JP2017204219A - Model learning apparatus, word extraction apparatus, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To learn a model for accurately extracting a word corresponding to a question sentence if there are other domains of a plurality of learning objects.SOLUTION: A series model featuring part 38 extracts, with regard to each of a plurality of original domains, a feature of an original domain on each word included in a question sentence by a table constituent element, and duplicates as a feature of a common domain. A series model learning part 40 learns a series model 42 according to the feature of the original domain and the feature of the common domain obtained with regard to each of the plurality of original domains. A regression model featuring part 44 extracts a regression model feature. A regression model learning part 46 learns a regression model 48.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a model learning device, a word extraction device, a method, and a program, and more particularly, to a model learning device, a word extraction device, a method, and a program for extracting a word necessary for a question answer 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 pair, the method of measuring a similarity after vectorizing a word in a semantic space is known (refer nonpatent literature 3).

  In addition, even when there is no learning data in the target domain, there is known a method for realizing highly accurate word string extraction at low cost by diverting learning data when there is learning data in another domain ( Non-patent document 5).

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). Hal Daume III and Daniel Marcu. 2006. Domain adaptation for statistical classifiers. Journal of Artificial Intelligence Research, pages 101-126 Yazdani, Majid and Henderson, James, “A Model of Zero-Shot Learning of Spoken Language Understanding”, EMNLP2015

  However, since the method described in Non-Patent Document 5 is a method that assumes that other domains are independent, the accuracy is low.

  The present invention has been made in view of the above circumstances, and model learning that can learn a model that can accurately extract a word corresponding to a question sentence when there are a plurality of learned other domains. An object is to provide an apparatus, a method, and a program.

  It is another object of the present invention to provide a word extraction apparatus, method, and program capable of accurately extracting a word corresponding to a question sentence.

  To achieve the above object, the model learning device according to the present invention extracts a word representing an entry corresponding to a question sentence of an input target domain from a word representing a database entry of an unlearned target domain. A model learning device for providing a set of question sentences in the original domain, to each of a plurality of original domains, a word indicating a database entry of the original domain and a label indicating a table component are given Based on the meaning vector of each word created in advance, the similarity between the word included in the question sentence and the word representing the entry of the table component is calculated for each word included in the question sentence. A sequence model feature generating unit that extracts the domain feature as a domain feature and replicates the original domain feature as a common domain feature; Based on the features of the original domain and the features of the common domain obtained for each word included in the question sentence for each of a plurality of original domains, and the labels given for each word included in the question sentence A sequence model learning unit for learning a sequence model for extracting a table constituent element corresponding to a word chunk, and for each of the plurality of original domains, each word based on a set of question sentences of the original domain For each of the word chunks to which the label is assigned, the similarity with each of the words representing the entry of the table component represented by the label assigned to the word chunk is used for the regression model. A regression model feature extracting unit that extracts the feature, and a word channel to which the label is assigned to each of the plurality of original domains. A regression model for learning a regression model for extracting a word representing an entry corresponding to a word based on the features for the regression model obtained for each of the keywords and the label given for each of the word chunks And a learning unit.

  The model learning method according to the present invention is a model learning 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. A query method of the original domain, in which the feature modeling unit for a sequence model is assigned a word indicating a database entry of the original domain and a label indicating a table component for each of the plurality of original domains Based on a set of sentences, using a semantic vector of each word created in advance, each word included in the question sentence is similar to a word representing an entry of the table component for each table component Are extracted as features of the original domain, and the features of the original domain are duplicated as features of the common domain. A learning unit is assigned to each of the plurality of original domains, the original domain feature and the common domain feature obtained for each word included in the question sentence, and each word included in the question sentence. Based on the label, a sequence model for extracting a table constituent element corresponding to the word chunk is learned, and the regression model featureizing unit asks each of the plurality of original domains a question of the original domain. Each word representing an entry in the table element represented by the label attached to the word chunk, for each word chunk assigned the label, using a semantic vector of each word based on a set of sentences And the regression model learning unit assigns the label to each of the plurality of original domains. Learning a regression model for extracting a word representing an entry corresponding to a word based on the features for the regression model obtained for each of the word chunks and the label given for each of the word chunks To do.

