JP6058563B2 - Model learning device, filter device, method, and program - Google Patents

Description

  The present invention relates to a model learning device, a filter device, a method, and a program, and more particularly, to a model learning device, a filter device, a method, and a program for determining whether or not an input sentence is eligible as an utterance sentence.

  There are roughly two types of dialogue systems: task-oriented dialogue systems and non-task-oriented dialogue systems. The former achieves a specific task by interaction with the system, and is used, for example, in a flight reservation system or a weather information retrieval system. Since these systems can assume what is spoken in advance, the system generates a utterance by maintaining a database of utterances created manually or by applying weather information extracted from the database to manual templates. (Non-Patent Document 1).

  A non-task-oriented dialogue system handles a dialogue with no purpose, and the content of the dialogue is a so-called chat. Since various topics are spoken in the chat, it is impossible to assume the content that is spoken in advance. Therefore, utterance generation is a very difficult task. In order to deal with a wide range of user input, recent technologies include a database of texts such as the web and Twitter (registered trademark), and a system utterance is selected by selecting the one similar to the user utterance ( Non-patent document 2, Non-patent document 3).

Ryuichiro Higashinaka, Katsuhito Sudoh, Mikio Nakano, `` Incorporating Discourse Features into Confidence Scoring of Intention Recognition Results in Spoken Dialogue Systems '', Speech Communication, Volume 48, Issues 3-4, pp.417-436, 2006. Shibata, M., Nishiguchi, T., and Tomiura, Y. (2009). ″ Dialog system for open-ended conversation using web documents. ″ Infomatica, 33 (3), pp. 277-284. Bessho, F., Harada, T., and Kuniyoshi, Y. (2012). ″ Dialog System Using Real-Time Crowdsourcing and Twitter Large-Scale Corpus. ″ In Proc. SIGDIAL, pp. 227-231.

  In a non-task-oriented dialog system, when a sentence such as web or Twitter (registered trademark) is extracted and used as a system utterance, a problem may occur as the utterance. For example, if a sentence depending on the context is extracted and is suddenly transmitted to the user as a system utterance, the meaning cannot be understood by a user who does not share the context. For example, even if the system extracts the sentence "It's fun to go" and tells the user about this, if the location information is not shared, the user can't understand this sentence because it doesn't know where to go. .

  In addition, even when extracting verbal statements that are often used as explanations of things and using them as system utterances, it is difficult for the user to take the meaning because there is no basic content of “what happened”. There's a problem. For example, even if the system utters “simple taste ramen”, it is difficult for the user to understand because “slight taste ramen” is unknown.

  In the sentence “It is fun to go,” the object “where” is missing. In short, the necessary grammatical elements are missing. “Simple taste ramen” lacks the “what” and “what” elements. The problem is that both lack grammatical elements.

  It is sufficient that such a sentence can be detected and filtered from the system utterance candidates, but there is a problem that it is difficult to automatically detect such a sentence. This is because sentences on the web, especially Twitter (registered trademark), often do not conform to the grammar of written words, such as omission of case particles, so grammatical analysis cannot often be performed properly, so It is difficult to detect if an element is missing.

  In addition, there is a problem that it is often difficult to determine whether the necessary grammatical elements are satisfied. This is because verbs require different grammatical elements depending on the situation in which they are used. If you can't live without eating, you don't need an object to eat, but if you want to eat, you can't understand the meaning without an object. “Shinjuku Station is too crowded and lost” is a noun phrase sentence that ends with a narrative, but the meaning of “what happened” is properly included, and there is no problem as an utterance.

  In the present invention, a model learning device, a method, and a method that can learn a model that is highly accurate as to whether or not it is a sentence that is qualified as an utterance sentence, which is made to solve the above problems. The purpose is to provide a program.

  It is another object of the present invention to provide a filter device, a method, and a program capable of accurately determining whether or not a sentence is qualified as an utterance sentence and outputting a sentence that is determined as an utterance sentence. And

  In order to achieve the above object, the model learning device according to the first invention is added with positive example information indicating that it is qualified as an utterance sentence or negative example information indicating that it is not eligible as an utterance sentence. For each input sentence that has been morphologically analyzed, a dependency analysis is performed, each word node representing the part of speech of the word corresponding to each word included in the input sentence is included, and an edge corresponding to the dependency relation of the word A tree structure connecting the word nodes, for each of the word nodes, a node representing a notation of a word corresponding to the word node, a node representing a standard notation of a word corresponding to the word node, and the word A dependency analysis unit for creating the tree structure in which at least one of nodes representing word termination forms corresponding to nodes is added as a child node of the word node; and the dependency analysis unit The tree structure based on a plurality of subtrees obtained from the tree structure created for each of the input sentences and positive example information or negative example information added to each of the input sentences And a model learning unit that learns a model for determining whether or not the sentence corresponding to is a sentence that is qualified as an utterance sentence.

