JP6334587B2 - Word extraction apparatus, method, and program - Google Patents

Description

  The present invention relates to a word extraction device, method, and program, and more particularly, to a word extraction device, method, and program for extracting a word necessary for question answering from a sentence.

  2. Description of the Related Art Conventionally, a technique for extracting a word string close to an expression existing in a database from a sentence using a given sentence and a database storing knowledge is known. For example, in the case of a triple database structure, if it is found that the query sentence includes two expressions in the database, the remaining one can be presented as an answer.

  Moreover, when calculating the similarity of a word, the method of measuring a similarity after vectorizing a word in a semantic space is known (refer nonpatent literature 3).

K. Yao et al. “Recurrent Conditional Random Field for Language Understanding”, ICASSP2014 Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." ArXiv preprint arXiv: 1301.3781 (2013). Lafferty, John, Andrew McCallum, and Fernando CN Pereira. "Conditional random fields: Probabilistic models for segmenting and labeling sequence data." (2001).

  When there is learning data that is annotated in advance as to which database item is associated with a word string to be extracted in a sentence, the above word string is obtained by using a general-purpose machine learning method (see Non-Patent Document 1). Can be extracted. However, creating learning data corresponding to a database is expensive.

  In addition, for example, when obtaining the similarity of words individually, it is difficult to determine which features should be weighted at the time of learning, compared to the classification-based approach.

  Further, in Non-Patent Document 3, if the original domain unsupervised text and the target domain unsupervised text are expressed by different vector models, the vector space itself is completely different, so The weight to be used at the time of application may not be equivalent. It is also possible to use a large amount of unsupervised text unrelated to the domain, but because words are used in various meanings, if multiple domains are mixed, the vector space cannot be used appropriately in the target domain There is.

  Moreover, in question words and the like, there are cases where the character string in the sentence and the character string in the DB are not similar in terms of surface or semantics. For example, it is a case where “XYZ briefing session is <when>?” → “date”, and the surface layer and the semantics are not similar.

  Further, when the original domain and the target domain are far from each other, not only the above-mentioned problem can be solved but also data resources may be insufficient in the first place. To prepare for such a difficult situation, we want to obtain supplementary information that can be used universally in any domain.

  The present invention has been made in view of the above circumstances, and provides a word extraction device, method, and program capable of extracting a word corresponding to a question sentence when a database of unlearned domains is used. Objective.

  In order to achieve the above object, a word extraction device according to a first aspect of the present invention is a word representing an entry corresponding to a question sentence of an input target domain from a word indicating an entry in an unlearned target domain database. For each word included in the question sentence, for each table component of the database of the target domain, using a semantic vector of each word created in advance. Learning in advance from a set of question sentences in the original domain, to which the similarity of the word representing the entry is extracted as a feature, and the word representing the entry in the learned original domain database The sequence model for extracting the table constituent elements corresponding to the word chunks, and the quality by the sequence model feature conversion unit A sequence model applying unit that assigns a label representing the table component to each word chunk included in the question sentence based on the features of each table component extracted for each word included in the sentence; For each of the word chunks to which the label is given by the sequence model application unit using the semantic vector of each word, each of the words representing the entry of the table component represented by the label attached to the word chunk; Corresponding to a word learned in advance from a set of question sentences in the original domain, to which a feature representing a regression model is extracted as a feature and a word representing an entry in the original domain database A regression model for extracting a word representing an entry and the label extracted by the regression model featureizing unit The on the basis of the respective feature word chunks, corresponding to the question, is configured to include a, a regression model application unit for extracting a word representing an entry in the database of the target domain.

  Further, in the word extracting device according to the first aspect, the regression model feature adding unit is configured such that, for each of the word chunks to which the label is given by the sequence model applying unit, the label given to the word chunk is Extracting each of the at least one feature representing a similarity to each of the words representing the entry of the table component to represent, and representing the entry of the table component represented by the label for each of the at least one feature You may make it perform the relativity of the said feature showing the similarity with each of a word.

  In the word extracting device according to the first aspect, the semantic vector of each word may be created in advance based on the text of the target domain and the text of the original domain.

  Further, in the word extracting device according to the first invention, the regression model feature converting unit is similar to the word chunk for each of the word chunks to which the label is given by the sequence model applying unit. Similarity with a word included in a question sentence of a domain, similarity with a word representing the entry given to a word included in a question sentence of the original domain that is most similar to the word chunk, and the word chunk Combining the similarity between the most similar word representing the entry given to the word contained in the question sentence of the original domain and the word representing the entry of the table component represented by the label given to the word The similarity feature between domains may be further extracted.

