JP5426292B2 - Opinion classification device and program - Google Patents

Description

  The present invention relates to an opinion classification device that classifies an input sentence or input document written in a natural language into a predetermined category and a program thereof.

In order to efficiently analyze feedback information from users and viewers for content such as products, services and broadcasts, a technology that automatically classifies sentences written in natural language according to their opinion type is effective It is. In the conventional technique, an input document is divided into words, and the documents are classified by analyzing the distribution of the appearance frequency of those words.
This method was called the bag-of-words method and provided reasonable performance.

  The technique described in Patent Document 1 registers a positive expression representing a positive evaluation and a negative expression representing a negative evaluation based on the number of occurrences of a word in the text. The evaluation expression and the connection expression indicating the connection relationship are extracted from the text, the registered evaluation expression is detected from the extracted evaluation expressions, and the polarity with respect to the detected evaluation expression is determined.

Japanese Patent No. 3962382

The bag-of-words method described as background technology uses only the surface layer of divided words as a feature. Therefore, the first problem is that synonyms of words cannot be identified, and even if a feature to be classified with a certain expression is learned, another expression using synonyms cannot be understood, and classification performance does not improve. It was. Secondly, because a collocation (a lump of words consisting of a plurality of words and having a single meaning) cannot be identified, each individual word constituting a collocation is processed, and is unique as a lump such as an idiom. There is a problem that misclassification occurs when it has meaning, that is, when the meaning is not constructive. Third, because the word modification relationship cannot be identified, it is impossible to determine what the negative expression denies or which is more important when there are multiple candidate opinion predicates, and the classification performance is There was a problem of not going up.
That is, it is required to be able to determine whether or not a sentence belongs to a predetermined sentence category based not only on the statistical characteristics of the surface layer of the expression but also on the meaning.

  The present invention has been made based on the above problem recognition, and provides an opinion classification device that can classify sentences corresponding to synonyms. Further, the present invention provides an opinion classification device that can classify sentences corresponding to meanings such as idioms. It is another object of the present invention to provide an opinion classification device that can correctly classify a sentence even when a plurality of predicates are included or the sentence has a modification relationship. The present invention also provides these programs.

  [1] In order to solve the above problem, an opinion classification apparatus according to an aspect of the present invention is based on a morpheme analysis unit that reads input sentence data and performs morpheme analysis processing, and the input sentence data that has been subjected to morpheme analysis processing. A dependency analysis unit that performs dependency analysis processing and outputs parse tree data; and an opinion property predicate that includes a predicate category representing a type of the opinion property predicate detected from the parse tree data A predicate detection unit for adding information to the parse tree data, and a function representation of the input sentence data based on the input sentence data subjected to morpheme analysis processing by the morpheme analysis unit, and the detected function A functional expression detection unit that adds functional expression information including a functional expression type representing a type of expression to the parse tree data; and Is read and the functional expression detection unit adds the functional expression information, and whether the parse tree data corresponds to a predetermined sentence category using the characteristics of the read parse tree data And a determination unit for determining whether or not.

Here, the functional expression includes modalities of natural language and other functional expressions. Modality is a kind of functional expression.
According to the above configuration, the dependency analysis unit outputs data of the parse tree representing the sentence structure. Then, the predicate detection unit detects the predicate and adds information on the predicate category to the data of the parse tree. Therefore, the data of the parse tree stores not only the surface layer of the predicate, but also information on the predicate category representing the type of the predicate, and information on the position of the predicate in the sentence structure. Then, the function expression detection unit detects the function expression and adds information on the function expression type to the data of the parse tree. Therefore, the parse tree data holds information on the function expression type indicating the type of function expression, and also holds information on the position of the function expression in the sentence structure. The decision part reads the data of the parse tree including such abundant information and uses it as a feature. And a judgment part judges whether a sentence corresponds to a predetermined sentence category based on such a feature. That is, the determination unit classifies sentences into sentence categories as such.

