JP6857754B2 - Answer candidate generation device, answer candidate generation method, and program - Google Patents

Description

The present invention relates to an answer candidate generation device, an answer candidate generation method, and a program, and more particularly to an answer candidate generation device for generating an answer candidate used in a dialogue system, an answer candidate generation method, and a program.

In general, when talking to a dialogue system such as a robot, the user often speaks and responds. However, recently, there is something like a chat dialogue that the system talks to, and it is being developed on a scenario basis. In the case of an utterance such as asking something from the system, it is necessary for us to recognize the answer to the other party's answer and speak again. When generating such a scenario, it is often the case that the possible answers to the system questions are manually predicted and created in the scenario.

Japanese Unexamined Patent Publication No. 2014-219872

Construction of a natural dialogue platform and application to voice dialogue toys. Kanako Onishi et al. Materials of the Japanese Society for Artificial Intelligence, SIG-SLUD-B502-14, 2015

However, since the response pattern of the other party is enormous, it is difficult to cover most of the answers expected in the dialogue system, and there is a problem that the cost is enormous to execute.

The present invention has been made to solve the above problems, and provides an answer candidate generation device, an answer candidate generation method, and a program capable of obtaining appropriate and various answer candidates for a question sentence. The purpose.

In order to achieve the above object, the answer candidate generation device according to the present invention expresses the concept of the word by representing each of the answer words to the question sentence and each of the words similar to each of the words included in the question sentence. A predetermined language model score and the extended answer for the combination of the extended word candidate generation unit, which is obtained based on the score of the concept vector and used as the extended answer word candidate, and the question sentence and the extended answer word candidate. Clustering based on the concept vector The word set including the answer word candidate having the highest integrated score and the answer word for the question sentence is clustered based on the score of the concept vector representing the concept of the word for the word candidate. It is composed of parts and parts.

Further, according to the answer candidate generation device, the answer candidate generation method, and the program of the present invention, it is possible to obtain an appropriate and diverse answer candidate for the question sentence.

It is a block diagram which shows the structure of the answer candidate generation apparatus which concerns on embodiment of this invention. It is a figure of the comparison example of the search image of the extended word with and without clustering. It is a flowchart which shows the answer candidate generation processing routine in the answer candidate generation apparatus which concerns on embodiment of this invention.

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

<Overview of Embodiments of the Present Invention>

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

In response to the above-mentioned problems, in the present embodiment, statistical information using a large amount of text, language information of the question, and statistical information are used to predict the answer content of the user to the question sentence made by the dialogue system. By doing so, it is possible to support the creation of dialogue scenarios.

The answer candidate generation device of the present embodiment expands the answer words and asks a question by giving a question sentence issued by the system and several words (answer seeds) to answer the question sentence with respect to the scenario of the dialogue system. It is a device that outputs words that give various answers to.

In this embodiment, a large amount of text such as a blog article is statistically processed to create a concept vector expressing the meaning of a word. The extended word candidates to be output are determined based on the similarity of the concept vectors. In addition, the question sentence is linguistically analyzed to generate a sentence pattern. By applying the extended word candidates output based on the concept vector to the sentence pattern and determining whether it is valid in terms of the language model, it is possible to output words that are valid answers to the question with higher accuracy. There is a feature.

Further, a wider variety of extended word candidates can be obtained by re-inputting the set of extended answer words output by the above method as an answer seed. At this time, there is a feature that more diverse extended word candidates can be obtained by performing clustering in which extended word candidates to be input again are grouped together with words having similar meanings and expanding the words in each cluster.

<Structure of Answer Candidate Generation Device According to Embodiment of the Present Invention>

Next, the configuration of the answer candidate generation device according to the embodiment of the present invention will be described. As shown in FIG. 1, the answer candidate generation device 100 according to the embodiment of the present invention includes a CPU, a RAM, a ROM that stores a program for executing an answer candidate generation processing routine described later, and various data. It can be configured with a computer that includes. The answer candidate generation device 100 functionally includes an input unit 10, a calculation unit 20, and an output unit 50 as shown in FIG.

The input unit 10 accepts the input of the question sentence and each of the answer words to the question sentence (hereinafter, also referred to as "(original) answer seed"). In the present embodiment, a case where "what food do you want to eat freshly made?" Is accepted as a question sentence and "bread, fried food, rice, noodles, and grilled meat" are accepted as each of the answer words will be described as an example.