  In the word extraction device according to the first aspect, the sequence model feature estimation unit estimates, for each word included in the question sentence, a question content of the question sentence or a superordinate concept of the word learned in advance. In order to further extract the estimation result of the question sentence or the estimation result of the word estimated using a domain-independent general-purpose model as a feature, the regression model featureization unit applies the series model For each of the word chunks to which the label is assigned by the unit, the estimation result of the question sentence or the estimation result of the word estimated using the general-purpose model may be further extracted as a feature.

  The word extraction method according to the second invention is a word extraction apparatus for extracting a word representing an entry corresponding to an inputted question sentence of a target domain from a word representing a database entry of an unlearned target domain. In the extraction method, the sequence model feature generating unit uses, for each word included in the question sentence, for each table component of the database of the target domain for each word included in the question sentence. The step of extracting the similarity with the word representing the entry of the table component as a feature, and the question sentence of the original domain to which the word representing the entry of the database of the learned original domain is assigned by the series model application unit A sequence model for extracting table components corresponding to word chunks learned in advance from a set of A label representing the table component is given to each word chunk included in the question sentence based on the feature for each table component extracted for each word included in the question sentence by the feature making unit And the regression model feature generating unit uses the meaning vector of each word to represent each of the word chunks to which the label has been added by the sequence model applying unit, as indicated by the label attached to the word chunk. A step of extracting the similarity with each of the words representing the entries of the table component as a feature, and a query of the original domain to which the regression model applying unit is given a word representing an entry of the database of the original domain A regression model for extracting a word representing an entry corresponding to the word learned in advance from a set of sentences; Extracting a word representing an entry in the database of the target domain corresponding to the question sentence based on the feature of each of the word chunks to which the label has been extracted, extracted by a feature unit for It is characterized by including.

  A program according to a third invention is a program for causing a computer to function as each part of the word extracting device according to the first invention.

  According to the word extraction device, method, and program of the present invention, the similarity between each word included in the question sentence and the word representing the entry is determined for each table component of the database using the semantic vector of each word. Based on the series model for extracting table features corresponding to words, extracted as features, and the extracted features of each word included in the question sentence, each word chunk included in the question sentence, Assigning a label representing a table component, and using each word's semantic vector, the similarity between each word chunk and each word representing an entry in the table component represented by the label attached to the word chunk, Regression model for extracting the word representing the entry corresponding to the word extracted as a feature and each extracted word chunk of the labeled word chunk Based on the above, it is possible to extract the word corresponding to the question sentence when the database of the unlearned domain is used by extracting the word representing the entry of the unlearned domain database corresponding to the question sentence. An effect is obtained.

It is an abstract figure showing the relationship of the table component in the database of the original domain of learning object, and the unlearned object domain. It is a block diagram which shows the structure of the model learning apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the similarity feature between domains at the time of learning. It is a figure which shows an example of relativity of a feature. It is a block diagram which shows the structure of the word extraction apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the similarity feature between domains at the time of application. It is a flowchart which shows the model learning process routine in the model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the feature-processing routine for series models which concerns on embodiment of this invention. It is a flowchart which shows the feature-processing routine for regression models which concerns on embodiment of this invention. It is a flowchart which shows the word extraction processing routine in the word extracting device which concerns on embodiment of this invention.

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

<Outline according to Embodiment of the Present Invention>

  First, an outline of the embodiment of the present invention will be described.

  Here, a problem to be solved by the embodiment of the present invention will be described. As shown in FIG. 1, during learning, the original domain to be learned is a toy domain. Input sentence at learning “How much is this book?”, Annotation result of input sentence at learning (eg, “Book⇒subject-picture book”, “How much⇒predicate-Price”) In original domain DB consisting of three table components Based on this, the word extraction model is learned. The table component is an element that constitutes a triple database of subject / predicate / object. The object here refers to the value of the entry where subject and predicate intersect in the table. In this embodiment, a case of a triplet will be described as an example, but the present invention can be applied to a triplet or more.

  Then, the learned model is applied to the home appliance domain, which is an unlearned target domain. When applying the input sentence “How much is a vacuum cleaner” and the input of the target domain DB, for example, ““ Vacuum cleaner ”⇒subject-vacuum”, ““ How much ”⇒predicate-price” Get correspondences between sentence words and table elements and entries.

  In the embodiment of the present invention, a solution divided into two stages is adopted for the above problem. In the first stage, an extraction model with a high level of abstraction and the same semantic structure (table constituent element) level of the database is used.

  In the second stage, the table components are subdivided and linked to the entries, so the classification approach is used. Here, if the learning database of the learned original domain and the database of the target domain are different, it is impossible to take a direct classification-based approach to the learning database. Take an approach to find the similarity to. In this way, it is possible to extract words associated with entries even in unknown domains.