As described above, by using the predicate category and the function expression type as features, classification based on the type of the predicate and the function expression can be performed as well as the surface layer of the expression. For example, a sentence including a synonym of a word can be classified more appropriately.
In addition, the parse tree holds information on word chunks such as collocations, so it identifies the word chunks, where they are located in the sentence structure, and what role they play Sentences can be classified using features such as.
In addition, since the parse tree also holds information on word modification relationships, not only the frequency of appearance of opinion opinion predicates, but also statements that take into account the role played by opinion property predicates in modification relationships. Classification can be performed.

  [2] Further, in the opinion classification device according to one aspect of the present invention, in the above configuration, the function expression detection unit detects a modality as a function expression of the input sentence data, and represents a type of the detected modality. The modality detection unit adds functional expression information including an expression type to the parse tree data.

  [3] Also, in the opinion classification device according to one aspect of the present invention, in the above configuration, the determination unit uses information on a subtree of the read parse tree data as the feature data. The machine comprises a model that has been machine-learned in advance with the parse tree data to which the opinion predicate information and the function expression information are added, and the sentence category corresponding to the correct example sentence, and refers to the model To determine whether the parse tree data corresponds to the predetermined sentence category.

When the above configuration is used, a correct example sentence and a sentence category corresponding thereto are given in advance. And a judgment part performs machine learning using a correct example sentence beforehand, and builds a model. And a judgment part classifies a sentence using the model constructed | assembled in that way.
With such a configuration, the device designer and the user do not need to analytically grasp the relationship between the feature space of the parse tree and the sentence category, and classification can be performed with high accuracy simply by giving examples.
As a machine learning method, for example, a tree kernel SVM is used.

  [4] A program according to an aspect of the present invention includes a morpheme analysis unit that reads input sentence data and performs morpheme analysis processing, and performs dependency analysis processing based on the input sentence data that has been subjected to morpheme analysis processing. A dependency analysis unit that outputs parse tree data; and an opinion predicate information including a predicate category representing a type of the detected opinion predicate in the parse tree data; A predicate detection unit to be added, and a function expression type representing a type of the detected function expression by detecting a function expression of the input sentence data based on the input sentence data subjected to morpheme analysis processing by the morpheme analysis unit A functional expression detection unit for adding functional expression information including parse tree data to the parse tree data; Judgment of reading the parse tree data to which the function expression information is added by the detection unit, and determining whether the parse tree data corresponds to a predetermined sentence category using features of the read parse tree data The computer is made to function as an opinion classification device having a section.

  [5] Further, in the program according to one aspect of the present invention, the function expression detecting unit detects a modality as the function expression of the input sentence data, and includes function expression information indicating a type of the detected modality. The computer is caused to function as an opinion classification device which is a modality detection unit for adding to the parse tree data.

  According to the present invention, it is possible to classify sentences accurately and efficiently based not only on the surface layer of the expression but also on the structure of the sentence and the role played by predicates and modalities in the structure.

It is the block diagram which showed the function structure of the opinion classification device by one Embodiment of this invention. It is the schematic which shows the data structural example of the dependency parsing tree which the dependency analysis part by the same embodiment outputs. FIG. 3 is a schematic diagram illustrating a parse tree corresponding to the parse tree data of FIG. 2. It is the schematic which shows the structure and example of data of predicate dictionary data by the embodiment. It is the schematic which shows the structure and data example of semantic category dictionary data by the embodiment. It is the schematic which shows the structural example of parsing tree data with which the predicate detection part added opinion property predicate information and the modality detection part added modality information by the embodiment. FIG. 7 is a schematic diagram illustrating a parse tree corresponding to the parse tree data of FIG. 6. It is the schematic which shows the parsing tree at the time of performing a dependence structure grammar analysis by the modification of the embodiment. It is a table which shows the correct rate of sentence classification | category by the verification experiment result of the opinion classification | category apparatus 1 by the embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a functional configuration of the opinion classification apparatus according to the embodiment. As shown in the figure, the opinion classification apparatus 1 includes a morphological analysis unit 11, a dependency analysis unit 12, a predicate detection unit 14, a modality detection unit 15 (functional expression detection unit), and an integrated determination unit 17 (determination unit). It is comprised including.

