JP2019012457A - Information processing device, information processing method, and program - Google Patents

Abstract

To provide an information processing device, an information processing method, and a program for supporting realization of accuracy improvement of word category classification.SOLUTION: A classification model is learned that includes: extracting words from a plurality of sentences; determining, based on a category dictionary including information on categories corresponding to the words, the category of the words registered in the category dictionary among the extracted words; acquiring, based on a plurality of sentences, a distributed representation of words whose categories have been determined; and classifying, based on the acquired distributed representation and the determined category, categories of words represented by distributed representation.SELECTED DRAWING: Figure 5

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program.

In a chat bot or the like that automatically responds to an inquiry from a user, natural language processing technology is used to learn from text data such as FAQs and inquiry correspondence logs, and to interpret the meaning of sentences. In such natural language processing, a technique for analyzing the structure (morpheme, dependency structure, etc.) and meaning (predicate term structure, semantic role, semantic frame, etc.) of the sentence is required. For example, in the semantic analysis system, a category is assigned to a phrase obtained by performing morphological analysis and dependency analysis of an input sentence, and then a category is assigned to the phrase, and then the category information of the predicate and the related phrase is used as a key. Some predicate term structure analysis is performed with reference to the semantic frame dictionary, and then the meaning of the predicate (semantic frame) is identified and the semantic roles of the terms in the relationship are given. However, when a word that is not registered in the category dictionary is included in the sentence, there is a problem that the semantic category cannot be assigned and the semantic role cannot be assigned because the category of the phrase that is a key cannot be determined as it is. In other words, the range of sentences that can be semantically analyzed is limited by the range of words registered in the category dictionary. Proper determination of word categories is an important step.
If the number of words registered in the category dictionary is increased, the category of the word can be determined more correctly, but manual word registration requires specialized knowledge regarding the meaning of the word. In an area (domain) in which the number of technical terms is large or increases day by day, it takes labor to create a category dictionary by hand. Therefore, a method of increasing the number of registered words by extracting synonyms of words registered in the category dictionary and designating the same category as the words for the synonyms can be considered.
As a synonym extraction technique, for example, Patent Document 1 discloses a technique for extracting a synonym of an input word based on word co-occurrence data. Since the technique disclosed in Patent Document 1 is based on the co-occurrence frequency of an input word and a synonym candidate, there is a problem that a word that co-occurs although it is not related semantically may be determined as a synonym, and accuracy is low. There is.
Further, in Patent Document 2, two words of an arbitrary word set are learned by assigning a label to a feature vector generated with respect to a word set extracted from a text with reference to a synonym dictionary and learning. A technique for determining whether or not is a synonym is disclosed. The synonym determination technique disclosed in Patent Document 2 is a kind of supervised learning, and a large amount of correct answer data is required to improve accuracy. Further, since the synonym dictionary is data in which word pairs that are synonyms are stored, there is a problem that the category dictionary cannot be used as correct answer data.

Japanese Patent No. 5611173 Japanese Patent No. 5356197

  Due to the problems described above, the category of a word may fail because the category of the word cannot be determined or an incorrect category is determined. Therefore, an object of the present invention is to support improvement in accuracy of word category classification.

  Therefore, the information processing apparatus of the present invention is based on an extraction unit that extracts a word from a plurality of sentences and a category dictionary that includes information on a category corresponding to the word, and among the words extracted by the extraction unit, Determination means for determining a category of a word registered in a category dictionary, acquisition means for acquiring a distributed representation of a word whose category is determined by the determination means based on the plurality of sentences, and acquisition by the acquisition means Learning means for learning a classification model for classifying the category of the word represented by the distributed expression based on the determined distributed expression and the category determined by the determining means.

  According to the present invention, it is possible to support improvement in accuracy of word category classification.

FIG. 1 is a diagram illustrating an example of a hardware configuration of the information processing apparatus. FIG. 2 is a diagram illustrating an example of a category dictionary. FIG. 3 is a diagram illustrating an example of a semantic frame dictionary. FIG. 4 is a diagram illustrating an example of a functional configuration of the information processing apparatus. FIG. 5 is a flowchart illustrating an example of processing of the information processing apparatus. FIG. 6 is a flowchart illustrating an example of processing of the information processing apparatus. FIG. 7 is a flowchart illustrating an example of processing of the information processing apparatus. FIG. 8 is a diagram illustrating an example of a relationship between words in the semantic space. FIG. 9 is a diagram illustrating an example of the relationship between words.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
<Embodiment 1>
(Outline of processing of this embodiment)
An overview of the processing of this embodiment will be described. In the present embodiment, it is assumed that the information processing apparatus 100 is a main subject of processing. The information processing device 100 is an information processing device such as a personal computer (PC), a server device, a tablet device, or a smart phone.
In this embodiment, the information processing apparatus 100 extracts a word in a sentence from corpus data including a plurality of ticket data in the problem management system. The corpus data may include text other than ticket data. Then, the information processing apparatus 100 refers to the category dictionary and determines a category for the words registered in the category dictionary among the words extracted from the corpus data. A category is a classification for classifying the characteristics of words. Then, the information processing apparatus 100 acquires a distributed representation of each word whose category has been determined. Then, the information processing apparatus 100 learns a category classification model used for category classification based on the distributed representation of each word for which the category has been determined and the category of each word for which the category has been determined. The category classification model is, for example, a classifier that accepts a distributed expression of a word as an input and returns the category of the word as an output.
The information processing apparatus 100 uses the learned classifier to determine a category of words that are not registered in the category dictionary among words extracted from the corpus data. Thereby, the information processing apparatus 100 can determine a category also for a word that is not registered in the category dictionary among words extracted from the corpus data.