(Relativity of features)

  The problem that it is difficult to determine which features should be weighted is handled by relativizing the features. A relative score is assigned to an entry that is another classification candidate. Thereby, it can be considered to give the same amount of information as the classification problem.

(Semantic vector model across domains)

  In response to the problem that the weight at the time of learning and the weight at the time of application may not be equal to each other, the unsupervised text in the original domain and the unsupervised text in the target domain are learned simultaneously to obtain a semantic vector model. Thereby, the weight with respect to the said vector obtained at the time of learning can be used as it is also at the time of application.

(Interdomain similarity feature)

  For issues where the strings may not be similar, measure the similarity between the target domain text string and the source domain text string (eg <when is the XYZ party>), Based on the correspondence with the entry of the original domain (for example, corresponding to the entry “date” in the original domain DB), the similarity between the entry of the original domain DB and the entry of the target domain DB (“date” and “date”) The similarity is calculated. Thereby, even in-sentence notations that are not superficially and semantically similar can be linked to entries in the DB.

(General model features)

  In order to obtain general-purpose supplemental information, a general-purpose model learned from general-purpose teacher data that is roughly used in any domain is used as a clue for estimation in the target domain. Thereby, even if the original domain and the target domain are greatly different, they can be linked to the entry in the DB.

<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. 2, the model learning device 100 according to the 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 and a calculation unit 20 as shown in FIG.

  The input unit 10 receives the annotated source domain question sentence set 21, the source domain DB 22, and the source domain unsupervised text 23 for the source domain to be learned. Annotated source domain question text is, for example, a morpheme for a question text like "<subj = picture book> book </ subj> where <pred = reseller> where / where can I buy </ pred>?" This is a question sentence to which a slash indicating a boundary and an association between a table component and an entry are given.

  Further, the input unit 10 receives the target domain unsupervised text 24 and the general model supervised text for the target domain at the time of application.

  The arithmetic unit 20 includes an annotated original domain question sentence set 21, an original domain DB 22, an original domain unsupervised text 23, a target domain unsupervised text 24, a general model supervised text 25, a semantic vector learning unit 30, The semantic vector model 32, the general-purpose model learning unit 34, the general-purpose model 36, the series model feature making unit 38, the series model learning unit 40, the series model 42, the regression model feature making unit 44, and the regression A model learning unit 46 and a regression model 48 are included.

  The semantic vector learning unit 30 learns a semantic vector model 32 composed of semantic vectors for each word based on the original domain unsupervised text 23 and the target domain unsupervised text 24 received by the input unit 10. The semantic vector model 32 is a model that can output, for example, that the semantic similarity between “book” and “picture book” is 0.5. The semantic vector model 32 learns by simultaneously vectorizing words included in the original domain unsupervised text 23 and the target domain unsupervised text 24. The existing modeling method described in Non-Patent Document 2 is used for modeling. As a modification, learning may be performed by further adding general-purpose text that does not depend on the domain to the input.

  The general-purpose model learning unit 34 learns the general-purpose model 36 based on the general-purpose model supervised text 25 received by the input unit 10. The general-purpose model 36 is a model that can identify that “NTT” is a company name, for example. Here, a general-purpose model that does not depend on the domain is learned using supervised text. There are two general-purpose models to be learned in this embodiment: a sentence classification model and a word string extraction model. In this embodiment, a sentence classification model is learned. The sentence classification model is a model that uses an existing classification technique (such as Support Vector Machines) from data that provides an identification result for one sentence and the data. For example, the sentence classification model gives an estimated label that “price” is heard from a sentence “this is how much”. Note that when learning a word string extraction model, data that gives an identification result for each word in a sentence and a model that uses an existing sequence labeling method (CRF or the like) are learned from the data. For example, “<company name> NTT </ company name> <amount> line usage fee </ amount> is <amount question> how much </ amount question>?” It is done.

  The sequence model feature conversion unit 38 uses the semantic vector model 32 learned by the semantic vector learning unit 30 to configure the table configuration of the original domain DB 22 for each word included in the question sentence in the annotated original domain question sentence set 21. For each element, the similarity with the word representing the entry of the table constituent element is extracted as a feature. In addition, it is a premise that the annotation data of the question sentence is mapped to the entry. The sequence model feature generating unit 38 uses only the mapping information to the table components in the annotation data.

  For example, using the output label of the sequence model as a table component of the original domain DB 22, the sequence model feature generating unit 38 performs the following first to third processes for each word included in the question sentence for each table component. Extract features.