The morpheme analysis unit 11 reads input sentence data to be analyzed, performs morpheme analysis processing, and outputs a morpheme string as a result. Here, the input sentence data is text data including sentences and sentences written in a natural language such as Japanese. The morphological analysis process itself can be performed using various widely known methods. For example, using a character sequence as input, generate a lattice structure that is a graph of possible morpheme candidates while referring to morpheme dictionary data, and select the optimal path based on the part-of-speech information and the appearance frequency of n-grams. By searching, morphological analysis can be performed.
The dependency analysis unit 12 reads the result of the above morphological analysis, performs dependency analysis, and outputs the data of the dependency structure analysis tree as a result. Note that the dependency analysis process itself can be performed using various well-known methods.

  The predicate detector (predicate detector) 14 refers to the list of opinion-predicate (opinion-holding predicate) pre-registered in the dictionary data based on the data of the dependency structure analysis tree. Opinion predicates are detected by processing that takes into account word usage and collocation processing. The opinion predicate is a predicate having opinion. Then, the predicate detection unit 14 adds information regarding the detected opinion predicate to a corresponding node of the structural analysis tree. Here, the opinion predicate information added to the structural analysis tree includes a predicate category representing the type of detected opinion predicate.

  The modality detector 15 detects Japanese modalities based on the dependency structure analysis tree data and assigns meanings such as “deny”, “desire”, and “question”. Then, the modality detection unit 15 applies information related to the detected modality and other functional expressions (in the following description, the modality and other functional expressions may be simply referred to as “modalities”) to correspond to the structural analysis tree. Append to node. Here, the modality information added to the structural analysis tree (similarly, in the following description, information on the modality and other functional expressions may be simply referred to as “modality information”) is the type of the detected modality. (In the following description, the type of modality and other functional expression types may be simply referred to as “modality type”). As will be described later, the modality detection unit 15 does not need to use the dependency structure information, and it is sufficient to use the morpheme string information included in the data of the structural analysis tree as an input. That is, the modality detection unit 15 may directly use the result of the morphological analysis instead of reading the data of the structural analysis tree. In any of these cases, the modality detection unit 15 does not change the processing based on the input sentence data obtained as a result of the morphological analysis processing performed by the morphological analysis unit 11.

  The integrated judgment unit 17 uses the predicate detection unit 14 and the modality detection unit 15 respectively to obtain the data of the structural analysis tree to which the information on the opinion predicate (including the predicate category) and the information on the modality (including the modality type) are added. The data is read, and this is used as a feature (as a feature space) for classification, and as a result, sentence classification category (sentence category) data (classification result data) is output. The integration determining unit 17 performs processing using a tree kernel SVM (Support Vector Machine) technique. The tree kernel SVM is a machine learning technique, and a model learning process is performed in advance by giving a large number of correct answer case data to the integrated judgment unit 17.

  Next, details of main processes among the above-described units and data used for the processes will be described.

FIG. 2 is a schematic diagram illustrating a configuration example of the dependency structure analysis tree data output by the dependency analysis unit 12. As shown in the figure, this data is text data and is expressed using an S-expression. In the illustrated data example, the dependency structure analysis unit 12 outputs corresponding to the input sentence data “The theme is good, but the content is boring.”, And the entire S expression is the structure analysis for the input sentence. Represents a tree. The Japanese sentence corresponding to the above input sentence is “the theme is good but the content is boring”.
In this data, for example, the expression “(word but)” corresponds to a node of the word “but” in the structural analysis tree. Further, the expression “(chunk (word the) (word theme))” corresponds to a node of a chunk consisting of a plurality of words “The theme”. The expression (element 1 element 2 element 3) represents a structure in which the child nodes of the node of element 1 are the node of element 2 and the node of element 3 in this order from the left. The tree structure of the parse tree is expressed by nesting the relationship between the parent node and the child node in multiple layers.
Note that the line feed position in the S-expression data can be arbitrarily changed. In the present embodiment, the parse tree is expressed in the form of an S expression, but the parse tree may be expressed by another data structure.