Further, the information processing apparatus 100 detects a word to which an inappropriate category is assigned for each word for which a category has been determined. The information processing apparatus 100 specifies the word A and the word B that are similar to each other. Then, the information processing apparatus 100 specifies a word C having a certain relationship with the word A and a word D having a similar relationship to the word B. The information processing apparatus 100 detects that the categories of the words C and D may be inappropriate based on the positional relationship in the semantic space of the distributed expressions of the identified words A, B, C, and D. .
The information processing apparatus 100 corrects the categories for the words C and D, and learns the category classification model again based on the words whose categories are corrected.

(Hardware configuration of information processing device)
FIG. 1 is a diagram illustrating an example of a hardware configuration of the information processing apparatus 100.
The information processing apparatus 100 includes a CPU 101, a main storage device 102, an auxiliary storage device 103, a network I / F 104, and an input / output I / F 105. Each element is communicably connected to each other via a system bus 106.
The CPU 101 is a central processing unit that controls the information processing apparatus 100. The main storage device 102 is a storage device such as a random access memory (RAM) that functions as a work area for the CPU 101 or a temporary storage location for data.

The auxiliary storage device 103 is a storage device that stores various setting information, various programs, corpus data, teacher data, various dictionary data, various model information, various threshold information, and the like. In the present embodiment, the auxiliary storage device 103 stores corpus data including a plurality of ticket data in the assignment management system that is text data. The auxiliary storage device 103 stores category dictionary data. A plurality of words are registered in the category dictionary data, and information on the category of each registered word is included. The category dictionary data is, for example, table format data indicating correspondence between words and categories. FIG. 2 is a diagram illustrating an example of category dictionary data. In the example of FIG. 2, the word item indicates each registered word, and the category item indicates the category of the corresponding word.
The auxiliary storage device 103 stores semantic frame dictionary data. In the semantic frame dictionary data, a plurality of predicates and clause categories related to the predicates are registered, and semantic frames (semantic roles and meanings) corresponding to the registered “predicates + clause categories related to the predicates”. Etc.). The semantic frame dictionary data is, for example, data in a table format indicating correspondence between predicates, clause categories related to the predicates, and semantic frames. FIG. 3 is a diagram illustrating an example of semantic frame dictionary data. In the example of FIG. 3, the predicate item indicates a registered predicate, and the “term” item indicates clause information related to the corresponding predicate. The “category” item indicates the category of the word in the corresponding phrase, the “case” item indicates the particle following the word in the corresponding phrase, and the “semantic role” item indicates the corresponding phrase. Indicates the semantic role of In addition, the item “predicate meaning” indicates the meaning of the corresponding predicate. The auxiliary storage device 103 is configured as a storage medium such as a read only memory (ROM), a hard disk drive (HDD), a solid state drive (SSD), or a flash memory.
The network I / F 104 is an interface used for communication with an external device via a network such as the Internet or a LAN. The input / output I / F 105 is an interface used for inputting information from an input device such as an operation unit of a mouse, a keyboard, or a touch panel. The input / output I / F 105 is an interface used to output information to an output device such as a display, a display unit of a touch panel, and a speaker.
When the CPU 101 executes processing based on a program stored in the auxiliary storage device 103, functions of the information processing apparatus 100 described later with reference to FIG. 4 and processing of flowcharts described with reference to FIGS. .

(Functional configuration of information processing device)
FIG. 4 is a diagram illustrating an example of a functional configuration of the information processing apparatus 100.
The information processing apparatus 100 includes an acquisition unit 401, a learning unit 402, an extraction unit 403, and an update unit 404.
The acquisition unit 401 extracts words from sentences in the corpus data stored in the auxiliary storage device 103. Then, the acquisition unit 401 acquires a distributed representation of the extracted word based on the corpus data.
The distributed expression is a technique for expressing a word with a real vector of a plurality of dimensions (for example, 100 to 300 dimensions). There is a distribution hypothesis that the meaning of a word in a sentence is determined from surrounding words (context words). Given the distribution hypothesis, a word can be represented as a vector where each element indicates the appearance probability of each context word. Since the number of context words is enormous (1 trillion or more), the size of this vector also becomes enormous (1 trillion or more dimensions), and cannot be stored in a PC memory or the like. However, most of the elements of this vector are zero. Therefore, this vector can be compressed (for example, compressed to a size of 100 to 300 dimensions). In the distributed expression, the word is expressed as a vector compressed in this way, on the premise of the distribution hypothesis.

The words expressed in the distributed representation become closer vectors as the meaning of the words is closer. Due to this property, the vector indicated by the distributed representation of the word extracted by the acquisition unit 401 becomes a closer vector as the meaning is closer.
In addition, the words expressed in the distributed expression also have the following properties. That is, a vector indicating the difference between the word (1) and the word (2) as the relationship between the word (1) and the word (2) and the relationship between the word (3) and the word (4) are similar. And the vector indicating the difference between the word (3) and the word (4) is a close vector. In the processing described later with reference to FIG. 7, the information processing apparatus 100 uses this property to detect words whose categories are not correctly classified.