  A word extraction device according to the present invention is a word extraction device that extracts 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. The applied time series model feature extracting unit that extracts the features of the common domain, the sequence model learned by the model learning device, and each word included in the question sentence by the applied time series model feature generating unit Based on the extracted features of the common domain for each table component, each word chunk included in the question sentence is A sequence model application unit that assigns a label representing the table component, and a word that is assigned to the word chunk by the sequence model application unit using the semantic vector of each word. A regression model feature conversion unit for extracting a similarity with each of the words representing the entry of the table component represented by the label as a feature for regression model; and the regression model learned by the model learning device; , Based on the regression model features of each of the word chunks with the label extracted by the application regression model featureization unit, the database entry of the target domain corresponding to the question sentence is And a regression model application unit that extracts a word to be expressed.

  The word extraction method according to the present invention is a word extraction method in a word extraction device for extracting a word representing an entry corresponding to an inputted question sentence of a target domain from a word representing an entry in an unlearned target domain database. The applied time-series model feature generating unit uses, for each word included in the question sentence for each table constituent element of the database of the target domain, using the semantic vector of each word created in advance. Similarity with a word representing a table component entry is extracted as a feature of a common domain, and the sequence model application unit learns the sequence model learned by the above model learning method, and the applied time series model features Based on the features of the common domain for each table component extracted for each word included in the question sentence by the section A label representing the table component is assigned to each word chunk included in the question sentence, and the applied regression model featureizing unit uses the semantic vector of each word to generate the label by the series model applying unit. For each of the word chunks to which the word chunk is assigned is extracted as a feature for regression model by extracting the similarity with each of the words representing the entry of the table component represented by the label assigned to the word chunk, and a regression model application unit Is based on the regression model learned by the model learning method, and the features for the regression model of each of the word chunks to which the label has been extracted, extracted by the feature model for regression model at the time of application. A word representing an entry in the target domain database corresponding to the question sentence is extracted.

  The program according to the present invention is a program for causing a computer to function as each part of the model learning device or the word extracting device.

  According to the model learning device, method, and program of the present invention, for each of a plurality of original domains, the similarity to a word representing an entry of a table component is extracted as a feature of the original domain, A sequence for replicating a feature as a common domain feature and extracting table constituent elements corresponding to word chunks based on the original domain features and common domain features obtained for each of a plurality of original domains By learning the model, there is an effect that it is possible to learn a model that can accurately extract a word corresponding to a question sentence when there are a plurality of other domains to be learned.

  According to the word extraction device, method, and program of the present invention, it is possible to extract the word corresponding to the question sentence with high accuracy.

It is the abstract figure showing the relationship of the table structural element in the database of the some 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 block diagram which shows the structure of the word extraction apparatus which concerns on embodiment of this invention. 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 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, at the time of learning, a plurality of original domains to be learned are a toy domain and a travel domain. The learning input sentence “How much is this book?”, The learning input sentence annotation result (eg, “book ⇒ subject-picture book”, “how much ⇒ predicate-price”), original domain 1DB consisting of three table components, A word extraction model is learned based on the original domain 2DB. 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.

  Further, in the conventional technique, when there are a plurality of learned other domains, it cannot be fully utilized. Therefore, in the embodiment of the present invention, the technology of Non-Patent Document 4 which is the prior art of transfer learning is used. And focus on common information between domains during learning. Thereby, when there are a plurality of other domains, the accuracy in the target domain can be improved.