The calculation unit 20 includes a concept vector model generation unit 22, a concept vector model 24, a language model generation unit 26, a language model 28, a sentence pattern generation unit 30, an extended word candidate generation unit 32, and an extended answer word candidate. A reranking unit 34, a clustering unit 36, and a condition determination unit 38 are included.

The concept vector model generation unit 22 generates a concept vector representing the concept of a word for each of the words included in the text set, and stores it as the concept vector model 24. The concept vector is based on the appearance information of words from the text set, and in the present embodiment, the concept vector creation method uses a neural network (Word2vec). Further, any one may be used, such as one using singular value decomposition. However, since the extended word candidate generation unit 32, which will be described later, performs a process of synthesizing a plurality of word vectors, the concept vector adopted by the concept vector model generation unit 22 can also be used as a synthesis vector. The synthesis method is not specified.

The language model generation unit 26 generates a language model 28 based on N grams of words included in the text set. At this time, the method of creating the language model is not specified, but in order to utilize the model efficiently, auxiliary verbs such as "ha" and "ga" are replaced with "*". However, the language model 28 created here is in a format in which the sentence pattern generated by the sentence pattern generation unit 30 described later can be used.

The sentence pattern generation unit 30 generates a sentence pattern with a tag to be replaced with a word from the question sentence received by the input unit 10. Specifically, first, the sentence pattern generation unit 30 parses the input question sentence. The parsing here is a dependency analysis. First, the 5W1H expressions such as "why", "what", "when", "where", and "who" are searched for the interrogative analysis. The corresponding 5W1H expression is replaced with a tag (<cand>) for replacing it with a word. Next, find the phrase that contains the verb or verb stem in the dependency analysis result. With this phrase as the standard phrase, the remaining phrases are divided into before and after the standard phrase, and the phrases are synthesized according to the dependency relationship for each. Here, when synthesizing with the phrase before the standard phrase, first add "<cand> *" including the tag (* is an arbitrary particle) before the standard phrase, and then the phrase by the dependency relationship. Is synthesized. To combine with the phrase after the reference phrase, add the replacement tag "<cand>" to the end of the last phrase of the question sentence (remove punctuation marks and question marks), and repeat the composition from there according to the dependency relationship.

The standard phrase of the question "What food do you want to eat freshly?" Is the "I want to eat" part. In the composition with the phrase before this standard phrase, two sentence patterns "<cand> * I want to eat" and "<cand> * I want to eat freshly made" are generated, and by combining with the phrase after this, "Food is Three sentence patterns are generated: "<cand>", "the food you want to eat is <cand>", and "the food you want to eat freshly is <cand>". Therefore, a total of five sentence patterns are generated. The replacement tag <cand> may be set freely, or a sentence pattern in which particles such as "ha" are replaced with * may be created.

The following is an output example of a sentence pattern. In this example, particles such as "ha" are replaced with *.

Sentence pattern 1: <cand> * Sentence pattern 2: <cand> * Freshly made * Sentence pattern 3: Food * <cand>
Sentence pattern 4: Food you want to eat * <cand>
Sentence pattern 5: Freshly made * Food you want to eat * <cand>

The extended word candidate generation unit 32 obtains each of the answer words to the question sentence received by the input unit 10 and each of the words similar to each of the words included in the question sentence based on the concept vector model 24, and extends the words. Candidates for answer words. Further, for each cluster obtained by the clustering unit 36 described later, each of the words belonging to the cluster and each of the words similar to each of the words included in the interrogative sentence are obtained based on the concept vector model 24 and expanded. Candidates for answer words.

The extended word candidate generation unit 32 uses the concept vector of the word of the concept vector model 24 to generate a word that is a candidate for the extended answer word to be output. Let α be a set of answer words (answer seeds) to be input. Prepare a set Q of the results of morphological analysis of the question sentence. Here, the composite vector of the concept vector of the word _{set α + q} when a certain word q in Q is added to α

Then, the score of the concept vector of a certain word w can be calculated by the following equation (1).

・・・（１）

... (1)

here,

Is a composite vector created for each of the words q

And the concept vector of the word w

The degree of similarity with. Any method of calculating the degree of similarity can be used as long as it indicates that the larger the numerical value, the higher the similarity. The above scores are calculated for all the words included in the conceptual vector model, and the top n words when the scores are arranged in descending order are used as extended answer word candidates in the next extended answer word candidate reranking unit 34. ..

