JP2023149601A - Inference apparatus, inquiry reply apparatus, interactive apparatus, and inference method - Google Patents

Abstract

To provide an inference apparatus that can operate at a high speed and with a sufficient accuracy using few computational resources.SOLUTION: An inference apparatus 50 includes a first neural net 80 that outputs a vector representation of a first input, and a second neural net that outputs a vector representation of a second input, and when there is a predetermined relation between the vector representation of the first and second inputs using learning data of the first and second inputs which has a predetermined relation, causes the first and second neural nets to be learned so as to be located close to each other in a vector space, and clusters the vector representation which is an output of the learned second neural net. The inference apparatus 50 further includes a database 84 constructed in advance so as to enable retrieval extraction of clusters on the basis of the vector representation of the first input, and infers an output on the basis of information 88 of the clusters retrieved and extracted from the database 84 on the basis of the vector representation of the input by the first neural net 80 with respect to an input 60.SELECTED DRAWING: Figure 1

Description

The present invention relates to a natural language reasoning device, question answering device, dialogue device, and reasoning method.

A conventionally known question answering system is a system (WISDOM X) developed by the applicant of the present invention and provided on the web. This system mainly extracts a plurality of content words, which form a keyword group, from an input question. Based on this keyword group, several passages (groups of sentences consisting of about seven consecutive sentences) collected from the Internet and the like are selected. By inputting the obtained passage group together with the questions into the neural network, the passages are classified from the viewpoint of whether or not they contain answers to the questions. For passages that are determined to contain answers, phrases and words that serve as answers to questions are extracted from them, formatted, and output as answers.

Due to the nature of the question-and-answer system, it is necessary to answer questions quickly. To this end, in the above system, the web is crawled in advance and data on the web is stored in a local storage device. The search range of the keyword-based search engine described above is the information stored in this storage device, resulting in a large number of answer candidates. Therefore, the process of determining whether each answer candidate includes an answer to the question involves processing a large amount of data. Furthermore, since each passage to be processed has a certain amount of data and requires natural language processing, a large-scale neural network is used for the processing. As a result, the total amount of data processing that must be ultimately performed becomes extremely large. Therefore, there is a problem in that large computational resources are required to operate the conventional question answering system.

These circumstances are not limited to question-and-answer systems. The applicant has also developed a dialogue system that responds not only to questions but also to general input. In this dialog system, in order to create response candidates, a plurality of questions are generated from input, and the questions are input into the above-described question answering system to obtain a plurality of answers. Thereafter, a response to the input utterance is generated from those answers, and further processing is performed to select one optimal response from the plurality of generated responses. The amount of data processing required for this exceeds the amount of data processing for a question answering system. As a result, there is a problem in that very large computational resources are required to operate the dialogue system.

Therefore, it is an object of the present invention to provide an inference device, a question answering device, a dialogue device, and an inference method that can operate at high speed and with sufficient accuracy using fewer computational resources than conventional ones.

An inference device according to a first aspect of the invention includes a first neural network supplied with a first input and outputting a vector representation of the first input, and a first neural network supplied with a second input and outputting a vector representation of the first input. a second neural network that outputs a vector representation of the input of the first input and a second neural network that outputs the vector representation of the first input and the second input using at least learning data of the first and second inputs having a predetermined relationship. The first and second neural networks are trained to be located close to each other in the vector space when the vector representations of , further includes a database constructed in advance to enable clustering based on the position on the vector space and search and extraction of clusters based on the vector representation of the first input; Outputs having a predetermined relationship are inferred based on cluster information searched and extracted from a database based on the vector representation of input by the network.

Preferably, the database is searched using a centroid of vector representations of the outputs included in each cluster.

More preferably, the first neural network is additionally trained based on a first input and a vector representation associated with a cluster to which a second input having a predetermined relationship with the first input belongs.

A question answering device according to a second aspect of the invention includes any of the above-mentioned inference devices, and the predetermined relationship includes a question and an answer to the question.

A dialogue device according to a third aspect of the invention includes any of the above-mentioned inference devices, and the predetermined relationship includes an utterance and a response to the utterance.

An inference method according to a fourth aspect of the invention includes a first neural network that is supplied with a first input and outputs a vector representation of the first input; a second neural network that outputs a vector representation of an input; and a second neural network that outputs a vector representation of an input of a step of training the first and second neural networks so that the vector representations of the two inputs are located close to each other in the vector space when they have a predetermined relationship; and an output of the trained second neural network. Clustering the vector representations based on the positions on the vector space, and constructing a database in advance to enable search and extraction of clusters based on the vector representation of the first input; and inferring an output having a predetermined relationship based on cluster information searched and extracted from a database based on a vector representation of an input by the neural network.