  A model learning method according to a second invention is a model learning method in a model learning apparatus including a dependency analysis unit and a model learning unit, wherein the dependency analysis unit is qualified as an utterance sentence. Dependent analysis is performed on each input sentence that has been subjected to morpheme analysis to which positive example information or negative example information indicating that it is not eligible as a spoken sentence is added, and corresponding to each word included in the input sentence. Each word node representing the part of speech of the word, and a tree structure connecting the word nodes with edges according to the dependency relationship of words, and for each of the word nodes, a word corresponding to the word node At least one of a node representing a notation of a word, a node representing a standard notation of a word corresponding to the word node, and a node representing an end form of a word corresponding to the word node. The tree structure added as a child node of the input sentence, and the model learning unit includes a plurality of subtrees obtained from the tree structure created for each of the input sentences in the dependency analysis unit, and the input sentence Based on the positive example information or the negative example information added to each, a model for determining whether or not the sentence corresponding to the tree structure is a sentence suitable as an utterance sentence is learned.

  According to the first and second inventions, morphological analysis to which the dependency analysis unit is added with positive example information indicating that it is eligible as an utterance sentence or negative example information indicating that it is not eligible as an utterance sentence Dependency analysis is performed for each input sentence, a tree structure is created, and the model learning unit adds multiple subtrees obtained from the tree structure created for each input sentence to each input sentence. Based on the positive example information or the negative example information, a model for determining whether or not the sentence corresponding to the tree structure is a sentence suitable as an utterance sentence is learned.

  Thus, a tree structure was created for each input sentence to which positive example information indicating that it is eligible as an utterance sentence or negative example information indicating that it is not eligible as an utterance sentence was created. Based on a plurality of subtrees obtained from a tree structure and information on positive examples or information on negative examples, it is determined with high accuracy whether or not a sentence corresponding to the tree structure is an appropriate sentence as an utterance sentence. You can learn the model.

  In the model learning device according to the first aspect, the dependency analysis unit may include, for each of the word nodes, a node representing semantic information of a word corresponding to the word node and a word corresponding to the word node. The tree structure may be created by further adding at least one of nodes representing essential cases as a child node of the word node.

  A filter device according to a third aspect of the present invention performs dependency analysis on an input sentence that has been subjected to morphological analysis, includes each word node that represents a part of speech of the word corresponding to each word included in the input sentence, A tree structure connecting the word nodes with edges according to the dependency relationship, for each of the word nodes, a node representing a notation of a word corresponding to the word node, a standard of a word corresponding to the word node A dependency analysis unit for creating the tree structure in which at least one of a node representing a notation and a node representing a word end form corresponding to the word node is added as a child node of the word node; and the dependency analysis unit A plurality of sub-trees obtained from the tree structure created in step 1 and a sentence corresponding to the tree structure are determined in advance to determine whether or not the sentence is a qualifying sentence. Based on the model, a model application unit that calculates a score representing the degree that the input sentence is a sentence that is eligible as an utterance sentence, and based on the score calculated by the model application part, the input sentence is an utterance sentence And a filter unit that determines whether or not the sentence is qualified and outputs the input sentence that is determined to be a qualified sentence as an utterance sentence.

  A filtering method according to a fourth aspect of the present invention is a filtering method in a filter device including a dependency analysis unit, a model application unit, and a filter unit, wherein the dependency analysis unit is configured for an input sentence that has undergone morphological analysis. A tree structure that includes dependency analysis, includes each word node representing the part of speech of the word corresponding to each word included in the input sentence, and connects the word nodes with edges according to the dependency relationship of the words For each of the word nodes, a node representing a word notation corresponding to the word node, a node representing a standard word notation corresponding to the word node, and an end form of the word corresponding to the word node The tree structure is created by adding at least one of the representing nodes as a child node of the word node, and the model application unit is created by the dependency analysis unit The input sentence is based on a plurality of subtrees obtained from the notation structure and a pre-learned model that determines whether the sentence corresponding to the tree structure is a sentence that is qualified as an utterance sentence. A score representing the degree of qualification as an utterance sentence is calculated, and the filter unit determines whether or not the input sentence is qualification as an utterance sentence based on the score calculated by the model application unit. The input sentence determined to be an appropriate sentence as an utterance sentence is output.