FIG. 3 is a schematic diagram visually showing a parse tree corresponding to the data shown in the previous figure. In this figure, Japanese expressions corresponding to English sentences are shown at the same time. As shown in the figure, in this example, “ga / but” is a root node, and there are two child nodes “ha / is” and “ga / is” directly connected to the root node from the left side. Further, the child nodes “ha / is” have two child nodes, “theme / The theme” and “good / good” from the left side. Further, the child nodes “ga / is” have two child nodes “content / the content” and “boring / boring” from the left side.
The parsing trees shown in FIGS. 2 and 3 are obtained as a result of phrase structure grammar analysis by the dependency analysis unit 12.

The predicate detection unit 14 reads the data of the S expression corresponding to the parse tree and analyzes the data to detect the opinion predicate in the input sentence. At this time, the predicate detection unit 14 detects the opinion predicate by referring to the predicate dictionary data.
FIG. 4 is a schematic diagram illustrating a configuration of predicate dictionary data and a data example. As shown in the figure, the predicate dictionary data is configured as tabular data, and has two data items of opinion predicate and predicate category. The opinion item predicate data item stores character strings of words such as adjectives, adjective phrases, verbs, and verb phrases having opinion. The data item of the predicate category stores information for identifying the category corresponding to the predicate. In the example shown in the figure, an adjective entry “good” is associated with “category 2”, and an adjective entry “boring” is associated with “category 1”.
The content of the predicate dictionary data itself is prepared in advance. As an example, the number of entries in the predicate dictionary data may be around 500 entries, but more abundant entries may be prepared.
The predicate detection unit 14 adds information on the detected opinion predicate to the data of the parse tree. In other words, the predicate detection unit 14 performs processing for pasting information on opinion opinion predicates to the nodes of the parse tree. The resulting data will be described later.

  The modality detection unit 15 reads the data of the S expression corresponding to the parse tree and analyzes the data to detect the modality in the input sentence. However, the modality detection unit 15 does not use the structure of the parse tree for the modality detection process itself. The modality detection unit 15 obtains a sequence of words of the input sentence based on the read S-expression, and detects the modality using this sequence.

FIG. 5 is a schematic diagram showing a configuration of the semantic category dictionary data and a data example. As shown in the figure, the semantic category dictionary data is configured as tabular data, and has two data items of type and semantic category. In the data example shown in the figure, semantic categories such as request, recommendation, intention, desire,... Belong to the type “modality”. There are also other types of expressions similar to modalities, and these are collectively referred to as functional expressions. As shown in FIG. 5, the modality is one of the types included in the functional expression. Functional expressions other than modalities include connection method expressions, state-related, tense-related, modal-related, case-related, and other types, and the relationship between these types and the semantic categories they contain is also a semantic category dictionary. Stored in the data.
By referring to the semantic category dictionary data, the modality detection unit 15 can perform the modality and other functional expressions in the same manner. Therefore, in the following description, the detection of modality and the detection of other functional expressions are not particularly distinguished.
Note that one of “type” and “semantic category” shown in FIG. 5 or a combination of the two is a modality type (function expression type) representing the type of function expression.

  Modality and other functional expressions in Japanese are expressed as a sequence of auxiliary verbs following the verb in the sentence. In some cases, a sequence subsequent to a verb and including a verb or a noun expresses a modality. Such a sequence represents the modality, aspect, tense and state shown in FIG.

  To detect sentence modalities and other functional expressions correctly, it is necessary to resolve two types of ambiguity, meaning ambiguity and range ambiguity in Japanese. Semantic ambiguity is that an expression can represent a plurality of different modalities. Range ambiguity is a problem with Japanese as an agglutinating word, and there are multiple possible ranges of expressions following a verb (how to identify and delimit how far the sequence follows the verb). is there.