  In the first process, the series model feature obtaining unit 38 first extracts the surface layer similarity and the semantic similarity with each entry, and for each table component of the original domain DB 22, for each entry of the table component The maximum value among the surface layer similarity and the semantic similarity is taken as the feature of the table component. For example, if there are three types of table elements, subject, predicate, and object, the surface layer similarity and the semantic similarity are extracted for each of the three types.

  The surface layer similarity is the edit distance between the target word in the question sentence and each entry, and the maximum value of the surface layer similarity with each entry of the table component is extracted as the surface layer similarity of the table component Is done. For example, the similarity is calculated based on the target word and entry results such as “main subject: picture book editing distance = 1, character duplication rate = 0.5, word matching rate = 0”, and the surface layer similarity is obtained. The semantic similarity is a similarity calculated using a semantic vector model for the target word and entry pair, and the maximum value of the semantic similarity with each entry of the table component is the table. Extracted as the semantic similarity of components. For example, the similarity is calculated based on the result of the target word and entry such as “main subject: picture book semantic similarity = 0.5”, and the maximum value is determined from the surface layer of the target word and the table component. Extract as similarity. An example of the output of the feature data for the series model of the table constituent elements is as follows. If the target word of interest is `` book '', `` correct answer label = B-subj subj edit distance = 1 subj semantic similarity = 0.5 pred edit distance = 2 pred semantic similarity = 0.1 obj edit distance = 2 obj semantic similarity = 0.1 ". If the target word of interest is “ha”, then “correct answer label = O subj edit distance = 2 subj semantic similarity = 0.1 pred edit distance = 2 pred semantic similarity = 0.1 obj edit distance = 2 obj semantic similarity = 0.1 " If the target word of interest is `` where '', `` correct answer label = B-pred subj edit distance = 6 subj semantic similarity = 0.1 pred edit distance = 6 pred semantic similarity = 0.1 obj edit distance = 6 obj semantic similarity = 0.1 ". Here, B / I / O given to the header of the correct label represents O = a word that is not extracted except for B = the head of the target word string to be extracted, I = the head of the target word string to be extracted.

  In the second process, the series model feature obtaining unit 38 estimates the question contents of the question sentence using the general model 36 that does not depend on the domain for estimating the question contents of the question sentence for each question sentence. Then, the estimation result of the question sentence is extracted as a feature of a specific word included in the question sentence. For example, as the estimation result of the sentence classification model that is the general-purpose model 36, an estimation result that the sentence is listening to “amount” is obtained. This estimation result indicating "what is heard as a whole sentence" is often an important clue to the content word at the end and the first question word. As given. When the general model 36 is a word string extraction model, the estimation result of the word estimated using the domain-independent general model 36 for estimating the superordinate concept of the word is extracted as the feature of the word. You can do it. For example, the estimation result is given as a feature to each word in the estimated word string range (eg, “NTT⇒company name”, “line usage fee⇒amount”).

  In the third process, the series model feature generating unit 38 extracts an inter-domain similarity feature for each word included in the question sentence. At the time of learning, the similarity between a target word and a word representing an entry given to the same notation as the target word by annotate data is set as an inter-domain similarity feature. In the learning data, when there are a plurality of types of entries given to the same notation as the target word, the target word and the corresponding one of the pairs of entries given to the same notation as the target word A ratio of occurrence of a pair with an entry assigned to the target word is defined as an inter-domain similarity feature. For example, as shown in FIG. 3, the inter-domain similarity feature (similarity b) for the target word “how much” and the annotated entry “price” is included in the question sentence of the annotated source domain question sentence set 21. Of the pairs of all appearing words “how much” and annotated entries, the ratio of the pair of the word “how much” and the entry “price” may be used as the similarity feature between domains.

  The sequence model learning unit 40 extracts the features for each table component extracted for each word included in the question sentence in the annotated source domain question sentence set 21 by the series model feature converting part 38 and the annotated source domain question sentence. Based on the labels given to the question sentences in the set 21, a sequence model 42 for extracting table constituent elements is learned using CRF (Non-Patent Document 3) or the like of the existing method. The series model 42 is a weight parameter for each feature corresponding to each label (table constituent element). Depending on the sequence model to be learned, it is a word string that concatenates one word or two or more words, such as "<subj> book </ subj> is <pred> where / where / buy </ pred>". A label can be assigned to a certain word chunk.

  The regression model feature conversion unit 44 uses the meaning vector of each word learned by the semantic vector learning unit 30 to include a labeled word chunk included in the question sentence in the annotated source domain question sentence set 21. For each, the similarity between the word chunk and the word representing each entry candidate of the table component represented by the label is extracted as a feature.