A question answering device according to a fifth aspect of the invention includes a database including a plurality of records, each of which includes a link destination of an answer candidate for a question and an identifier of a cluster to which a vector representation of the answer candidate belongs. a neural network for converting a question into a question vector that is a vector representation of the question in response to input of a question; and a cluster of semantic expression vectors of answer candidates. Cluster selection means for selecting a predetermined number of clusters whose representative vector of the cluster is closest to the question vector; and a link destination included in a record having an identifier of one of the predetermined number of clusters selected by the cluster selection means in the database. , and an answer selection means for selecting an answer to the question text in a predetermined procedure from among the answer candidates collected by the answer candidate collecting section.

Preferably, the question answering device further includes an utterance history storage means for storing the utterance history of the questioner who has uttered the question, and an utterance history storage means for storing the utterance history of the questioner who has uttered the question. topic identification means for identifying one or more topic candidates in the past utterance history of the questioner who uttered the question based on the utterance history stored in the utterance history storage means; The method further includes topic adding means for adding information representing one or more topics specified by the topic specifying means to the question text.

More preferably, each of the answer candidates includes a plurality of consecutive sentences.

A dialogue device according to a sixth aspect of the invention includes the above-described question answering device, topic estimation means for estimating the topic of an input utterance, and information indicating the topic estimated by the topic estimation means for the input utterance. and a topic adding means for adding a question to the question answering device and inputting it as a question to the question answering device, and a response generating means for generating a response to the input utterance by formatting the output of the question answering device in response to the utterance to be suitable for dialogue. .

A method for training a question-answering model according to a seventh aspect of the present invention is based on a plurality of question sentences and a group of answer candidates for each of the plurality of question sentences. A step of preparing learning data by generating a positive example consisting of a combination of a question sentence and a negative example consisting of a combination of a question sentence and an answer candidate that does not correspond to the question sentence; A question conversion neural network for converting into a question vector, which is a representation, and an answer conversion neural network, which converts an answer candidate into an answer candidate vector, which is a vector representation of the answer candidate, are trained using a Siamese network. and converting an answer candidate included in the answer candidate group into an answer candidate vector that is a vector representation of the answer candidate using the answer conversion neural network trained in the step of training. Clustering the generated answer candidate vectors into a predetermined number of clusters, assigning to each answer candidate an identifier of the cluster to which the answer candidate belongs, and determining an answer candidate corresponding to the question for each of the plurality of question sentences. , and a predetermined index representing the distance between the question vector output by the question conversion neural network for each of the plurality of question sentences and the cluster to which the answer candidate corresponding to the question belongs is small. and a step of additionally learning the question conversion neural network so that the question conversion neural network becomes

Preferably, the method for training a question answering model further includes the step of collecting a plurality of question sentences from the Internet, prior to the step of preparing training data.

More preferably, the collecting step includes, prior to the step of preparing the learning data, collecting a plurality of question sentences from the Internet together with sentences before or after the question sentences, and for each of the plurality of question sentences, the question sentence estimating a topic related to the question based on previous or subsequent sentences collected for the question, and the step of preparing learning data includes estimating a topic related to the question for each of the plurality of question sentences. A positive example consisting of a combination of a question sentence and an answer candidate corresponding to the question sentence, a positive example consisting of a combination of a question sentence and an answer candidate corresponding to the question sentence, from a plurality of question sentences to which topics have been assigned, and an answer candidate group, and preparing learning data by generating negative examples consisting of combinations with non-corresponding answer candidates.

More preferably, the method for training a question answering model further includes, prior to the step of preparing learning data, collecting a group of answer candidates for each of the plurality of question sentences from the Internet; storing, for each candidate answer, an identifier of a cluster to which the candidate answer belongs, and a URL on the Internet of the candidate answer, in association with each of the candidate answers. and adding a new record to the database containing the above URL.

The above and other objects, features, aspects and advantages of the present invention will become apparent from the following detailed description of the invention, understood in conjunction with the accompanying drawings.

FIG. 1 is a functional block diagram of a question answering device according to a first embodiment of the present invention. FIG. 2 is a functional block diagram of a learning device that learns the question BERT in the question answering device according to the first embodiment. FIG. 3 is a functional block diagram of the BERT learning section shown in FIG. 2. FIG. 4 is a schematic diagram illustrating problems that may occur due to question BERT before additional learning. FIG. 5 is a schematic diagram illustrating problem solving by question BERT after additional learning. FIG. 6 is a functional block diagram of the additional learning section shown in FIG. 2. FIG. 7 is a functional block diagram of an interaction device according to a second embodiment of the invention. FIG. 8 is a functional block diagram of a learning device that performs learning of the dialog history management section in the dialog device according to the embodiment in FIG. FIG. 9 is a functional block diagram of the additional learning section shown in FIG. 8. FIG. 10 shows the external appearance of a computer system that realizes a question answering device 50 according to a first embodiment of the present invention, an interaction device 350 according to a second embodiment, and a learning device for learning a neural network used therein. It is a diagram. FIG. 11 is a block diagram showing the hardware configuration of the computer system shown in FIG. 10.