<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 an annotated source domain 1 question sentence set 21A, an annotated source domain 2 question sentence set 21B, an original domain 1DB 22A, and an original domain 2DB 22B for a plurality of original domains 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. For example, the original domain 1 is a toy domain, and the original domain 2 is a travel domain.

  Further, the input unit 10 receives a semantic vector model 32.

  The computing unit 20 includes an annotated original domain 1 question sentence set 21A, an original domain 1DB 22A, an annotated original domain 2 question sentence set 21B, an original domain 2DB 22B, a semantic vector model 32, and a sequence model featureizing part 38. A series model learning unit 40, a series model 42, a regression model feature conversion unit 44, a regression model learning unit 46, and a regression model 48.

  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 may be learned in advance by making a vector model from words included in the unsupervised text. The existing modeling method described in Non-Patent Document 2 is used for modeling.

  Using the semantic vector model 32, the sequence model feature generating unit 38 uses the table component for each table component of the original domain 1DB 22A for each word included in the question sentence in the annotated original domain 1 question sentence set 21A. The degree of similarity with the word representing the entry is extracted as a feature of the original domain 1 and is also copied as a feature of the common domain. 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. In addition, a header (for example, “toy-”) representing the original domain 1 is assigned to the feature of the original domain 1, and a header (for example, “common-”) representing the common domain is assigned to the feature of the common domain.

  For example, using the output label of the sequence model as a table component of the original domain 1 DB 22A, the sequence model feature generating unit 38 performs the following processing for each word included in the question sentence for each table component by the following processing. And extract common domain features.

  The series model feature generating unit 38 extracts the surface layer similarity and semantic similarity with each entry, and for each table component of the original domain 1DB 22A, the surface layer similarity and semantic similarity with each entry of the table component The maximum value in degrees is the feature of the original domain 1 of the table constituent element. 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 the similarity calculated using the semantic vector model 32 for the target word and entry, and the maximum value of the semantic similarity with each entry of the table component is Extracted as the semantic similarity of the table elements. 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. Extracted 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”, the following features of the original domain 1 and the features of the common domain are extracted.

Correct answer label = B-subj
toy-subj edit distance = 1 toy-subj semantic similarity = 0.5 toy-pred edit distance = 2 toy-pred semantic similarity = 0.1 toy-obj edit distance = 2 toy-obj semantic similarity = 0.1
common-subj edit distance = 1 common-subj semantic similarity = 0.5 common-pred edit distance = 2 common-pred semantic similarity = 0.1 common-obj edit distance = 2 common-obj semantic similarity = 0.1

  If the target word of interest is “ha”, the following features of the original domain 1 and features of the common domain are extracted.

Correct label = O
toy-subj edit distance = 2 toy-subj semantic similarity = 0.1 toy-pred edit distance = 2 toy-pred semantic similarity = 0.1 toy-obj edit distance = 2 toy-obj semantic similarity = 0.1
common-subj edit distance = 2 common-subj semantic similarity = 0.1 common-pred edit distance = 2 common-pred semantic similarity = 0.1 common-obj edit distance = 2 common-obj semantic similarity = 0.1

  If the target word of interest is “where”, the following features of the original domain 1 and features of the common domain are extracted.

Correct answer label = B-pred
toy-subj edit distance = 6 toy-subj semantic similarity = 0.1 toy-pred edit distance = 6 toy-pred semantic similarity = 0.1 toy-obj edit distance = 6 toy-obj semantic similarity = 0.1
common-subj edit distance = 6 common-subj semantic similarity = 0.1 common-pred edit distance = 6 common-pred semantic similarity = 0.1 common-obj edit distance = 6 common-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 toy domain, “toy-” and “common-” are added to the feature header as described above.

  In addition, the sequence model feature generating unit 38 uses the semantic vector model 32 to generate, for each table component of the original domain 2DB 22B, the original domain 2DB 22B for each word included in the question sentence in the annotated original domain 2 question sentence set 21B. Similar to the feature 1, the similarity to the word representing the entry of the table constituent element is extracted as the feature of the original domain 2 and duplicated as the feature of the common domain. In addition, a header (for example, “travel-”) representing the original domain 2 is assigned to the feature of the original domain 2, and a header (for example, “common-”) representing the common domain is assigned to the feature of the common domain.