Further, in the extended word candidate generation unit 32, for example, after clustering by the clustering unit 36, a set Q _k of extended answer word candidates for each cluster is obtained, so the above score is calculated for each cluster for the word w. Then, the top n words when the scores calculated for all the clusters are arranged in descending order are used as the extended answer word candidates in the next extended answer word candidate reranking unit 34. A score obtained by summing the scores calculated for each cluster for the word w may be used. Further, a calculation method may be used in which the number of words belonging to the cluster to which the extended answer word candidate belongs is introduced as a coefficient into the calculation of the above equation (1). For example, if the number of words in all clusters is 100 and the number of words in a certain cluster is 5, 5/100 may be used as a coefficient.

The extended answer word candidate re-ranking unit 34 has the score of the language model 28 and the score obtained when the tag included in the sentence pattern is replaced with the extended answer word candidate for each of the upper n extended answer word candidates. The integrated score is calculated based on the score of the concept vector calculated for the extended answer word candidates, and the extended answer word candidates having the top N integrated scores are obtained. Similarly, after clustering by the clustering unit 36, for each of the extended answer word candidates of the upper n words, the score of the language model 28 obtained when the tag included in the sentence pattern is replaced with the extended answer word candidate, And the integrated score based on the score of the concept vector calculated for the extended answer word candidate is calculated, and the extended answer word candidates having the top N integrated scores are obtained. The integrated score after clustering is obtained based on the score of the language model 28 and the score obtained by multiplying the score of the concept vector by the number of words belonging to the cluster to which the extended answer word candidate belongs as a coefficient.

Specifically, in the extended answer word candidate re-ranking unit 34, first, for each of the extended answer word candidates, the replacement tag <cand> part of the sentence pattern is replaced with the extended answer word candidate, so that the extended answer word candidate can be replaced. Apply each word of to a sentence pattern. Then, the replaced sentence is input to the language model 28 to calculate the score of the language model 28. Here, any method can be used for calculating the score of the language model 28. In the present embodiment, for example, the conditional probability of a word is used as the score in a general language model, and the score Score (w) of the extended answer word candidate w is calculated as shown in the following equation (2).

・・・（２）

... (2)

Here, in the above equation (2), w represents a certain extended answer word candidate, and p represents a sentence pattern. In the above equation (2), Score (w, p) is calculated for each combination of the extended answer word candidate w and all sentence patterns p, and long (p) is multiplied as a weight and totaled. lang (p) represents the word length of the sentence pattern p. The longer the word length, the greater the weight, and by considering the word length, the influence of the long sentence pattern becomes greater. Since a long sentence pattern is closer to a question sentence containing modifiers, words that are more likely to appear in a long sentence pattern are more related to the question sentence.

Further, as an alternative to the language model 28, a Web search using the sentence pattern p in which the extended answer word candidate w is replaced as a query (q (w, p)) is executed, and the number of search hits is used as a score as described below (3). It is also possible to calculate Sore (w, p) as in the formula).

・・・（３）

... (3)

Here, H (q (w, p)) indicates the number of search hits when a Web search is performed with the query q (w, p). At this time, using the Sore (w, p) calculated by the above equation (3), the score Score (w) of the extended answer word candidate w is calculated according to the right side of the above equation (2).

Then, for each of the extended answer word candidates w, the final integration is performed using the score Score (w) of the language model 28 and the score of the concept vector calculated for the extended answer word candidate w by the extended word candidate generation unit 32. Calculate the score. Here, as the method for calculating the integrated score, any method can be used as long as both the score of the language model 28 and the score of the concept vector are taken into consideration. The average value of the score of the language model 28 and the score of the concept vector may be used, or the ranking of the scores arranged in descending order by the score of the language model 28 and the ranking when arranged in descending order by the score of the concept vector. You may use the average value of.

When the extended answer word candidates are arranged in descending order of the integrated score, the top N items are output to the output unit 50. Further, it is a word set used in the next clustering unit 36. Here, the number of outputs n of the extended answer word candidates and the final number of outputs N of the extended word candidate generation unit 32 may be freely set if n> N. Further, the integrated score calculation method when clustering is used may be the same calculation method as above. In addition to the above calculation method, for example, the extended answer word candidate is added to the score of the concept vector of the extended answer word candidate. A calculation method may be used in which the number of words belonging to the cluster to which the word belongs is introduced as a coefficient. This is because most of the extended answer word candidates output by the extended answer word candidate reranking unit 34 match the answer seed and the question sentence, but some of them include extended answer word candidates that may be incorrect. There is. Since such extended answer word candidates are often words that tend to differ significantly from other outputs, they tend to form independent clusters. If the number of words is used as a coefficient for processing, the score of the extended answer word candidates output from a small cluster (cluster with a small number of words) tends to be small, so many such extended answer word candidates are output. It is possible to prevent it from being done.