  In the regression model feature conversion unit 44, differently from the series model feature conversion unit 38, specifically, the first to fourth processes are performed below for each word chunk to which a label included in the question sentence is given. A feature is performed for each entry candidate of the table constituent element represented by the label given to the word chunk and the table constituent element of the original domain DB 22.

  First, in the first process, the regression model feature converting unit 44 selects, for each word chunk included in the question sentence, the entry candidate of the original domain DB 22 of the table component represented by the label attached to the word chunk. The surface layer similarity and the semantic similarity with each of them are extracted as features in the same manner as in the series model feature converting unit 38.

  In the second process, the regression model feature generating unit 44 estimates the question content of the question sentence using the general-purpose model 36 for each question sentence, as in the case of the series model feature providing unit 38, Is extracted as a feature of a specific word chunk included in the question sentence. In addition, since the second processing of the regression model feature generating unit 44 is performed for each entry, first, the mapping of which entry in the original domain DB 22 is closest to the label classified by the general-purpose model 36 is performed. deep. For example, if there is “Price” in the entry of the original domain DB 22 and “Price” exists in the classification target of the general-purpose model 36, they are mapped as being close. In the regression model, all entry candidates are featured, so for each entry candidate, a pair between the mapped generic model label and the actually estimated generic model estimation label is extracted as a feature. To do. For example, when learning is performed on the entry candidate “price” of the original domain DB 22 during learning and the estimated label (sentence classifier) of the general-purpose model 36 outputs “money”, “{DB candidate map = A feature of “price & general model estimation result = amount} = 1” is extracted.

  In the third process, the regression model feature converting unit 44 extracts the inter-domain similarity feature for each of the word chunks included in the question sentence in the same manner as the sequence model feature generating unit 38.

  In the fourth process, the regression model feature-imparting unit 44, for each of the word chunks to which the label is assigned, has a surface layer with each of the words representing the entry candidate of the table component represented by the label attached to the word chunk. The features are relativized with respect to similarity, semantic similarity, and interdomain similarity. In the fourth process, for each of the word chunks, for each of the entry candidates, for each of the three types of surface layer similarity, semantic similarity extracted in the first process, and interdomain similarity extracted in the third process, The extracted value having the maximum value is set as the relativization feature, and the relativistic surface layer similarity, the relativistic semantic similarity, and the relativistic domain similarity are added to the features of all entry candidates. As shown in FIG. 4, for example, in the table element predicate represented by the label given to the word chunk, there are entry candidates “price”, “weight”, “sales date”, and the maximum value of semantic similarity is “weight”. "0.2" is added to all entry candidates "price", "weight", and "sales date" as a relative semantic similarity.

  The regression model learning unit 46 extracts the features for each entry candidate extracted for each word chunk included in the question sentence in the annotated source domain question sentence set 21 by the regression model featureizing part 44 and the annotated source domain question sentence. Based on the labels given to each question sentence in the set 21, a regression model 48 for extracting a word representing an entry is learned using a logistic regression model or the like of an existing method. Specifically, the value 1 is assigned to a correct annotate pair (for example, “1: book-picture book”) of a word chunk (surface character string) with a label in a question sentence and a word representing an entry. Regression learning is performed by assigning 0 to other than (for example, “0: book-stuffed toy”).

<Configuration of Word Extraction Device According to Embodiment of the Present Invention>

  Next, the configuration of the word extraction device according to the embodiment of the present invention will be described. As shown in FIG. 5, a word extraction device 200 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 word extraction processing routine described later. Can be configured with a computer. Functionally, the word extraction device 100 includes an input unit 210, a calculation unit 220, and an output unit 250 as shown in FIG.

  The input unit 210 receives a target domain question sentence that is a question sentence about an unlearned target domain and a target domain DB 225 of the target domain. In the following explanation, the target domain question sentence is described as a question sentence.

  Further, the input unit 210 receives the annotated source domain question sentence set 21 and the source domain DB 22 for the learned source domain.

  The calculation unit 220 includes an annotated original domain question sentence set 21, an original domain DB 22, a semantic vector model 32, a general model 36, a series model 42, a regression model 48, a series model featureizing unit 238, A series model application unit 240, a regression model feature conversion unit 244, and a regression model application unit 246 are included.

  The annotated original domain question sentence set 21, the original domain DB 22, the semantic vector model 32, the general model 36, the series model 42, and the regression model 48 are stored in the same manner as the model learning apparatus 100. Yes.