  The sequence model learning unit 40 extracts the features of the original domain 1 for each table component and the common domain extracted for each word included in the question sentence in the annotated source domain 1 question sentence set 21A by the series model feature conversion unit 38 , The label assigned to the question sentence in the annotated source domain 1 question sentence set 21A, and each word included in the question sentence in the annotated source domain 2 question sentence set 21B by the sequence model feature conversion unit 38 CRF (non-patented) of the existing method based on the features of the original domain 2 and the common domain of each table component and the labels given to the question sentences in the annotated original domain 2 question sentence set 21B. The sequence model 42 for extracting the table constituent elements is learned using the literature 3). 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 making unit 44 uses the semantic vectors of each word included in the semantic vector model 32 to each of the labeled word chunks included in the question sentence in the annotated source domain 1 question sentence set 21A. 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 for the regression model.

  Unlike the series model feature unit 38, the regression model feature unit 44 performs the following processing, and for each word chunk to which a label included in the question sentence is assigned, It is a table component represented by the given label, and is featured for each entry candidate of the table component of the original domain 1DB 22A.

  Specifically, the regression model feature making unit 44, for each of the word chunks included in the question sentence, with each of the entry candidates in the original domain 1DB 22A of the table constituent element represented by the label attached to the word chunk. The surface layer similarity and the semantic similarity are extracted as regression model features in the same manner as the series model feature converting unit 38.

  In addition, the regression model feature making unit 44 uses the semantic vector of each word included in the semantic vector model 32 to add a labeled word chunk included in the question sentence in the annotated source domain 2 question sentence set 21B. , The similarity between the word chunk and the word representing each entry candidate of the table constituent element represented by the label is extracted as a regression model feature.

  The regression model learning unit 46 extracts the regression model features for each entry candidate extracted for each word chunk included in the question sentence in the annotated source domain 1 question sentence set 21A by the regression model feature making part 44 and the annotated Labels assigned to each question sentence in the original domain 1 question sentence set 21A and entries extracted for each word chunk included in the question sentence in the annotated original domain 2 question sentence set 21B by the regression model featureizing unit 44 Based on the features for the regression model for each candidate and the labels assigned to each question sentence in the annotated source domain 2 question sentence set 21B, a word representing an entry is extracted using the logistic regression model of the existing method. The regression model 48 for learning is learned. 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. 3, 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 200 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.

  The arithmetic unit 220 includes a semantic vector model 32, a series model 42, a regression model 48, an applied time series model feature making unit 238, a series model applying unit 240, an applied time regression model feature making unit 244, And a regression model application unit 246.

  The semantic vector model 32, the series model 42, and the regression model 48 store the same ones as the model learning apparatus 100.

  The applied time-series model feature generating unit 238 represents an entry of the table constituent element for each table constituent element of the target domain DB 225 for each word included in the question sentence, using the semantic vector model 32 created in advance. The similarity with the word is extracted as a feature of the common domain. Specifically, the same processing as the above and the sequence model feature converting unit 38 of the model learning device 100 is performed to extract the features, but the points described below are different.

  The applied time series model feature generating 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 maximum value among the surface layer similarity and the semantic similarity is the feature of the table constituent element. Further, the feature to be extracted is the common domain feature, and a header (for example, “common-”) representing the common domain is added.