Based on the category dictionary data stored in the auxiliary storage device 103, the learning unit 402 determines the category of the word for which the distributed representation has been acquired by the acquisition unit 401. At this time, the learning unit 402 cannot determine a category for words that are not registered in the category dictionary data. Then, the learning unit 402 learns a category classification model used for classification of the category of the word represented by the distributed expression based on the distributed expression of the word for which the category has been determined.
The learning unit 402 learns a semantic relationship classification model used for classification of semantic relationships between words. A semantic relationship between words is a relationship between the meaning of one word and the meaning of the other word. The semantic relationship between words includes, for example, a superordinate concept / subordinate concept relationship (for example, a relationship between “country” and “Japan”), a synonymous relationship (for example, “Japan” which is a subordinate concept of “country”) And "US"), inclusion relations (for example, "car" and "tyre", one of which includes the other).
The extraction unit 403 extracts words that satisfy a specific condition from the corpus data. The extraction unit 403 extracts two words that are similar to each other from the corpus data, or extracts a word that has a semantic relationship with a certain word.
The update unit 404 updates the word category.

(Category classification model learning process)
FIG. 5 is a flowchart illustrating an example of processing of the information processing apparatus 100. An example of learning processing such as a category classification model used for classifying a word category will be described with reference to FIG.
In step S <b> 501, the acquisition unit 401 extracts words from the corpus data stored in the auxiliary storage device 103.
In S <b> 502, the acquisition unit 401 determines a category of a word registered in the category dictionary among the words extracted in S <b> 501 based on the category dictionary stored in the auxiliary storage device 103.
In step S <b> 503, the acquisition unit 401 acquires a distributed representation of all the words whose categories are determined in step S <b> 502 based on the corpus data stored in the auxiliary storage device 103. For example, the acquisition unit 401 acquires a distributed representation of a word by unsupervised learning (word2vec or the like) based on a word in sentence data in corpus data and a word around the word.
In S504, the learning unit 402 learns a category classification model based on the distributed representation acquired in S503 and the category of the word whose category is determined in S502.
For example, the learning unit 402 uses the distributed representation acquired in S503 and the categories corresponding to the respective distributed representations acquired in S503 as teacher data, inputs the distributed representation of the word, and outputs the category of the word. The classifier is learned as a category classification model. The learning unit 402 stores this teacher data in the auxiliary storage device 103. The learning unit 402 uses, for example, naive Bayes or a support vector machine (SVM) as a discriminator.
In the present embodiment, the learning unit 402 learns the category classification model by using the distributed representation of all the words whose categories are determined in S502 and the corresponding categories in the processing of S504 as teacher data. However, the learning unit 402 learns a category classification model by using, as teacher data, a distributed representation of a predetermined number of randomly sampled words and corresponding categories among the words whose categories are determined in S502. It is good to do.

As described above, the information processing apparatus 100 has learned the category classification model used for classifying the categories of words represented in the distributed representation by the processing of FIG. Thereby, the information processing apparatus 100 can enable classification of the category of words expressed in a distributed expression. For example, the information processing apparatus 100 selects a word whose category is undetermined from words extracted from corpus data, and acquires a distributed representation of the selected word in the same manner as in S503. The information processing apparatus 100 can determine the category of the word by inputting the acquired distributed expression to the classification model learned in the process of FIG.
Further, the information processing apparatus 100 can assign a semantic role to a predicate or a word in a sentence based on the word for which the category is determined. The information processing apparatus 100 extracts a predicate and a clause related to the predicate from the sentence. Then, when the category of the word included in the extracted phrase is undecided, the information processing apparatus 100 acquires a distributed representation of the word, and inputs the acquired distributed representation into the learned classification model so that the word Determine the category. Then, the information processing apparatus 100, for example, in the extracted predicate or the extracted clause from the semantic frame dictionary of FIG. 3 based on the extracted predicate, the extracted clause, and the category of the word included in the clause. The meaning role and meaning of the word can be determined.

(Semantic relationship classification model learning process)
FIG. 6 is a flowchart illustrating an example of processing of the information processing apparatus 100. An example of learning processing such as a semantic relation classification model used for classification of semantic relations between words will be described with reference to FIG.
In step S601, the learning unit 402 determines whether to perform learning of the semantic relationship classification model. For example, the learning unit 402 determines whether to perform learning of the semantic relationship classification model based on information indicating whether to perform learning of the semantic relationship classification model stored in advance in the auxiliary storage device 103. judge. For example, the learning unit 402 learns the semantic relationship classification model when the information indicating whether to execute the learning of the semantic relationship classification model stored in the auxiliary storage device 103 is information indicating the execution. Is determined to be executed. For example, when the information indicating whether or not to learn the semantic relationship classification model stored in the auxiliary storage device 103 is information indicating that the learning unit 402 does not execute the semantic relationship classification model, It is determined that no learning is performed.
For example, the CPU 101 determines whether to execute learning of the semantic relation classification model stored in the auxiliary storage device 103 based on a user operation via an input device connected via the input / output I / F 105. The information shown can be updated. For example, when the semantic relation classification model learned in advance is stored in the auxiliary storage device 103, the user may think that learning of the semantic relation classification model is not required. In such a case, the user updates the input device so as to update the content of the information indicating whether or not the learning of the semantic relation classification model stored in the auxiliary storage device 103 is to be executed. The information processing apparatus 100 is instructed.