  The series model feature generating unit 238 uses, for each word included in the question sentence, a word representing an entry of the table constituent element for each table constituent element of the target domain DB 225 using the semantic vector model 32 created in advance. Are extracted as features. Specifically, the features are extracted by performing the first to third processes in the same manner as the above and the sequence model feature converting unit 38 of the model learning apparatus 100, but the points described below are different.

  In the first process, the series model feature obtaining unit 238 extracts the surface layer similarity and the semantic similarity with each entry of the target domain DB 225, and for each table component of the target domain DB 225, the table component The maximum value of the surface layer similarity and the semantic similarity with each entry is the feature of the table component. In addition, in the third process, the series model feature generating unit 238 extracts features obtained by combining the following similarities a to c as interdomain similarity features. The similarity a is a similarity to a word included in the question sentence of the annotated source domain question sentence set 21 that is most similar to the target word included in the question sentence in the target domain. The similarity b is a similarity with a word representing an entry in the original domain DB 22 assigned to a word included in a question sentence of a learned domain that is most similar to a word in the target domain. The similarity c is a word representing an entry in the original domain DB 22 given to a word included in a question sentence of a learned domain that is most similar to a word in the target domain, and each entry of the table component of the target domain DB 225 Among the similarities to the word representing, it is the maximum similarity. Moreover, what is necessary is just to refer to the annotated former domain question sentence set 21 and former domain DB22 about the word showing the entry provided to the word contained in the question sentence of the learned domain. FIG. 6 shows an example of the similarities a to c.

  The sequence model application unit 240 extracts a table component corresponding to a word learned in advance from a set of questioned sentences in the learned domain, to which a word representing an entry in the learned domain database is assigned. 42 and the feature for each word included in the question sentence extracted by the series model feature generating unit 238, the word chunk included in the question sentence is labeled with a table element. Give. Application of the sequence model 42 may be performed using CRF (Non-patent Document 3) or the like of an existing method. For example, a label such as “<subj> vacuum cleaner </ subj> is <pred> where / where / buy </ pred>” is given.

  The regression model feature conversion unit 244 uses the semantic vector model 32 to determine entry candidates of the table constituent elements represented by the labels assigned to the word chunks for each of the word chunks assigned the labels by the sequence model application unit 240. Similarity with each word to be expressed is extracted as a feature. Specifically, the first to fourth processes are performed in the same manner as the regression model feature converting unit 44 of the model learning apparatus 100 to extract features. Further, in the third process, similar to the sequence model feature generating unit 238, for each word chunk, a feature obtained by combining the similarities a to c is extracted as the inter-domain similarity feature.

  The regression model applying unit 246 extracts a regression model for extracting a word representing an entry corresponding to a word learned in advance from a set of question sentences in the learned domain, to which a word representing an entry in the learned domain database is assigned. 48 for each table component, based on 48 and the features for each entry candidate of the table component represented by the label, extracted for each of the word chunks to which the label is assigned by the regression model featureizing unit 244. A word representing an entry in the target domain DB 225 corresponding to the sentence is extracted and output to the output unit 250. In the word extraction, a regression model is applied to the feature of each entry candidate of the table element represented by the label extracted for each word chunk to calculate a value for the pair of the word chunk and the entry candidate. For each table component, the entry candidate that obtained the highest output value is output as the final result. For example, for the word chunk “vacuum cleaner” with the subject label, the output value of the pair with the entry candidate is “1.0 vacuum cleaner-“ subj-vacuum cleaner ””, “0.2 vacuum cleaner-“ subj-washer ” If "", the former is output.

<Operation of Model Learning Device According to Embodiment of Present Invention>

  Next, the operation of the word extraction device 100 according to the embodiment of the present invention will be described. The input unit 10 accepts the annotated source domain question sentence set 21, the source domain DB 22, and the source domain unsupervised text 23 for the source domain to be learned, and the target domain unsupervised text 24 for the target domain at the time of application. When the general model supervised text is received, the word extraction device 100 executes a model learning processing routine shown in FIG.

  In step S <b> 100, the semantic vector model 32 is learned based on the original domain unsupervised text 23 and the target domain unsupervised text 24 received by the input unit 10.

  In step S102, the general-purpose model 36 is learned based on the general-purpose model supervised text 25 received by the input unit 10.

  In step S104, for each word included in the question sentence in the annotated source domain question sentence set 21, using the semantic vector model 32 learned in step S100, for each table constituent element of the original domain DB 22, the table constituent element The similarity with the word representing the entry is extracted as a feature.

  In step S106, the feature for each table component extracted for each word included in the question sentence in the annotated source domain question sentence set 21 in step S104 and the question sentence in the annotated source domain question sentence set 21 are given. The sequence model 42 for extracting the table constituent elements is learned based on the labeled labels.