The following is an output example of the top 10 extended answer word candidates in the integrated score when N = 10.

Fried chicken, fried chicken, hamburger steak, curry rice, fried noodles, side dish, rice bowl, pasta, pork cutlet, fried chicken

The following is an output example of the 10 extended answer word candidates with the highest integrated score obtained by clustering the above 10 cases and using the extended answer word candidates obtained for each cluster.

Croquette, fried rice, omelet rice, dumplings, cod roe spa, potato fried rice, fried rice, katsudon, fried rice, crab cream croquette

The clustering unit 36 clusters a word set including the top N extended answer word candidates and the answer words for the question sentence based on the concept vector model 24. Any method may be used for the clustering method. A method of setting the number of clusters in advance (k-means method) or a method of automatically determining the number of clusters (Affinity Propagation) may be used.

The reason for performing clustering in this embodiment will be described below.

In the present embodiment, an extended answer word set with an extended answer seed can be obtained by giving an input question sentence and an answer seed. However, the extended word candidate generation unit 32 and the extended answer word candidate up to this point can be obtained. By executing the process of the reranking unit 34 a plurality of times, a more diverse and large number of extended answer word candidates can be obtained. In that case, the N words output by the extended answer word candidate reranking unit 34 are used as the answer seed for the next extension. By using the set of output extended answer word candidates as the next answer seed, it is possible to obtain words that are different from the original answer seed that was initially input and the previous output, but the output N extended answer words. The set of candidates has more words than the original answer seed. When the extended answer word candidate of the extended word candidate generation unit 32 is calculated in this state, the composite vector of the concept vector of the word created from the answer seed becomes the average of the meanings of many words. , There was a problem that the variety of available words was lost.

In order to solve the above problem, the clustering unit 36 performs clustering based on the similarity of the concept vectors of the words of the concept vector model 24 for the N words used as the answer seeds, and divides the N words into a plurality of clusters. Then, the extended word candidate generation unit 32 performs the calculation of the above equation (1) for each cluster, and obtains the top n extended answer word candidates for all the clusters. Then, the extended answer word candidate reranking unit 34 calculates the score Score (w) of the extended answer word candidate w by the above equation (2) for each of the extended answer word candidates, and further calculates the integrated score. By doing so, the output words for the second and subsequent loops are determined. When clustering is performed, it is assumed that the word set is obtained by adding the word set α of the original answer seed to the newly used N word. The N words output by the extended answer word candidate re-ranking unit 34 are words that are an extension of the original answer seed, but because the question sentence and the language model are taken into consideration, they are a set of words that are more diverse than the answer seed. If a cluster is created in this state, the clusters can be easily distributed. In addition, some N words are output due to analysis errors, and if a cluster is created with a set of these inappropriate words, the answers output from the second time onward will be inappropriate words. There is a possibility that it will be occupied. Therefore, the word set α of the original answer seed is likely to have similar meanings, and it is definitely a correct word. Therefore, clustering including the original answer seed word set α improves the cohesiveness of the cluster. Moreover, it works to prevent a large number of inappropriate clusters from being generated.

By expanding words for each cluster, it is possible to obtain more diverse extended words. FIG. 2 shows a comparison of search images of extended words with and without clustering. Each black circle in FIG. 2 indicates an answer seed word expressed by a concept vector, and an asterisk indicates a composite vector. The range surrounded by the dotted line shows the range that can be searched by the similarity of the composite vector. In this example, it is assumed that the composite vector is represented by the centroid of the conceptual vector of each word.

The following is an example of cluster output.

Cluster (A): Curry, curry rice, Yakisoba cluster (B): Delicatessen cluster (C): Pasta cluster (D): Fried chicken, hamburger steak, rice bowl, pork cutlet, fried chicken

The condition determination unit 38 repeats the processes of the extended word candidate generation unit 32, the extended answer word candidate reranking unit 34, and the clustering unit 36 until a predetermined condition is satisfied. The condition may be, for example, the number of times or the number of output words.

<Operation of Answer Candidate Generation Device According to the Embodiment of the Present Invention>

Next, the operation of the answer candidate generation device 100 according to the embodiment of the present invention will be described. In the answer candidate generation device 100, the concept vector model generation unit 22 obtains the concept vector model 24, and the language model generation unit 26 obtains the language model 28 in advance. When the input unit 10 receives each of the question sentence and the answer word for the question sentence, the answer candidate generation device 100 executes the answer candidate generation processing routine shown in FIG.