  According to the third and fourth inventions, the dependency analysis unit performs dependency analysis on the input sentence that has been subjected to morphological analysis, creates a tree structure, and obtains the tree structure from the created tree structure by the model application unit. Based on a plurality of subtrees and a pre-learned model that determines whether the sentence corresponding to the tree structure is a sentence that is qualified as a spoken sentence, the input sentence is a sentence that is qualified as a spoken sentence A score representing the degree is calculated, and based on the calculated score by the filter unit, it is determined whether or not the input sentence is a sentence that is qualified as an utterance sentence. Is output.

  In this way, based on a plurality of subtrees obtained from the created tree structure and a pre-learned model that determines whether the sentence is qualified as an utterance sentence, by creating a tree structure for the input sentence The score is calculated, and based on the calculated score, it is determined whether the input sentence is an appropriate sentence as an utterance sentence. It is possible to output an input sentence that is determined to be a qualifying sentence.

  In the filter device according to the third aspect of the present invention, the dependency analysis unit includes, for each of the word nodes, a node representing semantic information of a word corresponding to the word node, and an essential word of the word corresponding to the word node. The tree structure in which at least one of nodes representing a case is further added as a child node of the word node may be created.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said model learning apparatus.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said filter apparatus.

  As described above, according to the model learning device, method, and program of the present invention, a tree structure is created for each input sentence to which positive or negative information is added, and the created tree structure Learns a model that determines with high accuracy whether or not a sentence corresponding to a tree structure is a sentence suitable as an utterance sentence, based on a plurality of subtrees obtained from be able to.

  Further, according to the filter device, method, and program of the present invention, a tree structure is created for an input sentence, and a plurality of subtrees obtained from the created tree structure and whether or not the sentence is qualified as an utterance sentence. The sentence is qualified as an utterance sentence by calculating a score based on a pre-learned model and determining whether the input sentence is an utterance sentence based on the calculated score. It is possible to output an input sentence determined to be a sentence that is qualified as an utterance sentence.

It is a block diagram which shows the functional structure of the model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the analysis result by JDEP. It is a figure which shows the example of the dependence structure between phrases. It is a figure which shows the example of the tree structure converted into the dependency structure between words. It is a figure which shows the example of the tree structure built centering on the part of speech node. It is a figure which shows the example of the tree structure which added the information of the semantic attribute and the essential case. It is a figure which shows the example of learning data. It is a figure which shows the example of the learned model. It is a block diagram which shows the functional structure of the filter apparatus which concerns on embodiment of this invention. It is a figure which shows the model learning process routine in the model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the filter processing routine in the filter apparatus which concerns on embodiment of this invention. It is a figure which shows the result of an experiment example. It is a figure which shows the example of the learning data at the time of using a score of 5 steps for the sentence to be input.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Principle of the present invention>
In the embodiment of the present invention, a machine learning technique is introduced to learn a classifier that distinguishes a sentence having a grammatical problem from a sentence having no grammatical problem. Specifically, a plurality of sentences extracted from the web or Twitter (registered trademark) are prepared, and a score indicating the suitability as an uttered sentence is given to these from a grammatical viewpoint. Then, a binary classifier is learned by using a high score as a positive example and a low score as a negative example. In learning, the result of syntactic analysis (specifically, the result of dependency analysis) is used as a feature value so that the grammatical relationship between words is included in the feature value.

<Configuration of Model Learning Device According to Embodiment of the Present Invention>
Next, the configuration of the model learning device according to the embodiment of the present invention will be described. As shown in FIG. 1, a model learning apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a model learning processing routine described later. Can be configured with a computer. Functionally, the model learning apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 adds positive example information to sentences that are extracted from the web, Twitter (registered trademark), etc., and qualifies as system utterances, and negative example information to sentences that are ineligible as system utterances. Each of the plurality of sentences is accepted.