In S <b> 602, the acquisition unit 401 acquires a distributed representation of words in the corpus data by unsupervised learning (word2vec or the like) based on, for example, the words in the sentence data in the corpus data and the words around the words. .
In S603, the learning unit 402 extracts a pair of two words and a semantic relationship between the words from a semantic relationship dictionary (for example, WordNet).
In S604, the learning unit 402 acquires a distributed expression of two words included in all the pairs extracted in S603 from the distributed expression acquired in S602. Then, the learning unit 402 outputs the semantic relationship with the input of the two words, using the set of the distributed representation of the acquired two words and the semantic relationship included in the set extracted in S603 as teacher data. The classifier is learned as a semantic relation classification model. The learning unit 402 uses, for example, a support vector machine (SVM) as a discriminator.
As described above, the information processing apparatus 100 has learned the semantic relationship classification model used for classification of the semantic relationship between two words expressed in a distributed expression by the processing of FIG. Thereby, the information processing apparatus 100 can enable classification of the semantic relationship between two words expressed in a distributed expression. For example, the information processing apparatus 100 can acquire the semantic relationship between the words by inputting the distributed representation of the two words extracted from the corpus data to the semantic relationship classification model learned in the process of FIG. 6. .
In the present embodiment, the learning unit 402 performs the semantic relationship classification model using the distributed representations of the two words included in all the pairs extracted in S603 and the corresponding relationships as teacher data in the processing of S604. I decided to learn. However, the learning unit 402, in the process of S604, among the pairs extracted in S603, a distributed representation of two words included in a predetermined number of randomly sampled pairs, and the corresponding relationships, It is good also as learning a semantic relation classification model by using as teacher data.

(Identification of inappropriate words in categories)
When there is a word having an inappropriate category corresponding to the word included in the teacher data or when the number of words included in the teacher data is not sufficient, the information processing apparatus 100 uses the category classification model learned in the process of FIG. It is not possible to properly classify the categories of words represented in a distributed representation.
In such a case, among the words whose categories are classified using the category classification model, the category of the word whose category classification result is inappropriate is updated, and the category classification is performed again based on the words whose category is updated. By learning the model, the classification accuracy of the category classification model can be improved. For example, when the category of the word included in the teacher data is updated, the classification accuracy of the category classification model can be improved by learning the category classification model again based on the teacher data. Further, for example, when a category of a word not included in the teacher data is updated, a category classification model is learned based on the teacher data by using the teacher data plus the word as new teacher data. The classification accuracy of the category classification model can be improved.
Also, if the teacher data contains words with inappropriate category classification results, remove these words from the teacher data and learn the category classification model again. Improvement can be expected.
In order to perform such processing and improve the accuracy of word category classification, it is necessary to identify a word that may have an inappropriate category.
Therefore, in order to improve the accuracy of word category classification, there has been a demand to specify a word whose category may be inappropriate.

The distributed representation of a word has a property that, as the relationship between two words and the relationship between two other words are similar, the vectors indicating these two relationships in the semantic space of the distributed representation are similar. For example, if the relationship between the word “Japan” and the word “baseball” is similar to the relationship between the word “US” and the word “hockey”, the word “Japan” in the semantic space of the distributed expression The vector indicating the relationship with the word “baseball” is similar to the vector indicating the relationship between the word “US” and the word “hockey” in the distributed expression semantic space.
The meaning space of the distributed expression is a space in which each coordinate in the space represents a vector indicated by the distributed expression of the word. The vector indicated by each coordinate is a vector having the origin of the semantic space as the start point and the coordinate as the end point. The number of dimensions in the semantic space of the distributed representation is the same as the vector indicated by the distributed representation of the word. For example, if the number of dimensions in the semantic space of the distributed representation is 2, the coordinates (x, y) on the semantic space of the distributed representation indicate a vector (x, y). The vector indicating the relationship between two words in the semantic space of the distributed representation is a vector indicating what relationship the words are, for example, a vector indicated by one word and a vector indicated by the other word And a vector indicating the difference between.

From this property, the following can be assumed. For example, it is assumed that there are a word (X) and a word (Y) that are similar to each other. Furthermore, a word (A) having a relationship with the word (X) and a word (B) having a relationship similar to the relationship between the word (X) and the word (A) in the word (Y). Suppose there is. A vector indicating the relationship between the word (B) and the word (Y) on the semantic space of the distributed representation, a vector indicating the relationship between the word (A) and the word (X) on the semantic space of the distributed representation, Are similar. In this case, it can be assumed that the word (A) and the word (B) are words having the same relationship as the words similar to each other and belong to the same category.
In the present embodiment, the information processing apparatus 100 classifies categories for words included in the corpus data stored in the auxiliary storage device 103 using the category classification model learned in the process of FIG. Then, the information processing apparatus 100 may use an assumption based on the nature of the distributed representation in the process described later with reference to FIG. 7, and the category may be inappropriate from words included in the corpus data in which the category is classified. Identify words. Then, the information processing apparatus 100 updates the category for a word whose determined category may be inappropriate, and learns the category classification model again using the word whose category has been updated as teacher data.

FIG. 7 is a flowchart illustrating an example of processing of the information processing apparatus 100. With reference to FIG. 7, an example of processing for determining a word that may have an inappropriate category, updating the category of the determined word, and learning the category classification model again will be described. The word (X) in the process of FIG. 7 is an example of the first word. The word (Y) in the process of FIG. 7 is an example of a third word. The word (A) in the process of FIG. 7 is an example of the second word. The word (B) in the process of FIG. 7 is an example of a fourth word.
In step S <b> 701, the extraction unit 403 extracts two words that are similar to each other from the corpus data stored in the auxiliary storage device 103. For example, the extraction unit 403 extracts two words that are similar to each other by extracting two words having the same semantic relationship with respect to a predetermined word. For example, the extraction unit 403 extracts words having a semantic relationship with a predetermined word using the semantic relationship classification model learned in S604. For example, the extraction unit 403 inputs two words, a predetermined word and another word, into the semantic relationship classification model learned in S604, and acquires the semantic relationship between the two words. Then, when the acquired semantic relationship is a relationship indicating a synonymous relationship, the extracting unit 403 extracts the two words as two words that are synonymous with each other. Further, the extraction unit 403 may extract words having a semantic relationship with a predetermined word using, for example, a semantic relationship classification model learned in advance. In the present embodiment, the extraction unit 403 includes a predetermined word (“country”), a word (X) (“Japan”) having the same semantic relationship (superordinate concept / subordinate concept relationship), and a word (Y ) ("US").