  In step S108, for each word chunk with a label included in each question sentence in the annotated source domain question sentence set 21, using the semantic vector of each word learned by the semantic vector learning unit 30, The similarity between the word chunk and the word representing each entry candidate of the table component represented by the label is extracted as a feature.

  In step S110, the features for each entry candidate extracted for each word chunk included in the question sentence in the annotated source domain question sentence set 21 by the regression model featureizing unit 44 and the annotated source domain question sentence set 21 Based on the label given to each question sentence, a regression model 48 for extracting a word representing an entry is learned using a logistic regression model or the like of an existing method.

  Specifically, step S104 is executed by a sequence model feature processing routine shown in FIG.

  In step S200, for each word included in each question sentence of the annotated original domain question sentence set 21, the surface layer similarity and the semantic similarity with each entry of the table constituent element of the original domain DB 22 are obtained for each table constituent element. The maximum value among the surface layer similarity and the semantic similarity with each entry of the table constituent element is extracted as the feature of the table constituent element.

  In step S202, for each question sentence in the annotated source domain question sentence set 21, the question content of the question sentence is estimated using the general model 36, and the estimation result of the question sentence is obtained as a specific word included in the question sentence. Extracted as a feature of.

  In step S204, the similarity between each word included in each question sentence of the annotated source domain question sentence set 21 and the word representing the entry given to the same notation by the annotate data is calculated between the domains related to the domains. Extracted as similarity feature.

  Specifically, step S108 is executed by a regression model feature processing routine shown in FIG.

  In step S300, for each of the word chunks with a label included in each question sentence of the annotated source domain question sentence set 21, the table constituent elements represented by the label attached to the word chunk are stored in the original domain DB 22. The surface layer similarity and the semantic similarity with each of the entry candidates are extracted as features in the same manner as the sequence model feature generating unit 38.

  In step S302, for each question sentence in the annotated source domain question sentence set 21, the question contents of the question sentence are estimated using the general-purpose model 36, and the estimation result of the question contents is determined as a specific item included in the question sentence. Extracted as a word chunk feature.

  In step S304, the domain similarity feature between domains is extracted for each of the labeled word chunks included in each question sentence of the annotated source domain question sentence set 21.

  In step S306, for each of the word chunks to which a label is attached, which is included in each question sentence of the annotated source domain question sentence set 21, a word representing an entry candidate of the table component represented by the label attached to the word chunk The features are relativized with respect to the surface layer similarity, the semantic similarity, and the inter-domain similarity with each.

<Operation of the word extraction device according to the embodiment of the present invention>

  Next, the operation of the word extraction device 200 according to the embodiment of the present invention will be described. The input unit 10 accepts a target domain question sentence that is a question sentence about an unlearned target domain and a target domain DB 225 of the target domain, and an annotated original domain question sentence set 21 about the learned original domain, Upon receipt of the domain DB 22, the word extraction device 200 executes a word extraction processing routine shown in FIG.

  First, in step S400, using the semantic vector model 32, for each word included in the question sentence, for each table component of the target domain DB 225, the similarity to the word representing the entry of the table component is used as a feature. Extract.

  Next, in step S402, based on the sequence model 42 and the feature for each table component extracted for each word included in the question sentence in step S400, the word constituent included in the question sentence is converted into a table constituent element. A label representing is attached.

  In step S404, using the semantic vector model 32, the similarity between each word chunk assigned a label in step S402 and each word representing an entry candidate of a table element represented by the label assigned to the word chunk. Are extracted as features.

  In step S406, based on the regression model 48 and the features for each entry candidate of the table element represented by the label extracted for each of the word chunks to which the label is assigned by the regression model featureizing unit 244, the table For each component, a word representing an entry in the target domain DB 225 corresponding to the question sentence is extracted and output to the output unit 250, and the process ends.

  Note that step S400 executes the same processing as the sequence modeling feature processing routine shown in FIG. In step S404, the same processing as the regression modeling feature processing routine shown in FIG. 9 is executed.