First, in step S100, a sentence pattern with a tag to be replaced with a word is generated from the question sentence received by the input unit 10.

Next, in step S102, each of the answer words to the question sentence received by the input unit 10 and each of the words similar to each of the words included in the question sentence are converted into a conceptual vector model calculated by the above equation (1). Obtained based on the score of the concept vector of words in 24, the top n words when the scores are arranged in descending order are set as extended answer word candidates.

In step S104, the score of the language model 28 calculated by the above equation (2) obtained when the tag included in the sentence pattern is replaced with the extended answer word candidate, and the concept vector calculated for the extended answer word candidate. Find the extended answer word candidates with the highest N (n> N) integrated scores based on the score.

In step S106, clustering is performed on a word set including the top N extended answer word candidates obtained in step S104 or the previous step S110 and the answer words for the question sentence based on the concept vector model 24.

In step S108, for each cluster, each of the words belonging to the cluster and each of the words similar to each of the words included in the interrogative sentence are converted into the word concept vector in the concept vector model 24 calculated by the above equation (1). The top n words when the scores are arranged in descending order are used as extended answer word candidates.

In step S110, the score of the language model 28 obtained when the tag included in the sentence pattern obtained in step S108 is replaced with the extended answer word candidate obtained for the cluster is calculated, and the score of the language model 28 is calculated by the above equation (2). The score of the calculated language model 28 and the integrated score based on the score of the conceptual vector calculated for the extended answer word candidates are obtained for the top N extended answer word candidates. Here, the integrated score is obtained based on the score of the language model 28 and the score obtained by multiplying the score of the concept vector by the number of words belonging to the cluster to which the extended answer word candidate belongs as a coefficient.

In step S112, it is determined whether or not the condition for ending the repetition is satisfied, and if the condition is satisfied, the process proceeds to step S114, and if not satisfied, the process returns to step S106 and the process is repeated.

In step S114, the extended answer word candidates obtained in the processes of steps S104 to S112 are output to the output unit 50 to end the process.

As described above, according to the answer candidate generation device according to the embodiment of the present invention, a sentence pattern with a tag to be replaced with a word is generated and included in each of the answer words to the question sentence and the question sentence. A language model obtained when each word similar to each of the words is obtained based on the score of the concept vector to be an extended answer word candidate, and the tag included in the sentence pattern is replaced with the extended answer word candidate. Score and concept of extended answer word candidates The integrated score based on the score of the vector finds the top N extended answer word candidates, and makes a word set containing the top N extended answer word candidates and the answer words to the question sentence. On the other hand, clustering is performed based on the concept vector, extended word candidate generation, extended word candidate reranking, and clustering process are repeated until predetermined conditions are satisfied, and each cluster is asked with each word belonging to the cluster. Each of the words similar to each of the words contained in the sentence is obtained based on the score of the concept vector to be the extended answer word candidate, and the tag included in the sentence pattern is replaced with the extended answer word candidate obtained for the cluster. By calculating the score of the language model obtained at the time, and finding the top N extended answer word candidates with the integrated score based on the score of the concept vector for the language model and the extended answer word candidates in all clusters. , You can get various answer candidates for use in the dialogue system.

By using the method of the embodiment of the present invention, the scenario creator of the dialogue system can give a question sentence uttered by the system and a word to answer the question as a few words (about 5 words) as an answer seed. , Various answer words (extended answer word candidates) can be obtained.

Further, in an open domain dialogue such as a chat dialogue, since the answers to the questions are diverse, it is difficult for the scenario creator to recall all the answers. However, by using the method of the embodiment of the present invention, Since the scenario creator does not have to think about many answers with confidence, there is an advantage that the cost of creating the scenario can be reduced. At the same time, in the dialogue scenario, the number of predictable answers increases, so that the possibility of being able to give a satisfactory response to the answers of more users increases, and thus the quality and satisfaction of the dialogue system can be expected to improve.

In addition, the method of the embodiment of the present invention extends the prediction of the user's response content to the dialogue scenario. In a dialogue scenario, the system generally asks a question, the user answers the question, and then the system receives the user's answer and speaks another word. Also in the embodiment of the present invention, the system creator creates the third utterance of the system by himself / herself in response to the extended answer. Here, the third utterance may be created with reference to the answer word output in the embodiment of the present invention. This is because, in the embodiment of the present invention, a large number of answer words are output from various viewpoints, and the answer words can be used not only for predicting the user's answer but also as the third utterance. It is also possible to automatically generate the dialogue scenario itself by combining the conventional automatic utterance generator with the creation of the third utterance.