  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 “vacuum cleaner”, the features of the following common domains are extracted.

common-subj edit distance = 1 common-subj semantic similarity = 0.5 common-pred edit distance = 2 common-pred semantic similarity = 0.1 common-obj edit distance = 2 common-obj semantic similarity = 0.1

  If the target word of interest is “ha”, the features of the following common domains are extracted.

common-subj edit distance = 2 common-subj semantic similarity = 0.1 common-pred edit distance = 2 common-pred semantic similarity = 0.1 common-obj edit distance = 2 common-obj semantic similarity = 0.1

  If the target word of interest is “where”, the features of the following common domains are extracted.

common-subj edit distance = 6 common-subj semantic similarity = 0.1 common-pred edit distance = 6 common-pred semantic similarity = 0.1 common-obj edit distance = 6 common-obj semantic similarity = 0.1

  The sequence model application unit 240 has a part corresponding to the common domain feature of the sequence model 42 for extracting table components corresponding to the words learned by the model learning device 100, and an applied time series model feature conversion unit. Based on the common domain feature for each table component extracted for each word included in the question sentence by 238, a label representing the table element is given to the word chunk included in the question sentence. Application of the sequence model 42 may be performed using CRF (Non-patent Document 3) or the like of an existing method. For example, applying the weight parameter for each of the common domain features corresponding to each table component to the extracted common domain features for each table component, the <subj> vacuum cleaner </ subj> < pred> where / where / buy </ pred> ”.

  The applied regression model feature generating unit 244 uses the semantic vector model 32 and, for each word chunk assigned a label by the sequence model applying unit 240, an entry of a table component represented by the label attached to the word chunk Similarity with each word representing a candidate is extracted as a feature for a regression model. Specifically, a feature is extracted by performing the same process as the regression model feature converting unit 44 of the model learning apparatus 100.

  The regression model application unit 246 has a regression model 48 for extracting a word representing an entry corresponding to the word learned by the model learning device 100, and a word given a label by the application regression model featureizing unit 244. A word representing an entry in the target domain DB 225 corresponding to the question sentence for each table component based on the regression model features for each entry candidate of the table component represented by the label extracted for each chunk. Extract and output to the output unit 250. The word extraction is performed by applying a regression model to the regression model feature for each entry candidate of the table element represented by the label extracted for each word chunk, and the value for the pair of the word chunk and the entry candidate. And the entry candidate that obtained the highest output value is output as the final result for each table component. 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 model learning device 100 according to the embodiment of the present invention will be described. The input unit 10 accepts the annotated source domain 1 question sentence set 21A and the original domain 1DB 22A for the original domain 1 to be learned, and the annotated source domain 2 question sentence set 21B for the original domain 2 to be learned; Upon receiving the original domain 2DB 22B, the model learning device 100 executes a model learning process routine shown in FIG.

  In step S100, for the original domain 1 to be learned, the annotated original domain 1 using the semantic vector model 32 based on the annotated original domain 1 question sentence set 21A and the original domain 1DB 22A received by the input unit 10 is used. For each word included in the question sentence in the one-question sentence set 21A, for each table constituent element of the original domain 1DB 22A, the similarity to the word representing the entry of the table constituent element is determined based on the features of the original domain 1 and the common domain. Extract as

  In step S102, for the original domain 2 to be learned, the annotated original domain 2 is received using the semantic vector model 32 based on the annotated original domain 2 question sentence set 21B and the original domain 2DB 22B received by the input unit 10. For each word included in the question sentence in the two-question sentence set 21B, for each table constituent element of the original domain 2DB 22B, the similarity to the word representing the entry of the table constituent element, the identity of the original domain 2 and the identity of the common domain Extract as

  In step S104, the features of the original domain 1 and the features of the common domain extracted for each word included in the question sentence in the annotated original domain 1 question sentence set 21A in step S100, and the annotated original domain The label attached to each question sentence in the one-question sentence set 21A and the original domain 2 for each table component extracted in step S102 for each word included in the question sentence in the annotated original domain-two question sentence set 21B Based on the features and the features of the common domain and the labels assigned to the question sentences in the annotated source domain 2 question sentence set 21B, the sequence model 42 for extracting the table constituent elements is learned.