In S702, the extraction unit 403 extracts, from the corpus data, a word (A) having a relationship with the word (X) extracted in S701. In the present embodiment, the extraction unit 403 extracts the word (A) (“baseball”) having a co-occurrence relationship with the word (X) (“Japan”) from the corpus data. For example, in the sentence included in the corpus data, the extraction unit 403 identifies what kind of word appears with what probability around the word (X). For example, the extraction unit 403 counts the number of appearances of each word that appears around the word (X) in the text included in the corpus data, and calculates the total number of appearances of the word (X) in the corpus data. By dividing by the number of appearances, the appearance probability of each word is acquired. Then, the extraction unit 403 extracts a word having a specified probability that is equal to or greater than a predetermined threshold as a word (A) having a co-occurrence relationship with the word (X).
Further, the extraction unit 403 may extract a word having a predetermined semantic relationship with the word (X), for example, using the semantic relationship classification model learned in S604. For example, the extraction unit 403 inputs the word (X) and another word to the semantic relationship classification model learned in S604, and the output semantic relationship is a predetermined semantic relationship. The word input to the semantic relationship classification model together with the word (X) is extracted as a word having a predetermined semantic relationship with the word (X). Examples of semantic relationships include inclusion relationships and synonymous relationships. For example, the extraction unit 403 may extract a word having a semantic relationship with the word (X) using a pre-learned semantic relationship classification model stored in the auxiliary storage device 103. Further, the extraction unit 403 may extract a word having a semantic relationship with the word (X) using, for example, a semantic relationship dictionary stored in advance in the auxiliary storage device 103.
Further, the extraction unit 403 may extract a word having a vector similar to the distributed expression (vector) of the word (X) as a distributed expression as a word (A) having a similar relationship with the word (X).
Hereinafter, the relationship between the word (X) extracted in S701 and the word (A) extracted in S702 is referred to as an expanded co-occurrence relationship. The expanded co-occurrence relationship is, for example, a relationship such as a co-occurrence relationship, a semantic relationship, or a similarity relationship of distributed expressions (vectors).
The process of S702 is an example of a first specifying process that specifies the second word.

In S703, the extraction unit 403 has a relationship similar to the expanded co-occurrence relationship between the word (X) and the word (A) between the word (Y) extracted in S701 from the corpus data. Then, the word (B) whose relationship with the word (Y) in the distributed representation semantic space is similar to the relationship between the word (X) and the word (A) in the distributed representation semantic space is extracted. In the present embodiment, the extraction unit 403 includes a vector indicating the relationship between the word (X) and the word (A) in the distributed expression meaning space (for example, a vector indicated by word (A) -word (X)), and the like. When the vector indicating the relationship between the word (Y) and the word (B) (for example, the vector indicated by the word (B) -the word (Y)) is similar, the word (X ) And the word (A) and the relationship between the word (Y) and the word (B) are similar.
Therefore, the extraction unit 403 includes a vector (for example, a vector indicated by the word (A) −the word (X)) indicating the relationship between the word (X) and the word (A) in the semantic space of the distributed representation, The word (B) is extracted so that the vector indicating the relationship between Y) and the word (B) (for example, the vector indicated by word (B) -word (Y)) is similar. The word (A) -word (X) vector is an example of a second relationship vector indicating the relationship between the first word and the second word in the distributed expression semantic space.

For example, the extraction unit 403 extracts the word (B) as follows. That is, for example, the extraction unit 403 extracts a word (B) having a co-occurrence relationship with the word (Y) from the corpus data. For example, in the sentence included in the corpus data, the extraction unit 403 specifies what word appears with a certain probability around the word (Y). For example, the extraction unit 403 adds up the number of appearances of each word that appears around the word (Y) in a sentence included in the corpus data, and calculates the total number of appearances of the word (Y) in the corpus data. By dividing by the number of appearances, the appearance probability of each word is acquired. Then, the extraction unit 403 extracts a word whose specified probability is equal to or greater than a predetermined threshold as a word (B) that has a co-occurrence relationship with the word (Y).
Further, the extraction unit 403 uses the semantic relationship classification model learned in S604, and a word having the same relationship as the relationship between the word (Y) and the word (X) and the word (A). May be extracted as a word (B) candidate. The extraction unit 403 may extract a word having a semantic relationship with the word (Y) using, for example, a previously learned semantic relationship classification model stored in the auxiliary storage device 103. The extraction unit 403 may extract a word having a semantic relationship with the word (Y) using, for example, a semantic relationship dictionary stored in advance in the auxiliary storage device 103. Further, the extraction unit 403 may extract a word having a vector similar to the distributed expression (vector) of the word (Y) as a distributed expression as a word (B) having a similar relationship with the word (Y).
Then, the extraction unit 403 determines whether the relationship between the word (B) candidate and the word (Y) in the distributed representation semantic space is the relationship between the word (X) and the word (A) in the distributed representation semantic space. Is identified as a word (B). For example, the extraction unit 403 uses a vector indicating the difference between the vector indicated by the word (A) and the vector indicated by the word (X) in the semantic space of the distributed expression as the word (X) and the word Obtained as a vector indicating the relationship with (A). Then, the extraction unit 403 acquires a vector obtained by adding the acquired vector and the vector indicated by the word (Y). This vector is a vector indicating a word whose relationship with the word (Y) in the distributed representation semantic space is the same as the relationship between the word (X) and the word (A) in the distributed representation semantic space.