In order to solve this problem and correctly detect modalities and other functional expressions, the modality detection unit 15 first determines the above 2 based on the word sequence of the input sentence obtained from the output of the dependency analysis unit 12. Search all the possibilities caused by the kind of ambiguity and generate the lattice structure of the word based on the result. This process refers to the function expression dictionary data prepared in advance, and among the word sequences included in the input sentence, the sequence that matches the entry in the function expression dictionary data is regarded as a function expression candidate in the input sentence. This is done by extending the edge corresponding to the candidate. Here, the function expression dictionary data is data having a large number of entries with a word sequence that can be a function expression as one entry.
The modality detection unit 15 selects a stochastic optimum route (maximum likelihood route) from among a large number of routes (routes connecting from the beginning to the end of the sentence) included in the generated lattice structure. Edges included on the optimum route are extracted as function expressions in the input sentence.
Here, a Markov model is used as a probabilistic model for selecting the optimum route from the lattice structure. And as a feature, a surface layer character string, a part of speech, and a semantic attribute of a functional expression are used. In addition, a first-order Markov model that handles 2-gram (bigram) is used, and Katz back-off smoothing is used for a low-frequency connection probability. Note that the Markov model is learned in advance using correct data to which a correct answer is given manually.

Then, the modality detection unit 15 refers to the above-described semantic category dictionary data, and selects only the functional expressions detected above that correspond to the semantic category defined in the semantic category dictionary data. That is, by appropriately adjusting the semantic categories stored in the semantic category dictionary data in advance, the modality detection unit 15 can select only necessary function expressions according to the application. For example, the modality detection unit 15 may select all the function expressions, or may select only the modality, or an expression corresponding to an arbitrary subset of all the function expressions. You may make it select only. When the semantic category dictionary data illustrated in FIG. 5 is used, the modality detection unit 15 selects all the modalities and some of the functional expressions.
The modality detection unit 15 adds information on the detected modality to the data of the parse tree. In other words, the modality detection unit 15 performs a process of pasting information on the modality type and the semantic category to the node of the parse tree. The resulting data will be described next.

FIG. 6 is a schematic diagram showing data obtained as a result of the predicate detection unit 14 and the modality detection unit 15 adding information about opinionality predicates and information about modalities, respectively, to the parse tree data shown in FIG. is there.
As shown in the figure, information “(pred-category 1) (pred-surface good)” related to the opinion predicate is added by the predicate detection unit 14 to the “(word good)” portion of the original parse tree data. Thus, “(chunk (word good) (pred-category 1) (pred-surface good))” is obtained. Similarly, information of “(pred-category 2) (pred-surface boring)” relating to opinion predicates is added to the “(word boring)” portion of the original parse tree data by the predicate detection unit 14. , “(Chunk (word good) (pred-category 2) (pred-surface boring))”.
Similarly, the modality detection unit 15 adds the information “(modity-type but + determine)” to the “(word but)” portion of the original parse tree data and adds “(chunk ( word but (modity-type but + determinate)) ".

  FIG. 7 is a schematic diagram visually showing a parse tree corresponding to the data shown in the previous figure. As shown in the figure, modality information “meaning“ but ”+ determinate” is shown in the root node “ga / but”. This corresponds to “(modity-type but + determine)” in the previous figure. In addition, opinion predicate information “opinion-holding predicate” is shown in each of the nodes “good” and “boring”. These correspond to “(pred-category 1) (pred-surface good)” and “(pred-category 2) (pred-surface boring)” in the previous figure, respectively.

As the above data shows, the input sentence according to this example includes two opinions “good” and “boring”, which are mutually contradictory. Moreover, the words “good” and “boring” that represent these opinions appear in the sentence only once. Therefore, the simple “bag-of-words” method cannot properly classify this sentence.
On the other hand, the opinion classification device 1 according to the present embodiment uses the opinion property predicates, opinion property modalities, and functional expressions detected by the predicate detection unit 14 and the modality detection unit 15, and their positions in the structure of the syntax tree. Make an integrated decision based on this information.
For this purpose, the data of the S expression representing the parse tree in which the opinion predicate information and the modality information are pasted becomes the input to the integrated judgment unit 17.