  As described above, according to the word extraction device according to the embodiment of the present invention, an entry is made for each table component of the target domain DB 225 for each word included in the question sentence using the semantic vector of each word. Based on the series model 42 for extracting a table component corresponding to the word and the extracted feature of each word included in the question sentence. Each word chunk included in the sentence is assigned a label representing a table component, and the meaning vector of each word is used to represent the entry of the table component represented by the label attached to the word chunk for each word chunk. The similarity with each word is extracted as a feature, a regression model 48 for extracting a word representing an entry corresponding to the word, and the extracted label is Corresponding to a question sentence when using an unlearned domain database by extracting a word representing an unlearned domain database entry corresponding to the question sentence based on the features of each given word chunk Can be extracted.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

10, 210 Input unit 20, 220 Arithmetic unit 21 Annotated source domain question sentence set 23 Original domain unsupervised text 24 Target domain unsupervised text 25 General model supervised text 30 Semantic vector learning unit 32 Semantic vector model 34 General model learning unit 36 General-purpose model 38, 238 Sequence model feature unit 40 Sequence model learning unit 42 Sequence model 44, 244 Regression model feature unit 46 Regression model learning unit 48 Regression model 100 Model learning device 200 Word extraction device 240 Sequence model application unit 246 regression model application unit 250 output unit

Claims (7)

A word extraction device for extracting a word representing an entry corresponding to a question sentence of an input target domain from a word representing an entry in an unlearned target domain database,
For each word included in the question sentence for each table component of the database of the target domain, the similarity with the word representing the entry of the table component is obtained using a semantic vector of each word created in advance. A sequence model feature extraction unit to extract as features,
A sequence model for extracting a table component corresponding to a word chunk, learned in advance from a set of question sentences in the original domain, to which a word representing an entry in a learned original domain database is assigned, and the sequence Based on the features for each table component extracted for each word included in the question sentence by the model feature conversion unit, each word chunk included in the question sentence has a label representing the table element. A series model application section to be assigned;
For each of the word chunks to which the label is given by the sequence model application unit using the semantic vector of each word, each of the words representing the entry of the table component represented by the label attached to the word chunk; A regression model featureizing unit that extracts the similarity of
A regression model for extracting a word representing an entry corresponding to a word learned in advance from a set of question sentences in the original domain, to which a word representing an entry in the original domain database is assigned, and the features for the regression model A regression model application unit that extracts words representing database entries of the target domain corresponding to the question sentence, based on the features of each of the word chunks to which the label is attached, extracted by the conversion unit;
Word extraction device including The regression model feature generating unit includes, for each word chunk assigned the label by the sequence model application unit, each word representing an entry in the table component represented by the label attached to the word chunk; Each of at least one feature representing the similarity of
The word extraction device according to claim 1, wherein each of the at least one feature is subjected to relativity of the feature representing a similarity to each of the words representing the entry of the table component represented by the label.   The word extraction device according to claim 1 or 2, wherein the semantic vector of each word is created in advance based on the text of the target domain and the text of the original domain. The regression model feature-imparting unit, for each of the word chunks that have been given the label by the series model application unit,
Similarity to the word contained in the question of the original domain that is most similar to the word chunk,
Included in the original domain question sentence that is most similar to the word chunk and similar to the word representing the entry given to the word contained in the original domain question sentence and the word chunk most similar to the word chunk The inter-domain similarity feature is further extracted by combining the similarity between the word representing the entry given to the word and the word representing the entry of the table component represented by the label given to the word. The word extraction device according to any one of claims 3 to 4. The sequence model feature generating unit uses a general-purpose model that does not depend on a domain for estimating a question content of a question sentence or a high-level concept of a word that has been learned in advance for each word included in the question sentence. The estimated result of the question sentence or the estimated result of the word is further extracted as a feature,
The regression model feature generating unit is configured to estimate the question sentence, or the word estimation result, estimated using the general-purpose model for each of the word chunks to which the label is assigned by the series model application unit. The word extraction device according to claim 1, further extracting as a feature. A word extraction method in a word extraction device for extracting a word representing an entry corresponding to a question sentence of an input target domain from a word representing an entry in an unlearned target domain database,
The sequence model feature recognition unit uses the semantic vector of each word created in advance to store the entry of the table component for each table component of the target domain database for each word included in the question sentence. Extracting the similarity to the word to be represented as a feature;
A sequence model application unit extracts a table component corresponding to a word chunk learned in advance from a set of question sentences in the original domain, to which a word representing an entry in a learned original domain database is assigned. Based on the sequence model and the features for each table component extracted for each word included in the question sentence by the sequence model feature conversion unit, the word chunk included in the question sentence includes the table. Providing a label representing the component;
The table constituent element represented by the label attached to the word chunk for each word chunk assigned the label by the sequence model applying unit using the semantic vector of each word Extracting the similarity with each of the words representing the entries as features,
A regression model for extracting a word representing an entry corresponding to a word learned in advance from a set of question sentences in the original domain, to which a regression model application unit is given a word representing an entry in the database of the original domain; The word representing the database domain entry of the target domain corresponding to the question sentence is extracted based on the features of each of the word chunks with the labels extracted by the regression model featureizing unit. Steps,
Word extraction method including   The program for functioning a computer as each part of the word extraction apparatus of any one of Claims 1-5.