  The calculation unit 20 includes a morphological analysis unit 30, a dependency analysis unit 40, a model learning unit 50, and a model storage unit 60.

  The morpheme analysis unit 30 performs morpheme analysis on each of the plurality of sentences received by the input unit 10. In the present embodiment, JTAG is used as a program for performing morphological analysis. Note that Chasen, Mecab, or the like may be used as a program for performing morphological analysis.

  The dependency analysis unit 40 performs dependency analysis (dependence structure analysis) for each of a plurality of sentences subjected to morphological analysis by the morpheme analysis unit 30. Specifically, for each of a plurality of sentences that have been morphologically analyzed, the sentences are grouped into clauses (a clause is a basic unit of Japanese consisting of a content word and a function word associated therewith). Determine dependencies. For example, for a sentence “I go to him and a movie”, four phrases “I”, “with him”, “to a movie”, and “go” are acquired from the result of the morphological analysis process. Then, by obtaining the dependency structure between the acquired phrases, it is possible to analyze that “I am”, “with him”, and “to the movie” are all structures related to “go”. In the present embodiment, JDEP is used as a program for performing dependency analysis. Here, JDEP is software that performs dependency analysis based on the output of JTAG. Cabocha or KNP may be used as a program for performing dependency analysis.

  Further, the dependency analysis unit 40 creates a tree structure for each of the plurality of sentences subjected to dependency analysis based on the structure of the dependency relationship of the clauses included in the sentence. Specifically, for each of the phrases included in the sentence, for each of the phrases included in the sentence, for each of the words included in the sentence, from the word, closest to the word, and within the phrase The word is located on the right side. Also, for each phrase, the rightmost word of the phrase is related to the word that is the main word in the clause at the destination. A word having no contact point (the last word in the sentence) is related to the root node (root) of the tree. Then, for each word included in the sentence, a node representing the part of speech of the word is created as a node representing the word (hereinafter referred to as a word node), and the word with an edge corresponding to the dependency relationship of the word Connect between nodes. For each word included in the sentence, a node representing each of the word notation, the standard notation, and the final form is added as a child node of the word node of the word.

  In addition, for each word included in the sentence, when a semantic attribute corresponding to the word exists, a node having information on the semantic attribute is added as a child node of the word node of the word. When there are a plurality of semantic attributes, nodes for all semantic attributes are added. Here, the semantic attribute refers to the semantic content of the word. A dictionary called Japanese vocabulary system stores correspondence between words and their semantic attributes. There are three types of semantic attribute information: general noun semantic attributes given to common nouns, proper noun semantic attributes given to proper nouns, and prescriptive meaning attributes given to predicates (mainly verbs). In the present embodiment, for each word included in the sentence for each sentence, semantic attributes corresponding to the words are acquired from the Japanese vocabulary system. The semantic attribute information is an example of word semantic information.

  Further, for each word included in the sentence, if the word is a predicate (mainly a verb or an adjective) and the word is present in the required case dictionary, a child node of the word node of the word As a result, a node having mandatory case information of the predicate is added. Here, the essential case dictionary is data holding a list of cases with predicates that can be constructed by analyzing large-scale text data. More specifically, large-scale text data is subjected to dependency analysis, and the frequency of case particles possessed by clauses related to predicates is counted. Then, case particles that frequently appear in a particular predicate are included in the essential case dictionary as essential cases of the predicate. A statistical test method may be used to determine whether or not a large number appears in a biased manner, and a chi-square test or a log likelihood ratio test may be used. In the required case dictionary used in the present embodiment, “dignity” is an indispensable case for “going”, “wo case” and “de case” are indispensable cases for “buy”, “ Is defined.