In the following description and drawings, identical parts are provided with the same reference numerals. Therefore, detailed description thereof will not be repeated.

1st Embodiment 1. Configuration A. Question Answering Device FIG. 1 shows a functional block diagram of a question answering device 50 according to a first embodiment of the present invention. Referring to FIG. 1, a question answering device 50 receives a question 60, and includes an answer candidate search section 64 for searching a passage group that becomes an answer candidate from the Internet 62, and a passage group searched by the answer candidate search section 64. and a passage DB (Database) 66 for storing. The number of passage groups stored in the passage DB 66 is limited compared to the conventional one. The question answering device 50 further includes an answer generation unit 68 for generating and outputting an answer 70 to the question 60 based on the passage group stored in the passage DB 66 using a method similar to a conventional method. The method by which the answer generation unit 68 generates the answer 70 in this embodiment is exactly the same as in the prior art. However, the number of passages to be processed by the answer generation unit 68 is much smaller than in the prior art.

A feature of this embodiment is that the number of passages to be searched from the Internet 62 at the time the question 60 is received is limited. Therefore, the number of passages stored in the passage DB 66 is small, and the calculation resources required by the answer generation section 68 are also far fewer than in the past.

The answer candidate search unit 64 includes a question BERT 80 for converting the question 60 and outputting a question vector 82 that is a vector expressing the question 60. The question BERT80 has been trained in advance, and the learning will be described later with reference to FIG. 2 and subsequent figures. Here, the question BERT etc. includes a neural network Bidirectional Encoder Representation from Transformers, and is a neural network including Transformer encoders such as BERT, RoBERTa, etc., and generally uses pre-learning using language resources. to the task Used after appropriate fine tuning. Fine tuning and the like in this embodiment will be described later.

The answer candidate search unit 64 further includes a centroid DB 84. The centroid DB 84 is generated simultaneously during the training process such as the query BERT 80. The method for generating the centroid DB 84 will be described later along with the explanation of learning. Simply put, the centroid DB 84 vectorizes and clusters many passages containing answers to various questions, and creates a database of centroids. Each record in the centroid DB 84 includes the cluster identifier of the cluster represented by the centroid and a vector representing the position of the centroid in the vector space. A vectorized passage is called a passage vector. The passage vector and the question vector 82 are vectors with the same number of dimensions.

The answer candidate search unit 64 further receives the question vector 82, searches the centroid DB 84 for records of a predetermined number of centroids closest to the question vector 82 in the vector space, and searches for clusters of the predetermined number of searched centroids. It includes an answer candidate cluster identification unit 86 for outputting identifiers 88, respectively. It is assumed that each cluster represented by these cluster identifiers 88 includes a passage vector corresponding to a passage that is likely to include an answer to the question 60. Why this is so will be made clear in the explanation regarding the training of question BERT80.

The answer candidate search unit 64 further includes an answer candidate DB 90 for storing in advance a large number of passage records representing passages that are answer candidates for various questions. Each record in the answer candidate DB 90 is created from information obtained by downloading a passage that is considered to contain an answer to a previously created question from the web. More specifically, each record includes a URL indicating the location where the passage was located and a cluster identifier of the cluster to which the passage vector belongs.

The answer candidate search unit 64 further receives the cluster identifier 88 output by the answer candidate cluster specifying unit 86, searches the answer candidate DB 90 for a record having a cluster identifier that matches the cluster identifier 88, and returns the URL of the searched record. Access the answer candidate search unit 92 for output as a candidate URL group 94 and each of the URLs included in the answer candidate URL group 94, download text containing passages that are considered to include answer candidates, and extract the passages. and a passage search unit 96 for storing in the passage DB 66.

As described above, the passage search section 96 accesses only URLs that belong to the cluster corresponding to the centroid specified by the answer candidate cluster specifying section 86. Therefore, there is no need to access large amounts of data downloaded from across the web as in the prior art. Only those downloaded from URLs belonging to the selected number of clusters are to be accessed. Therefore, even if the answer generation section 68 is the same as the conventional one, the required storage capacity and calculation resources will be much smaller.