Next, the process of the integrated determination unit 17 will be described in detail.
The integrated judgment unit 17 uses the parse tree to which opinion predicate information and modality information are added (hereinafter, the parse tree data is referred to as feature data for convenience) as input feature data. Are classified into predetermined sentence categories. For example, when the input sentence data is an opinion on a broadcast program, the integrated judgment unit 17 (1) a positive opinion (strong opinion), (2) a negative opinion (strong opinion), (3) watching the program What I thought (weak opinion), (4) What I learned by watching the program (weak opinion), (5) Requests for the program, (6) Questions about the program, (7) Other opinions (both above) The input sentence is classified into one of the eight sentence categories of (not applicable opinion) and (8) input sentence that is not an opinion. Here, the sentence categories are set to these eight, but the present invention is not limited to this, and in general, a set of arbitrary categories can be set.

As described above, the integrated determination unit 17 performs machine learning in advance using the tree kernel SVM technique, and classifies the feature data corresponding to the input sentence in a learned state.
Here, the process of classifying the feature data corresponding to the input sentence is reduced to a process of determining whether or not the feature data corresponds to each element belonging to the above-described category set.
The processing method adopted here is a kind of kernel method. In general, the kernel method is a method in which feature data (feature space) is mapped to the Euclidean space and the Euclidean space is partitioned by a hyperplane by applying SVM. In the present embodiment, it is this hyperplane that serves as a reference for determining whether or not individual feature data corresponding to an input sentence falls into each category. Here, the similarity between feature data is calculated as follows, assuming that it corresponds to the inner product between feature vectors in the Euclidean space. That is, the integrated determination unit 17 calculates the similarity S (T ₁ , T ₂ ) between the _first parse tree T ₁ and the second parse tree T ₂ using the following expressions (1) and ( Calculate according to 2).

Here, PT ₁ and PT ₂ are a set of all subtrees of the parse trees T ₁ and T ₂ , respectively. M (t _i , t _j ) is a function that returns a value according to whether or not the tree t _i and the tree t _j match. Here, the fact that the tree t _i and the tree t _j match each other means that they match completely, but a definition that the tree t _i and the tree t _j match may be given as appropriate.
That is, the similarity S (T ₁ , T ₂ ) is determined by the number of matching items between the subtree of the parse tree T _{1 and} the subtree of T ₂ .
As a specific process, the integrated determination unit 17 identifies all subtrees of the parse trees T ₁ and T ₂ in order to calculate the similarity S (T ₁ , T ₂ ), and brute force both parts. A process is performed to determine whether the elements match or do not match between the tree sets.

That is, the integration determining unit 17 uses data representing a set of subtrees of the read parse tree data as the feature data. Then, the integrated determination unit 17 performs machine learning in advance by using the parse tree data to which opinion predicate information and modality information are added to the given correct answer sentence and the sentence category (correct answer classification) corresponding to the correct answer sentence. Equipped with the model. That is, in the learning stage, the integrated judgment unit 17 is given a large amount of paired data consisting of analysis tree data obtained as a result of analyzing a case sentence and a correct answer of a category to which the case sentence belongs. At this time, the morphological analysis unit 11, the dependency analysis unit 12, the predicate detection unit 14, and the modality detection unit 15 may appropriately analyze the case sentence to obtain analysis tree data. And the integrated judgment part 17 performs machine learning by SVM using these correct answer data, and builds an internal model. Here, the model is physically a set of parameter values obtained as a result of the learning process and stored in the memory.
In a state where machine learning has been completed in advance, the integrated determination unit 17 refers to this model to determine whether or not the input sentence belongs to each category.

The above processing is summarized as follows. The predicate detection unit 14 detects an opinion predicate and outputs a predicate category corresponding to the predicate. Therefore, it is possible to make a determination using not only the surface character string of the predicate but also the predicate category as a feature. The modality detection unit 15 detects the modality and outputs the semantic category. Therefore, it is possible to make a judgment using the semantic category as a feature.
The opinion classification apparatus 1 including the integrated determination unit 17 as described above is information represented by the parse tree of the input sentence, that is, the surface layer of expressions (words and word sequences) included in the sentence, and sentence dependency. It is characterized by the type of opinion predicates contained in the sentence (predicate category), the type of modality contained in the sentence (modality type), and the position of those opinion predicates and modalities in the dependency structure of the sentence. While using it, it is necessary to make a judgment to identify the category to which the input sentence belongs by learning using cases without analytically determining the relationship between these features and sentence classification (category category). Can do.