  Specifically, the case where the dependency analysis unit 40 creates a tree structure for the sentence “I went to the movie with him” will be described. FIG. 2 shows a dependency analysis result for the sentence in JDEP. The asterisk symbol in FIG. 2 represents phrase information. After the asterisk, there are a phrase ID, a relational phrase ID, a relation type (usually written as D, and a parallel dependency as P), a main word, and a function word main word. Focusing on the first line in FIG. 2, the phrase “I am” is related to the phrase whose phrase ID is 0 and the phrase ID is 3, the main word is the 0th word in the phrase (ie, “I”), function It can be seen that the main word is the first word in the phrase (ie, “ha”). The end of the sentence is indicated by an EOS (End of Sentence) symbol. Lines that do not start with an asterisk and are not EOS are lines for individual words and consist of six fields. They are notation, part of speech, standard notation, final form (in the case of part of speech with utilization), reading, and semantic attribute information (indicated in [] in the order of general noun attribute, proper noun attribute, and predicate attribute). FIGS. 3 to 6 respectively show dependency structures between clauses, tree structures representing dependency relationships between words, tree structures built around word nodes representing parts of speech, nodes representing semantic attributes, and essential elements. The example of the tree structure which added the node showing a case is shown. Here, the semantic attributes are represented by numbers. In FIG. 6, N, P, and Y are given as prefixes to distinguish the general noun attribute, proper noun attribute, and prescriptive attribute, respectively. ing.

  The finally obtained tree structure is a structure in which a rough structure of a sentence is represented by a word node representing a part of speech, and the detailed information is held as a child node. Also, by having semantic attribute information as a child node, the rough semantic content pointed to by the word is held. Furthermore, it has information for verifying whether the necessary case information is provided by having the case information that the predicate should originally have from the essential case dictionary.

  For each sentence to which positive example information or negative example information is given, the model learning unit 50 creates a tree structure of the sentence created by the dependency analysis unit 40 and the positive example given to the sentence. Based on the information or the negative example information, a model for determining whether or not the utterance sentence of the system is eligible is learned. In the present embodiment, BACT is used as an algorithm used for model learning. This algorithm enumerates the subtrees included in the tree structure in the given data, and calculates the weight of how much the subtree contributes to the determination of positive and negative cases by statistical processing It is. The model learning result finally obtained is a set of pairs of subtrees and their weights. BACT is an algorithm used for general tree structure determination (Non-Patent Document 4: Taku Kudo, Yuji Matsumoto (2004) A Boosting Algorithm for Classification of Semi-Structured Text, EMNLP 2004.). As the learning algorithm, another algorithm may be used as long as it can make the presence of a subtree in the tree structure a feature amount and give a weight to the feature amount.

  An example of learning data is shown in FIG. +1 representing the positive example is added to the head of the positive example data, and -1 is added to the negative example data. The tree structure is an S expression which is an input form of BACT. This S expression includes the same contents as the tree structure described in FIG. The part indicated by “case” represents essential case information. An example of a model obtained as a result of learning is shown in FIG.

  The first line shown in FIG. 8 is an intercept, which is a weight given to a determination target regardless of what subtree is held. Hereinafter, the left column is a weight, and the right column is a character string representation of the subtree having the weight. In adjacent elements of the character string expression, “)” represents a sibling relationship, and the others are parent-child relationships. For example, “verb stem: A noun” represents a subtree composed of a parent node representing “verb stem: A” and a child node representing “noun”. A positive weight indicates that the subtree is likely to be a positive example, and a negative weight is the opposite. Here, the created weight is stored as a model in the model storage unit 60 and output to the output unit 90.

  The model storage unit 60 stores a model that is learned by the model learning unit 50 and that determines whether or not the utterance sentence of the system is eligible.

<Configuration of Filter Device According to Embodiment of the Present Invention>
Next, the configuration of the filter device according to the embodiment of the present invention will be described. As shown in FIG. 9, the filter device 200 according to the embodiment of the present invention is a computer that includes a CPU, a RAM, and a ROM that stores a program and various data for executing a filter processing routine to be described later. Can be configured. Functionally, the filter device 200 includes an input unit 210, a calculation unit 220, and an output unit 290 as shown in FIG.

  The input unit 10 receives a sentence to be filtered.

  The calculation unit 220 includes a morphological analysis unit 230, a dependency analysis unit 240, a model application unit 250, a model storage unit 260, and a filter unit 270.

  The morpheme analysis unit 230 performs morpheme analysis on the sentence received by the input unit 210 using JTAG, similarly to the morpheme analysis unit 30 of the model learning device 100.

  The dependency analysis unit 240 performs dependency analysis (dependent structure analysis) using JDEP for the sentence subjected to the morphological analysis by the morphological analysis unit 230, similarly to the dependency analysis unit 40 of the model learning device 100. Also, the dependency analysis unit 240 creates a tree structure based on the dependency structure of clauses included in the sentence for the sentence subjected to dependency analysis, as with the dependency analysis unit 40 of the model learning apparatus 100. To do.