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

For example, in the above-described embodiment, the case of configuring as an answer candidate generation device has been described as an example, but the present invention is not limited to this. For example, it is configured as a clustering device of only the clustering unit, accepts each of the question sentence and the answer word for the question sentence, and the upper answer word candidate obtained by the external device as input, and outputs the clustering result of the answer word candidate. You may do so.

10 Input unit 20 Calculation unit 22 Concept vector model generation unit 24 Concept vector model 26 Language model generation unit 28 Language model 30 Sentence pattern generation unit 32 Extended word candidate generation unit 34 Extended answer word candidate reranking unit 36 Clustering unit 38 Condition judgment unit 50 Output unit 100 Answer candidate generator

Claims (6)

Each of the answer words to the question sentence and each of the words similar to each of the words included in the question sentence are obtained based on the score of the concept vector representing the concept of the word, and the extended word is used as the extended answer word candidate. Candidate generator and
An extended answer word with a higher integrated score based on a predetermined language model score for the combination of the question sentence and the extended answer word candidate and a score of a concept vector representing the concept of the word for the extended answer word candidate. A clustering unit that clusters a word set including a candidate and an answer word to the question sentence based on the concept vector.
Answer candidate generator including. Further including a condition determination unit that repeats the processing of the extended word candidate generation unit and the clustering unit until a predetermined condition is satisfied.
In the iterative processing by the condition determination unit,
In the extended word candidate generation unit, for each cluster obtained by the clustering unit, each of the words belonging to the cluster and each of the words similar to each of the words included in the interrogative sentence are scored in the concept vector. And use it as the extended answer word candidate.
The answer candidate generation device according to claim 1. Each of the answer words to the question sentence and each of the words similar to each of the words included in the question sentence are obtained based on the score of the concept vector representing the concept of the word, and the extended word is used as the extended answer word candidate. Candidate generator and
An extension for finding the top N extended answer word candidates with a score obtained when a predetermined word in the question sentence is replaced with an extended answer word candidate and an integrated score based on the score of the concept vector for the extended answer word candidate. Answer word candidate re-ranking department and
A clustering unit that clusters a word set including the top N extended answer word candidates and answer words for the question sentence based on the concept vector.
A condition determination unit that repeats the processing of the extended word candidate generation unit, the extended answer word candidate reranking unit, and the clustering unit until a predetermined condition is satisfied.
Including
In the iterative processing by the condition determination unit,
In the extended word candidate generation unit, for each cluster obtained by the clustering unit, each of the words belonging to the cluster and each of the words similar to each of the words included in the interrogative sentence are scored in the concept vector. An answer candidate generation device that is obtained based on the above and is used as the extended answer word candidate. The extended word candidate generation unit finds each of the answer words to the question sentence and each of the words similar to each of the words contained in the question sentence based on the score of the concept vector representing the concept of the word, and extends it. Steps to make answer word candidates and
The clustering unit ranks higher in the integrated score based on the score of the predetermined language model for the combination of the question sentence and the extended answer word candidate and the score of the concept vector representing the concept of the word for the extended answer word candidate. A step of clustering a word set including the extended answer word candidate and the answer word for the question sentence based on the concept vector.
Answer candidate generation method including. The extended word candidate generation unit finds each of the answer words to the question sentence and each of the words similar to each of the words contained in the question sentence based on the score of the concept vector representing the concept of the word, and extends it. Steps to make answer word candidates and
The score obtained when the extended answer word candidate re-ranking unit replaces a predetermined word in the question sentence with the extended answer word candidate, and the integrated score based on the score of the concept vector for the extended answer word candidate are the top N. Steps to find extended answer word candidates and
A step in which the clustering unit clusters a word set including the top N extended answer word candidates and the answer word for the question sentence based on the concept vector.
A step in which the condition determination unit repeats the processes of the extended word candidate generation unit, the extended answer word candidate reranking unit, and the clustering unit until a predetermined condition is satisfied.
Including
In the iterative processing by the condition determination unit,
In the extended word candidate generation unit, for each cluster obtained by the clustering unit, each of the words belonging to the cluster and each of the words similar to each of the words included in the interrogative sentence are scored in the concept vector. A method of generating an answer candidate, which is obtained based on the above and is used as the extended answer word candidate. A program for causing a computer to function as each part of the answer candidate generation device according to any one of claims 1 to 3.

Images