  In step S106, a label included in each question sentence in the annotated source domain 1 question sentence set 21A using the semantic vector model 32 based on the annotated source domain 1 question sentence set 21A and the original domain 1 DB 22A. For each given word chunk, 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 for the regression model.

  Further, based on the annotated original domain 2 question sentence set 21B and the original domain 2 DB 22B, a label included in each question sentence in the annotated original domain 2 question sentence set 21B is assigned using the semantic vector model 32. For each word chunk, 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 regression model feature.

  In step S108, the features for the regression model for each entry candidate extracted for each word chunk included in the question sentence in the annotated source domain 1 question sentence set 21A in step S106, and the annotated source domain 1 question sentence set 21A. , The label assigned to each question sentence, the regression model features for each entry candidate extracted for each word chunk included in the question sentence in the annotated source domain 2 question sentence set 21B in step S106, and annotated Based on the labels given to each question sentence in the original domain 2 question sentence set 21B, 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, and the model The learning process routine is terminated.

<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. When 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, the word extraction device 200 executes a word extraction processing routine shown in FIG.

  First, in step S200, using the semantic vector model 32, for each word included in the question sentence, for each table element in the target domain DB 225, the similarity between the word representing the entry of the table element and the common domain is calculated. Extracted as a feature of.

  Next, in step S202, based on the part corresponding to the common domain feature of the series model 42 and the common domain feature for each table component extracted for each word included in the question sentence in step S200, A label representing a table component is given to the word chunk included in the question sentence.

  In step S204, using the semantic vector model 32, the similarity between each of the word chunks to which the label is assigned in step S202 and each of the words representing the entry candidate of the table element represented by the label attached to the word chunk. Are extracted as features for the regression model.

  In step S206, the regression model 48 and the feature for the regression model for each entry candidate of the table element represented by the label, extracted for each word chunk to which the label is assigned by the application regression model featureization unit 244, For each table 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 is terminated.

  As described above, according to the model learning device according to the embodiment of the present invention, the similarity between each of the plurality of original domains to be learned and the word representing the entry of the table component is determined based on the original domain. While extracting as a feature, the original domain feature is duplicated as a common domain feature and corresponds to a word chunk based on the original domain features and common domain features obtained for each of the multiple original domains By learning a sequence model for extracting table constituent elements, a model that can accurately extract a word corresponding to a question sentence can be learned when there are a plurality of other domains to be learned.

  Moreover, according to the word extracting device which concerns on embodiment of this invention, the word corresponding to a question sentence can be extracted with a sufficient precision.

  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.

  For example, the case where there are two original domains to be learned has been described as an example, but the present invention is not limited to this, and there may be three or more original domains to be learned.

10, 210 Input unit 20, 220 Operation unit 21A Annotated source domain 1 question sentence set 21B Annotated source domain 2 question sentence set 22A Original domain 1DB
22B Former domain 2DB
32 semantic vector model 38 sequence model feature unit 40 sequence model learning unit 42 sequence model 44 regression model feature unit 46 regression model learning unit 48 regression model 100 model learning device 200 word extraction device 238 applied time series model feature Unit 240 series model application unit 244 application regression model feature conversion unit 246 regression model application unit 250 output unit 225 target domain DB

Claims (7)