  The model application unit 250 uses a tree structure created by the dependency analysis unit 240 for the sentence received by the input unit 210 and a model for determining whether or not the utterance sentence of the system stored in the model storage unit 260 is appropriate. Based on this, a score indicating the degree of suitability of the sentence corresponding to the created tree structure as an utterance sentence of the system is calculated. Specifically, the subtrees included in the tree structure are enumerated, and according to the following formula (1), the model is referred to, and the weights of the respective subtrees are added together to qualify as an utterance sentence of the system. A score indicating the degree is calculated.

Here, t is the tree structure of the sentence, and score returns a score indicating the degree of qualification as an utterance sentence of the system regarding the tree structure. weight ₀ indicates an intercept, and subtrees is a function for enumerating subtrees from a tree structure. The weight is a function that refers to the model and returns the weight of the subtree.

  The model storage unit 260 stores the same model as the model that determines whether or not the utterance sentence of the system stored in the model storage unit 60 of the model learning device 100 is eligible.

  The filter unit 270 determines whether the score indicating the degree of qualification as the utterance sentence of the system calculated by the model application unit 250 for the sentence received by the input unit 210 exceeds a predetermined threshold value, and sets the threshold value. If it exceeds, it is determined that the sentence is eligible as an utterance sentence, and is output to the output unit 290. When the score indicating the degree of qualification as an utterance sentence of the system is equal to or less than the threshold value, the sentence is determined as an object to be removed and output to the output unit 290. According to the default setting in BACT learning, when the threshold value is set to 0, the determination performance is the highest, but the threshold value is set to a higher value when, for example, it is desired to increase the matching rate even if the recall rate is slightly reduced. Thus, it is possible to reduce a sentence that is ineligible as an utterance sentence of the system being determined as a sentence that is qualifying as an utterance sentence of the system. For example, when the threshold value is 0.005.

<Operation of Model Learning Device According to Embodiment of Present Invention>
Next, the operation of the model learning device 100 according to the embodiment of the present invention will be described. When each of the plurality of sentences to which the positive example information or the negative example information is given is received, the model learning device 100 executes a model learning process routine shown in FIG.

  First, in step S100, morphological analysis is performed using JTAG for each of a plurality of sentences to which positive example information or negative example information received by the input unit 10 is added.

  Next, in step S102, dependency analysis is performed using JDEP based on the morphological analysis result of the processing target sentence acquired in step S100.

  Next, in step S104, a tree structure is created based on the dependency analysis result of the processing target sentence acquired in step S102.

  Next, in step S106, it is determined whether or not the processing from step S102 to step S104 has been completed for all sentences accepted by the input unit 10. If the process from step S102 to step S104 is completed for all sentences, the process proceeds to step S108. If the process from step S102 to step S104 is not completed for all sentences, the process target is processed. Is changed to step S102.

  Next, in step S108, based on the tree structure for each of the plurality of sentences received in the input unit 10 acquired in step S104, and positive example information or negative example information given to each of the sentences. Then, a model for determining whether or not the utterance sentence of the system is eligible is learned and stored in the model storage unit 60.

  Next, in step S110, a model for determining whether or not the utterance sentence of the system acquired in step S108 is eligible is output to the output unit 90, and the process of the model learning process routine is terminated.

<Operation of Filter Device According to Embodiment of the Present Invention>
Next, the operation of the filter device 200 according to the embodiment of the present invention will be described. First, a model for determining whether or not the speech sentence of the system learned by the model learning device 100 is eligible is input from the input unit 210 and stored in the model storage unit 260. And if the sentence used as a process target is received, the filter apparatus 200 will perform the filter process routine shown in FIG.

  First, in step S <b> 200, a model for determining whether or not the utterance sentence of the system stored in the model storage unit 260 is eligible is read.

  Next, in step S202, as in step S100, the morphological analysis is performed on the sentence received by the input unit 210 using JTAG.

  Next, in step S204, as in step S102, dependency analysis is performed using JDEP based on the morphological analysis result of the processing target sentence acquired in step S202.

  Next, in step S206, similarly to step S104, a tree structure is created based on the dependency analysis result of the processing target sentence acquired in step S204.

