection means for the word that constitute the combination to collect a set of documents to be co-occur,
Using the collection of documents collected for each of the combinations by the document collection means as learning data, given the certain word and the context in the document of the word, the certain word in the context A classifier creating means for automatically creating a classifier that selects a semantic candidate optimum as a meaning from the semantic candidate set;
In the input sentence, based on a context in which the certain word is placed, classification execution means for selecting an optimum meaning of the certain word from the set of meaning candidates using the classifier Including
The document collection means includes a virtual document composed of documents existing on the Internet for each combination of the certain word and the meaning candidate in the meaning candidate set, using both the certain word and the meaning candidate as a search keyword. from corpus comprises search means for word that constitute the combination are searched to collect a set of documents that co-occur,
The classifier creating means includes:
Of the set of documents collected for each of the combinations by the document collection means, a document having a large number of documents included in the set is selected according to a predetermined criterion, and only semantic candidates corresponding to the set of documents are selected. Semantic candidate selection means for performing processing of selecting as a candidate of a semantic candidate set;
When a set of documents selected by the meaning candidate selecting means is used as learning data and given the word and the context in the document of the word, the meaning candidate most suitable as the meaning of the word in the context is given. And a machine learning means for automatically creating, by machine learning, a classifier that selects the semantic candidate set from the semantic candidate set. The machine learning means includes
For each document included in the document set selected by the document set selection unit, the context of the word in the document is determined from a word string existing in a predetermined range before and after the position of the word in the document. A feature amount vector calculating means for calculating a feature amount vector for learning having a predetermined configuration, which represents the feature amount of
A learning feature amount vector calculated by the feature amount vector calculation unit for each of the documents included in the document set selected by the document set selection unit, and a semantic candidate used when searching the document When learning data is created in pairs and machine learning using the learning data gives a feature vector for classification having the same configuration as the feature vector for learning, the feature vector for the classification 2. The natural language according to claim 1, further comprising: means for automatically creating a predetermined classifier that selects an optimum meaning of the certain word in the context corresponding to the meaning candidate set from the meaning candidate set. Ambiguity resolution device.   The meaning candidate selecting means selects a predetermined number of sets having a large number of documents included in the set from the set of documents collected for each of the combinations by the document collecting means, The ambiguity resolution device in natural language according to claim 1 or 2, comprising means for performing processing for selecting only a corresponding semantic candidate as the semantic candidate.   The meaning candidate selecting means selects a set in which the number of documents included in the set is larger than a predetermined threshold among the set of documents collected for each of the combinations by the document collecting means, and the documents The ambiguity resolution device in natural language according to claim 1 or 2, further comprising means for performing a process of selecting only the semantic candidates corresponding to the set of semantic candidates as the semantic candidates.   The means for collecting includes, for each combination of a certain word and a semantic candidate in the semantic candidate set, a word that constitutes the combination from a virtual corpus including web pages existing on the Internet. 2. The ambiguity resolution device in natural language according to claim 1, further comprising means for searching a set of web pages to be generated and collecting them as a set of elements having an upper limit of a predetermined constant. A computer program that, when executed by a computer, causes the computer to function as an ambiguity resolver in natural language according to any one of claims 1 to 5.

Images