The extraction unit 403 acquires the similarity between the acquired vector and the vector indicated by each word (B) candidate. The similarity between vectors is an index indicating the degree of similarity between the vectors, and includes, for example, a cosine similarity, an inner product, an absolute value of a vector indicating a difference between the vectors, and the like. In the present embodiment, the extraction unit 403 acquires the cosine similarity between this vector and the vector indicated by each candidate word (B).
Then, the extraction unit 403 identifies the acquired cosine similarity that is greater than or equal to a predetermined threshold. The extraction unit 403 specifies the maximum one when there are a plurality of items exceeding a predetermined threshold. The extraction unit 403 identifies, as the word (B), a word corresponding to the identified similarity among the candidates for the word (B). Thereby, the extraction unit 403 makes the vector indicating the relationship between the word (X) and the word (A) and the vector indicating the relationship between the word (Y) and the word (B) similar (that is, The word (B) can be specified such that the relationship between the word (X) and the word (A) is similar to the relationship between the word (Y) and the word (B).
In the present embodiment, it is assumed that the word specified as the word (B) is “hockey”.

FIG. 8 is a diagram illustrating an example of the relationship between the word (X), the word (Y), the word (A), and the word (B) in the semantic space of the distributed representation. A vector indicating the relationship between the word (X) and the word (A) (word (A) -word (X)) and a vector indicating the relationship between the word (Y) and the word (B) (word (B) -word (Y)) are similar to each other. As can be seen from FIG. 8, the word (X), the word (Y), the word (A), and the word (B) have an approximate parallelogram shape on the distributed expression semantic space. However, in the distributed representation semantic space, the four words having an approximate parallelogram shape are always in a relationship such as word (X), word (Y), word (A), and word (B). (Word (X) and word (Y) are similar, and there is a relationship (extended co-occurrence relationship) between word (X) and word (A). Word (Y) and word (B) It is not always true that there is a similar relationship between the two.
FIG. 9 is a diagram illustrating an example of the relationship between the word (X), the word (Y), the word (A), and the word (B). The word (X) and the word (Y) are subordinate concepts of the word “country”, and are shown to be similar to each other. Further, it is shown that the word (A) and the word (B) have a co-occurrence relationship with the word (X) and the word (Y), respectively.

The process of S703 is an example of a second specifying process that specifies the fourth word.
In S704, the extraction unit 403 determines whether the category of the word (A) extracted in S702 and the word (B) extracted in S703 are different.
If the extraction unit 403 determines that the categories of the word (A) extracted in S702 and the word (B) extracted in S703 are different, the categories of the word (A) and the word (B) are inappropriate. Is determined, and the process proceeds to S705. If the extraction unit 403 determines that the categories of the word (A) extracted in S702 and the word (B) extracted in S703 are the same, the categories of the word (A) and the word (B) are appropriate. As a result, the process of FIG. The processing of S704 is based on whether or not the category of the second word is different from the category of the fourth word, and whether or not the category may be inappropriate for the second word and the fourth word. It is an example of the determination process which determines these.

In step S <b> 705, the update unit 404 updates one or both categories of the word (A) and the word (B) that are determined to have an inappropriate category in step S <b> 704.
For example, when any one of the word (A) and the word (B) is determined, the update unit 404 updates the other category to the same category as the one category. For example, one of the word (A) and the word (B) is included in the dictionary data indicating the category of the word stored in advance in the auxiliary storage device 103, and the category matches the category indicated by the dictionary data. Then. In that case, the update unit 404 determines the one word as a word whose category has been determined, and updates the category of the other word to the same category as the category of the one word.

Further, the update unit 404 may perform the following processing. That is, the update unit 404 outputs information indicating that the category may be inappropriate for the word (A) and the word (B) to the output device connected via the input / output I / F 105. To do. For example, the update unit 404 indicates that there is a possibility that the category of the word (A) and the word (B) may be inappropriate on a display that is an output device connected via the input / output I / F 105. Output by displaying information. Thereby, the update part 404 can show that a category may be inappropriate about a word (A) and a word (B) with respect to a user. Furthermore, the update unit 404 may display, for example, a designation screen used for designation of categories for the word (A) and the word (B) on the display. The update unit 404 then selects one of the word (A) and the word (B) based on an operation on the designation screen by the user via the input device connected via the input / output I / F 105. Alternatively, specification of categories for both may be accepted. The update unit 404 updates either or both of the word (A) and the word (B) to the category indicated by the accepted designation.
The designation screen includes, for example, a category input field corresponding to the word (A) and the word (B). In this case, the update unit 404 can accept the designation of the category by acquiring information input in the category input field in the designation screen.

In S706, the learning unit 402 learns the category classification model again based on the word whose category has been updated in S705. If the word whose category has been updated in S705 is not included in the teacher data stored in the auxiliary storage device 103, the learning unit 402 performs category classification based on the teacher data and the word whose category has been updated in S705. Learn the model.
In addition, when the word whose category is updated in S705 is included in the teacher data stored in the auxiliary storage device 103, the learning unit 402 performs the following process. That is, the learning unit 402 updates the corresponding category in the teacher data for the word whose category is updated in S705 included in the teacher data, and learns the category classification model based on the updated teacher data.
The extraction unit 403 can also determine whether the categories of the word (X) and the word (Y) are different. When the extraction unit 403 determines that the categories of the word (X) and the word (Y) are different, the update unit 404 uses the same method as in S705 to indicate that the category (X) may have an inappropriate category. One or both categories are updated for the word and the word (Y). Then, the learning unit 402 learns the category classification model based on the word whose category has been updated.