  Next, in step S208, the utterance sentence of the tree structure system is qualified according to the above equation (1) based on the model acquired in step S100 and the tree structure of the sentence to be processed acquired in step S206. The score which shows the degree which is is calculated.

  Next, in step S210, it is determined whether or not the score indicating the degree of qualification as the system utterance sentence acquired in step S208 exceeds a predetermined threshold, and the degree of qualification as the system utterance sentence is determined. When the score indicating the value exceeds the threshold value, the process proceeds to step S214, and when the score indicates less than the threshold value, the process proceeds to step S212.

  Next, in step S212, the sentence received by the input unit 210 is determined as an exclusion target, output to the output unit 290, and the process ends.

  Next, in step S214, it is determined that the sentence received by the input unit 210 is eligible as an utterance sentence of the system, output to the output unit 290, and the process is terminated.

<Experimental example>
The graph of FIG. 12 shows the SVM (support vector) using the filter device 200 according to the embodiment of the present invention (Syntax in the drawing) and the feature of the word / part of speech N-gram (unigram, bigram, trigram) as a baseline. Machine) based filter performance comparison. FIG. 12 shows how the accuracy fluctuates when the threshold value of the filter unit is changed. The performance is better when the line is positioned at the top of FIG. As can be seen from FIG. 12, the precision (relevance rate) and recall (reproduction rate) are constantly higher than those of the filter device 200, and the effectiveness of the present invention was verified.

As described above, according to the model learning device according to the embodiment of the present invention, a tree structure is created for each input sentence to which positive or negative information is added, and the created tree structure A model that determines with high accuracy whether or not a sentence corresponding to a tree structure is an appropriate sentence as an utterance sentence of the system based on a plurality of subtrees obtained from Can learn.

  In addition, according to the filter device according to the embodiment of the present invention, a tree structure is created for an input sentence, and a plurality of subtrees obtained from the created tree structure and whether the sentence is qualified as a system utterance sentence A score indicating the degree of qualification as an utterance sentence of the system is calculated based on a pre-learned model to determine whether or not, and based on the score indicating the degree of qualification as an utterance sentence of the calculated system , By determining whether the input sentence is a sentence that qualifies as an utterance sentence of the system, it is determined with high accuracy whether or not it is a sentence that qualifies as an utterance sentence of the system. Can be output.

  Further, by using the present invention, since the utterance of the interactive system is easy for the user to understand, the system and the user can easily communicate with each other, and the interaction between the system and the user becomes smooth.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Further, in the present embodiment, the case where the information added to the sentence input to the model learning device 100 is binary information of positive example information or negative example information has been described, but the present invention is not limited to this. It is not something. For example, it is also possible to use a score such as 5 levels for an input sentence, and binarize it so that 1 or 2 points are negative examples and 4 or 5 points are positive examples. FIG. 13 shows an example in which a score is given by an external annotator.

  In the present embodiment, the case where a child node of a semantic attribute is added to a word node has been described. However, the present invention is not limited to this. For example, since it is only necessary to add information representing the meaning of a word as a child node, the Synnet ID in WordNet may be used. Moreover, you may add the number of the cluster allocated to each word obtained by clustering the several words in large-scale data according to the appearance method as a child node.

  An example of adding a node representing a word notation, a node representing a standard notation, a node representing an end form, a node having semantic attribute information, and a node having essential case information as child nodes of the word node However, the present invention is not limited to this. For example, as a child node of a word node, add at least one of a node representing the notation of the word, a node representing a standard notation, a node representing an end form, a node having semantic attribute information, and a node having essential case information You may make it do.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

DESCRIPTION OF SYMBOLS 10 Input part 20 Operation part 30 Morphological analysis part 40 Dependency analysis part 50 Model learning part 60 Model storage part 90 Output part 100 Model learning apparatus 200 Filter apparatus 210 Input part 220 Calculation part 230 Morphological analysis part 240 Dependency analysis part 250 Model Application unit 260 Model storage unit 270 Filter unit 290 Output unit

Claims (8)