A model learning 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 of a plurality of original domains, a word indicating an entry in the original domain database and a label indicating a table component are given, and based on a set of question sentences of the original domain, For each word included in the question sentence using a semantic vector, for each table component, a similarity with a word representing an entry of the table component is extracted as a feature of the original domain, and the original A sequence model feature part that replicates domain features as common domain features;
For each of the plurality of original domains, the features of the original domain and the features of the common domain obtained for each word included in the question sentence, and the label given for each word included in the question sentence A sequence model learning unit for learning a sequence model for extracting a table constituent element corresponding to a word chunk,
For each of the plurality of original domains, based on a set of question sentences of the original domain, each word chunk to which the label is attached is assigned to the word chunk using a semantic vector of each word. A regression model feature converting unit that extracts a similarity with each of the words representing the entry of the table component represented by the label as a feature for the regression model;
Corresponding to a word based on the features for the regression model obtained for each of the word chunks to which the label is attached and the labels assigned to each of the word chunks for each of the plurality of original domains. A regression model learning unit for learning a regression model for extracting a word representing an entry;
Model learning device including 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. An applied time series model feature extraction unit that extracts the common domain features;
The features of the common domain for each table component extracted by the sequence model learned by the model learning device according to claim 1 and each word included in the question sentence by the applied time series model feature conversion unit. And a series model application unit that assigns a label representing the table component to each word chunk included in the question 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; The feature model for the regression model when applied to extract the similarity of
Based on the regression model learned by the model learning device and the features for the regression model of each of the word chunks with the label extracted by the application regression model featureization unit, the question sentence A regression model application unit that extracts words representing entries in the database of the target domain corresponding to
Word extraction device including   The sequence model application unit includes a portion corresponding to the feature of the common domain of the sequence model learned by the model learning device, and each word included in the question sentence by the applied time series model feature conversion unit. The word extraction device according to claim 2, wherein a label representing the table component is assigned to each word chunk included in the question sentence based on the extracted feature of the common domain for each table component. A model learning method in a model learning 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,
Based on a set of question sentences of the original domain, wherein the feature modeling unit for the series model is given a word indicating a database entry of the original domain and a label indicating a table component for each of the plurality of original domains , Using the semantic vector of each word created in advance, for each table component, the similarity between the word representing the entry of the table component and the word of the original domain While extracting as a feature, the original domain feature is duplicated as a common domain feature,
The sequence model learning unit gives to each of the plurality of original domains the features of the original domain and the features of the common domain obtained for each word included in the question sentence, and each word included in the question sentence Learning a sequence model for extracting a table component corresponding to a word chunk based on the label
For each of the plurality of original domains, each of the plurality of original domains is provided with the label using the semantic vector of each word based on the set of question sentences of the original domain. , Extracting the similarity with each of the words representing the entry of the table component represented by the label given to the word chunk as a feature for regression model,
The regression model learning unit is based on the regression model features obtained for each of the word chunks to which the label is assigned and the labels given to each of the word chunks for each of the plurality of original domains. A model learning method for learning a regression model for extracting a word representing an entry corresponding to a word. 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,
For each word included in the question sentence, the applied time-series model feature generating unit uses a semantic vector of each word created in advance for each table constituent element of the target domain database. Extract the similarity to the word representing the entry as a common domain feature,
5. The table constituent element extracted by the sequence model application unit for each word included in the question sentence by the sequence model learned by the model learning method according to claim 4 and the applied time series model feature conversion unit Based on the features of each common domain, each word chunk included in the question sentence is given a label representing the table component,
The table represented by the label assigned to the word chunk, for each word chunk to which the label is assigned by the sequence model application unit, using the meaning vector of each word by the application regression model featureizing unit Extract the similarity to each of the words representing the entry of the component as a feature for the regression model,
A regression model application unit, the regression model learned by the model learning method; and a feature for the regression model of each of the word chunks to which the label is assigned, extracted by the application regression model featureization unit. A word extraction method for extracting a word representing an entry in the database of the target domain corresponding to the question sentence based on the word sentence.   When the sequence model application unit assigns a label, a part corresponding to the feature of the common domain of the sequence model learned by the model learning method and the query sentence by the applied time series model feature conversion unit 6. A label representing the table component is assigned to each word chunk included in the question sentence based on the features of the common domain for each table component extracted for each word included in the table. The word extraction method described.   The program for functioning a computer as each part of the model learning apparatus of Claim 1, or the word extraction apparatus of Claim 2 or Claim 3.