(effect)
As described above, in the present embodiment, the information processing apparatus 100 has learned the category classification model used for classifying the categories of the words in the distributed representation based on the teacher data including the words whose categories are known. Thereby, the information processing apparatus 100 can support the improvement of the accuracy of the category classification of the word by enabling the classification of the category of the word represented by the distributed expression.
In addition, the information processing apparatus 100 determines a word that may have an inappropriate category from words classified into categories. Thereby, the information processing apparatus 100 can support realization of an improvement in accuracy of word category classification. For example, assuming that the words to be synonymous are in the same category and simply using a synonym dictionary, the following processing can be performed. That is, if two similar words (for example, word (X) and word (Y) in the present embodiment) are extracted and the categories of the extracted words do not match, the categories of these words may be inappropriate. This is a process for determining a unique word. However, with this method, if the two words do not have a direct synonym, these words cannot determine whether the category may be inappropriate. However, with the processing of this embodiment, the information processing apparatus 100 determines that the two words have the same relationship (extended co-occurrence relationship) with each of the two words having a direct synonym relationship. Even if there is no direct synonym, it can be determined whether there is a possibility that the category is inappropriate.

In addition, the information processing apparatus 100 updates the category of a word whose determined category may be inappropriate, and learns the category classification model again based on the word whose category has been updated. Thereby, the information processing apparatus 100 can improve the accuracy of classification based on the category classification model.
For example, a special word may be used or a word may have a special meaning in technical terms or company terms in a specific field (industry or business). In such a case, for example, the word is not recorded in the dictionary, or the meaning of the word is different from the meaning described in the dictionary. For example, the word “host” has the meaning of “master (who has customers)” and belongs to the category “people”. However, the word “host” in the IT field means “host computer (providing service to another machine)” and belongs to the category “thing”.
In such a case, there is a possibility that the category of the word cannot be properly classified by the category classification model learned based on the teacher data including the word and the category of the general term. Even in such a case, the information processing apparatus 100 determines a word that may have an inappropriate category, updates the category of the determined word, and again determines a category classification model based on the updated word. By learning, it is possible to improve the accuracy of classification of words related to a specific field by the category classification model.
In this way, the information processing apparatus 100 does not need to use new dictionary data or the like for improving the classification accuracy by the category classification model, and can reduce the cost of preparing new dictionary data or the like. In addition, even when the meaning content of a word changes with time and a new technical term is born or is used with a meaning different from the conventional one, the information processing apparatus 100 adapts to such a situation change, Additional categorization models can be learned.

(Modification)
In the present embodiment, the information processing apparatus 100 is a single information processing apparatus. However, the information processing apparatus 100 may be configured as a system including a plurality of information processing apparatuses connected to be communicable with each other via a network (LAN or the Internet). In that case, the CPU of each of the plurality of information processing apparatuses included in the information processing apparatus 100 executes the processing in cooperation with each other based on the program stored in the auxiliary storage device of each information processing apparatus. And the processing of the flowcharts of FIGS.
In the present embodiment, the information processing apparatus 100 identifies a word that may have an inappropriate category from words in the corpus data classified using the category classification model learned in the process of FIG. . However, the information processing apparatus 100 may identify a word that may have an inappropriate category from words that have a category set in advance.

In this embodiment, the extraction unit 403 adds the vector indicated by the word (A) to the vector indicated by the word (Y) in the semantic space of the distributed representation, and subtracts the vector indicated by the word (X) in S703. The word (B) is specified as a word corresponding to a similar vector. However, in S703, the extraction unit 403 obtains the word from the vector obtained by subtracting the vector indicated by the word (X) from the vector indicated by the word (A) in the semantic space of the distributed expression and the vector indicated by each of the candidates for the word (B). The word (B) may be specified based on the similarity with the vector obtained by subtracting the vector indicated by (Y). The vector obtained by subtracting the vector indicated by the word (Y) from the vector indicated by the candidate for the word (B) is an example of a first relation vector indicating the relationship between the third word and the candidate word in the distributed expression semantic space. is there.
For example, the extraction unit 403 extracts a word (Y) from a vector obtained by subtracting a vector indicated by the word (X) from a vector indicated by the word (A) in the semantic space of the distributed representation and a vector indicated by each of the candidates for the word (B) The degree of similarity with the vector obtained by subtracting the vector indicated by is acquired. Then, the extraction unit 403 may specify a word (B) candidate whose acquired similarity is equal to or greater than a predetermined threshold as the word (B). In addition, when there are a plurality of word (B) candidates whose acquired similarity is equal to or greater than a predetermined threshold, the extraction unit 403 may specify the word (B) that has the maximum similarity.

In the present embodiment, the information processing apparatus 100 determines a word that may have an inappropriate category by performing the processes of S701 to S704. It can be assumed that the words of the superordinate concept and the subordinate concept belong to the same category. Therefore, the information processing apparatus 100 may determine a word that may have an inappropriate category by performing the following process in addition to the processes of S701 to S704.
That is, the information processing apparatus 100 identifies a word of a higher concept or a lower concept of a word based on a semantic classification model. Then, the information processing apparatus 100 compares categories of the identified word and the original word, and if they are not the same, these words may be determined as words that may have inappropriate categories. Then, the information processing apparatus 100 updates the category in the same process as in S705 for the word determined as a word that may have an inappropriate category, and learns the category classification model based on the updated word. To do. Thereby, the information processing apparatus 100 can further improve the classification accuracy based on the category classification model.