Dependency analysis is performed on each of the input sentences that have been subjected to morphological analysis to which positive example information indicating that the sentence is qualified or negative example information indicating that the sentence is not eligible is added, and the input sentence is analyzed. Each of the word nodes representing the part of speech of the word corresponding to each of the words, and a tree structure connecting the word nodes with an edge according to the dependency relationship of the words, for each of the word nodes At least one of a node representing a notation of a word corresponding to the word node, a node representing a standard notation of a word corresponding to the word node, and a node representing an end form of the word corresponding to the word node A dependency analysis unit for creating the tree structure added as a child node of
Based on a plurality of subtrees obtained from the tree structure created for each of the input sentences in the dependency analysis unit, and positive example information or negative example information added to each of the input sentences. A model learning unit for learning a model for determining whether or not a sentence corresponding to the tree structure is a sentence suitable as an utterance sentence;
Model learning device including   The dependency analysis unit determines, for each of the word nodes, at least one of a node representing semantic information of a word corresponding to the word node and a node representing an essential case of the word corresponding to the word node of the word node. The model learning apparatus according to claim 1, wherein the tree structure further added as a child node is created. Dependent analysis is performed on the input sentence that has been subjected to morpheme analysis, each word node representing the part of speech of the word corresponding to each word included in the input sentence is included, and the word at an edge corresponding to the dependency relation of the word A tree structure connecting nodes, and for each of the word nodes, a node representing a notation of a word corresponding to the word node, a node representing a standard notation of a word corresponding to the word node, and the word node A dependency analysis unit that creates the tree structure by adding at least one of nodes representing the end form of the corresponding word as a child node of the word node;
A pre-learned model for determining whether or not a plurality of subtrees obtained from the tree structure created in the dependency analysis unit and a sentence corresponding to the tree structure are qualifying sentences as utterance sentences Based on the above, a model application unit that calculates a score representing the degree to which the input sentence is a sentence that is qualified as an utterance sentence,
A filter unit that determines whether or not the input sentence is a sentence that is qualified as a spoken sentence based on the score calculated by the model application unit, and that outputs the input sentence that is determined to be a sentence that is qualified as a spoken sentence When,
Including a filter device.   The dependency analysis unit determines, for each of the word nodes, at least one of a node representing semantic information of a word corresponding to the word node and a node representing an essential case of the word corresponding to the word node of the word node. 4. The filter device according to claim 3, wherein the tree structure further added as a child node is created. A model learning method in a model learning device including a dependency analysis unit and a model learning unit,
The dependency analysis unit performs dependency processing for each input sentence that has been subjected to morphological analysis and is added with positive example information indicating that it is eligible as a spoken sentence or negative example information indicating that it is not eligible as a spoken sentence. A tree structure including each word node representing the part of speech of the word corresponding to each word included in the input sentence, and connecting the word nodes with an edge according to a dependency relation of the word, , For each of the word nodes, a node representing a notation of a word corresponding to the word node, a node representing a standard notation of a word corresponding to the word node, and a node representing an end form of the word corresponding to the word node Creating the tree structure with at least one added as a child node of the word node;
The model learning unit includes a plurality of subtrees obtained from the tree structure created for each of the input sentences in the dependency analysis unit, and positive example information or negative examples added to each of the input sentences. A model learning method for learning a model for determining whether or not a sentence corresponding to the tree structure is a sentence that is qualified as an utterance sentence based on the information. A filter method in a filter device including a dependency analysis unit, a model application unit, and a filter unit,
The dependency analysis unit performs dependency analysis on an input sentence that has been subjected to morphological analysis, includes each word node that represents a part of speech of the word corresponding to each word included in the input sentence, and includes a dependency relation between words A tree structure connecting the word nodes with an edge corresponding to each of the word nodes, each of the word nodes representing a notation of a word corresponding to the word node, and a standard notation of a word corresponding to the word node Creating the tree structure by adding at least one of a node and a node representing an end form of a word corresponding to the word node as a child node of the word node;
The model application unit determines whether or not a plurality of subtrees obtained from the tree structure created in the dependency analysis unit and a sentence corresponding to the tree structure are qualifying sentences as utterance sentences. , Based on a pre-learned model, calculate a score representing the degree to which the input sentence is a sentence that is eligible as a spoken sentence,
The filter unit determines, based on the score calculated by the model application unit, whether or not the input sentence is a sentence that is qualified as a spoken sentence, and the input sentence that is determined to be a sentence that is qualified as a spoken sentence Output filter method.   The program for functioning a computer as each part which comprises the model learning apparatus of Claim 1 or Claim 2.   The program for functioning a computer as each part which comprises the filter apparatus of Claim 3 or Claim 4.