  In this way, it is possible to detect meanings such as “denial”, “desire”, and “question” in the predicate in the sentence. For example, it is not enough to search only for the phrase “I want you to do it” when extracting a list of broadcast requests only from a large number of sentences. If you search for predicates containing, you can more appropriately search the list of requests. In this way, higher search performance and classification performance than the prior art can be obtained.

The opinion classification device 1 according to the present embodiment can be realized using an electronic circuit. Further, some or all of the functions of the apparatus may be realized by a computer.
When configuring a device using a computer, a program for realizing a necessary function is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed. It may be realized. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It is also possible to include those that hold a program for a certain time, such as a volatile memory inside a computer system serving as a server or client in that case. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  Although the embodiment has been described above, the present invention can also be implemented in the following modified example.

  <Modification 1> Modification 1 has a configuration in which the predicate detection unit 14 is excluded from the functional blocks illustrated in FIG. In this case, information about opinion predicates is not added to the nodes of the parse tree. However, the integrated determination unit 17 performs machine learning processing using data of the parse tree to which the modality information is added by the modality detection unit 15 and performs processing of classifying input sentence data using the learning result.

<Modification 2> In the above embodiment, the dependency analysis unit 12 performs a phrase structure grammar analysis and uses a parse tree obtained as a result. The parsing trees shown in FIGS. 2, 3, 6, and 7 are based on phrase structure grammar analysis. In the second modification, the dependency analysis unit 12 outputs parse tree data by performing dependency structure grammar analysis instead. FIG. 8 is a schematic diagram showing a parse tree obtained as a result of the dependency structure grammar analysis. As shown in the figure, this parse tree is composed of four nodes, “Theme is / The theme”, “Good is / good, but”, “The content is / the content”, and “Is boring / is boring”. ing. This parse tree has “is boring” as a root node, and has two child nodes “good but is good, but” and “content is / the content” connected to the root node. Has “the theme is / The theme” as a child node of “/ is good, but”. Note that opinion-predicate information (opinion-holding predicate) is added to each node of “good but is good, but” and “is boring”. In addition, function expression type information (meaning “BUT” + determinate) is also added to the node of “good but is good, but”.
In Japanese dependency analysis, it is also common to perform dependency structure grammar analysis. The tree structure obtained as a result of such dependency structure grammar analysis can also be expressed using S-expressions or other data structures.
Further, a parse tree having a different structure may be used by other analysis methods.
Regardless of which of these structures is used, each parse tree contains information that represents the syntax of the statement, as well as information on the predicate category and function expression type, and the position in the syntax of those predicates and function expressions. Contains information. Therefore, also in this modified example 2, similarly, the integrated determination unit 17 can appropriately classify the sentence.

  <Modification 3> In the above embodiment, the modality detection unit 15 detects not only the modality but also other functional expressions. In the third modification, the modality detection unit 15 detects only the modality. Therefore, the semantic category dictionary data used by the modality detection unit 15 stores only entries related to modalities, and does not store entries related to other functional expressions.

The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.
For example, in the above embodiment, the S-expression is used to represent the parse tree, but the parse tree may be represented using other types of data. For example, the parse tree may be represented by connecting the memory blocks corresponding to the nodes with pointers (data indicating the address of the point destination). Further, for example, a parsing tree may be expressed by using a relational database to hold the attributes of each node and the relationship between the nodes.

<Demonstration Experiment and Results> For the purpose of verifying the effectiveness of the present invention, an experiment for classifying an input sentence was performed by actually configuring the opinion classification apparatus according to the above-described embodiment using a computer.
FIGS. 9A to 9C are tables showing the correct rate of sentence classification based on the results of this experiment.
FIG. 9A shows the result of comparison with the method of the present invention using the bag-of-words method as a baseline. As shown here, the accuracy rate of the method of the present invention is higher both when the word surface layer, the predicate surface layer, and the predicate category are used as features, and when the predicate surface layer, the predicate category, and the modality type are used. Yes.
FIG. 9B shows the result of comparison with the method of the present invention using the case where only the surface layer of the word is used as a feature as a baseline. As shown here, the accuracy rate of the method of the present invention is higher. FIG. 9C also shows the result of comparison with the method of the present invention using only the surface layer of the word as a feature as a baseline. As shown here, the accuracy rate of the method of the present invention is higher.
Thus, the effectiveness of the technique of the present invention was demonstrated.