As mentioned above, although preferable embodiment of this invention was explained in full detail, this invention is not limited to the specific embodiment which concerns.
For example, part or all of the functional configuration of the information processing apparatus 100 described above may be implemented in the information processing apparatus 100 as hardware.

100 Information processing apparatus 101 CPU
402 Learning Department

Claims (14)

Extraction means for extracting words from a plurality of sentences;
A determining unit for determining a category of a word registered in the category dictionary among the words extracted by the extracting unit based on a category dictionary including information on a category corresponding to the word;
Obtaining means for obtaining a distributed representation of a word whose category is determined by the determining means based on the plurality of sentences;
Learning means for learning a classification model for classifying a category of words represented by the distributed expression based on the distributed expression acquired by the acquiring means and the category determined by the determining means;
An information processing apparatus. A classification means for classifying a category of words included in a clause related to an input sentence predicate;
Using a semantic frame dictionary, assign meaning to the previous description word based on the previous description word, the phrase, and the category of the word classified by the classification means, and assign a semantic role to the phrase Granting means;
The information processing apparatus according to claim 1, further comprising: First specifying means for specifying a second word having a relationship with the first word;
The first word in the semantic space has the relationship with the third word similar to the first word and the relationship with the third word in the semantic space of the distributed expression And second specifying means for specifying a fourth word similar to the relationship between the first word and the second word specified by the first specifying means;
Based on whether the category of the second word specified by the first specifying means is different from the category of the fourth word specified by the second specifying means, the second word and Determining means for determining whether the category may be inappropriate for the fourth word;
When it is determined by the determining means that there is a possibility that the category is inappropriate for the second word and the fourth word, one or both of the second word and the fourth word Updating means for updating the category,
Further comprising
The information processing apparatus according to claim 1, wherein the learning unit learns the classification model again based on the word whose category has been updated by the updating unit. The semantic space is a space in which each coordinate on the semantic space indicates a vector indicated by a distributed representation of a word,
The second specifying means uses a word having the relationship with the third word as a candidate word, and a first relationship vector indicating a relationship between the third word and the candidate word in the semantic space is The candidate word is specified as the fourth word when similar to a second relation vector indicating a relation between the first word and the second word in the semantic space. Information processing device.   The second specifying means may be configured such that the first relation vector, which is a vector indicating a difference between a vector indicated by the third word on the semantic space and a vector indicated by the candidate word, is displayed on the semantic space. The candidate word is specified as the fourth word when similar to the second relation vector, which is a vector indicating a difference between a vector indicated by the first word and a vector indicated by the second word. 4. The information processing apparatus according to 4. The semantic space is a space in which each coordinate on the semantic space indicates a vector indicated by a distributed representation of a word,
The second specifying means uses the word having the relationship with the third word as a candidate word, the vector indicated by the candidate word on the semantic space, and the vector indicated by the third word from the first word If the similarity between the vector indicated by the word and the vector added by the vector indicated by the second word specified by the first specifying means is equal to or greater than a predetermined threshold, the candidate word is The information processing apparatus according to claim 3, which is specified as the fourth word.   Output means for outputting information indicating whether the category of the second word specified by the first specifying means is different from the category of the fourth word specified by the second specifying means; The information processing apparatus according to claim 3, further comprising: When the category of the second word specified by the first specifying means is different from the category of the fourth word specified by the second specifying means, the second word and the fourth word And further comprising an accepting means for accepting designation of one or both categories of words,
When the category of the second word is different from the category of the fourth word, the updating unit determines one of the second word and the fourth word based on the designation received by the receiving unit. The information processing apparatus according to claim 3, wherein the category is updated for both.   The updating means is different from the fourth word category specified by the second specifying means in that the category of the second word specified by the first specifying means differs from the second word and the category The category for the other of the second word and the fourth word is updated to the one category when a category is determined for one of the fourth words. 8. The information processing apparatus according to any one of 7 to 7. First specifying means for specifying a second word having a relationship with the first word;
The first word in the semantic space has the relationship with the third word similar to the first word and the relationship with the third word in the semantic space of the distributed expression And second specifying means for specifying a fourth word similar to the relationship between the first word and the second word specified by the first specifying means;
Based on whether the category of the second word specified by the first specifying means is different from the category of the fourth word specified by the second specifying means, the second word and Determining means for determining whether the category may be inappropriate for the fourth word;
An information processing apparatus. An information processing method executed by an information processing apparatus,
An extraction step of extracting words from a plurality of sentences;
A determination step for determining a category of a word registered in the category dictionary among the words extracted in the extraction step based on a category dictionary including category information corresponding to the word;
An acquisition step of acquiring a distributed representation of the word whose category is determined in the determination step based on the plurality of sentences;
A learning step of learning a classification model for classifying a category of words represented by a distributed expression based on the distributed expression acquired by the acquiring step and the category determined by the determining step;
An information processing method including: An information processing method executed by an information processing apparatus,
A first specifying step of specifying a second word having a relationship with the first word;
The first word in the semantic space has the relationship with the third word similar to the first word and the relationship with the third word in the semantic space of the distributed expression And a second specifying step of specifying a fourth word similar to the relationship between the first word and the second word specified in the first specifying step;
Based on whether the category of the second word specified in the first specifying step is different from the category of the fourth word specified in the second specifying step, the second word and A determining step for determining whether the category may be inappropriate for the fourth word;
An information processing method including:   A program for causing a computer to function as each unit of the information processing apparatus according to any one of claims 1 to 9.   The program for functioning a computer as each means of the information processing apparatus of Claim 10.

Images