  The present invention can be used for the purpose of efficiently and accurately classifying a large amount of sentences written in a natural language. In particular, a business operator who sells a product or provides a service can use it for the purpose of efficiently and accurately classifying a large amount of opinion feedback fed back from the user. For example, it can be used in a broadcasting business.

DESCRIPTION OF SYMBOLS 1 Opinion classification apparatus 11 Morphological analysis part 12 Dependency analysis part 14 Predicate detection part 15 Modality detection part (functional expression detection part)
17 Integrated judgment part (judgment part)

Claims (5)

A morpheme analyzer that reads input sentence data and performs morpheme analysis processing;
A dependency analysis unit that performs dependency analysis processing based on the input sentence data that has been subjected to morphological analysis processing, and outputs parse tree data;
A predicate detector that detects an opinion predicate from the parse tree data and adds opinion predicate information including a predicate category representing the type of the detected opinion predicate to the parse tree data;
Based on the input sentence data subjected to morpheme analysis processing by the morpheme analysis unit, the function expression of the input sentence data is detected, and the function expression information including the function expression type indicating the type of the detected function expression is A function expression detection unit to be added to the parse tree data;
Reading the parse tree data to which the opinion predicate information is added by the predescript word detection unit and the function expression information being added by the functional expression detection unit, and the opinion predicate added to the read parse tree data A determination unit that determines whether or not the parse tree data corresponds to a predetermined sentence category using information and features of the function expression information ;
An opinion classification device you provided with a,
The determination unit includes a model that has been machine-learned in advance with the parse tree data to which the opinion predicate information and the function expression information are added and the sentence category corresponding to the correct case sentence for the correct case sentence. Determining whether or not the parse tree data corresponds to the predetermined sentence category by referring to the model;
An opinion classification device characterized by that. The function expression detection unit is a modality detection unit that detects a modality as a function expression of the input sentence data and adds function expression information including a function expression type representing a type of the detected modality to the parse tree data. ,
The opinion classification apparatus according to claim 1, wherein: The determination unit, Ru der those using information about the branch of the parse tree data read as the feature data,
3. The opinion classification apparatus according to claim 1 or 2, wherein A morpheme analyzer that reads input sentence data and performs morpheme analysis processing;
A dependency analysis unit that performs dependency analysis processing based on the input sentence data that has been subjected to morphological analysis processing, and outputs parse tree data;
A predicate detector that detects an opinion predicate from the parse tree data and adds opinion predicate information including a predicate category representing the type of the detected opinion predicate to the parse tree data;
Based on the input sentence data subjected to morpheme analysis processing by the morpheme analysis unit, the function expression of the input sentence data is detected, and the function expression information including the function expression type indicating the type of the detected function expression is A function expression detection unit to be added to the parse tree data;
Reading the parse tree data to which the opinion predicate information is added by the predescript word detection unit and the function expression information being added by the functional expression detection unit, and the opinion predicate added to the read parse tree data A determination unit that determines whether or not the parse tree data corresponds to a predetermined sentence category using information and features of the function expression information ;
A program that causes a computer to function as an opinion classification device comprising :
The determination unit includes a model that has been machine-learned in advance with the parse tree data to which the opinion predicate information and the function expression information are added and the sentence category corresponding to the correct case sentence for the correct case sentence. Determining whether or not the parse tree data corresponds to the predetermined sentence category by referring to the model;
program. The function expression detection unit is a modality detection unit that detects a modality as a function expression of the input sentence data and adds function expression information including a function expression type representing a type of the detected modality to the parse tree data. ,
The program according to claim 4, which causes a computer to function as the opinion classification device.

Images