B. Learning Device FIG. 2 shows a functional configuration of a learning device 150 for learning the question BERT 80 shown in FIG. 1 and simultaneously generating the centroid DB 84 and answer candidate DB 90. Referring to FIG. 2, the learning device 150 includes a question/passage collection unit 160 for collecting a large number of questions and a large number of passages including answer candidates for each question from the Internet 62; includes a question DB 162 for storing questions collected by the question/passage collection unit 160, and a passage DB 164 for storing passages collected by the question/passage collection unit 160. In FIG. 2, BERT learning and passage clustering are performed using the passages stored in the passage DB 164, but this is done in another DB, for example, a DB consisting of part of the passages stored in the passage DB 164. You can also use

Each record in the question DB 162 corresponds to each question. Each record includes, for example, a question identifier, a cluster identifier of the cluster to which the passage corresponding to the question mainly belongs, the text of the question, and the URL on the Internet where the question exists.

Each record of the question/passage collection unit 160 corresponds to each passage. Each record includes a passage ID, a corresponding question ID, and a URL on the Internet where the passage exists.

The learning device 150 further includes a training target question BERT 168 having the same configuration as the question BERT 80 shown in FIG. and a BERT learning unit 166 for simultaneously learning a question BERT 168 and an answer candidate BERT 170 using Siamese BERT Networks. The Siamese BERT network is described in the following references: An outline of the Siamese BERT network in this embodiment will be described later with reference to FIG. 3. Note that the present invention is not limited to the configuration using the Siamese BERT network, and there is no problem as long as the vector representations that are the outputs of the question BERT 80 and the answer candidate BERT 170 are close or distant but consistent with the learning data.

[References]
Nils Reimers and Iryna Gurevych, “Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks,” [online], August 27, 2019, Arxiv.org, [searched March 10, 2020], Internet ,＜https://arxiv.org/pdf/1908.10084＞
The learning device 150 further converts each passage included in the question/passage collection unit 160 into a passage vector using the answer candidate BERT 170 trained by the BERT learning unit 166, and further clusters the passages to create an answer candidate DB 90 and a centroid DB 84. It includes a passage clustering unit 172 for generating. Generation of the answer candidate DB 90 by the passage clustering unit 172 is performed according to the following procedure. First, by inputting each passage included in the question/passage collection unit 160 to the answer candidate BERT 170, it is converted into a vector obtained as an output of the answer candidate BERT 170. These vectors are passage vectors representing each passage. The passage clustering unit 172 further clusters these passage vectors into a predetermined number of clusters. The number of clusters may be determined in advance, or may be determined within a certain range without causing any problems based on predetermined criteria, such as minimizing the total variance of vectors within the cluster.

After clustering the passage vectors in this manner, the passage clustering unit 172 determines the centroid of each cluster. Furthermore, each centroid is given an identifier of the cluster represented by that centroid. The passage clustering unit 172 records the information regarding each centroid obtained in this manner and generates a centroid DB 84. Specifically, each record in the centroid DB 84 includes a cluster identifier and a vector of its centroid. Naturally, the centroid vector has the same dimension as the passage vector and the question vector.

Note that the passage clustering unit 172 generates each record in the answer candidate DB 90 as follows. The passage clustering unit 172 sets, for each record in the question/passage collection unit 160, the cluster identifier of the cluster to which the passage vector obtained from the passage of that record belongs and the URL of that record as a set and records it in the answer candidate DB 90. Register. The passage clustering unit 172 also performs, for each record in the question DB 162, a process of assigning the centroid ID of the cluster to which the passage obtained for that question DB 162 mainly belongs.

The learning device 150 further includes an additional learning unit 174 for performing additional learning of the question BERT 168 using the centroid DB 84 and the question DB 162. The question BERT168 that has been trained by the BERT learning unit 166 becomes the question BERT80 shown in FIG. 1 after undergoing additional learning by the additional learning unit 174. The meaning of this additional learning will be described later with reference to FIGS. 4 and 5.

FIG. 3 shows the configuration of the BERT learning section 166 shown in FIG. 2 in block diagram form. The teacher data for this learning includes a question, a passage, and a label indicating whether or not the passage includes an answer to the question. The label is, for example, 1 when the passage contains an answer to a question, and 0 otherwise.

Referring to FIG. 3, the BERT learning unit 166 inputs the question and answer candidate passage of the learning data into the question BERT 168 and the answer candidate BERT 170, respectively, and expresses the vector 200 (U) expressing the question and the answer candidate passage. After converting into a vector 202 (V), a cosine similarity calculation unit 204 for calculating the cosine similarity between the vector U and the vector V and the output of the cosine similarity calculation unit 204 ([-1, 1] range) to the range [0, 1], and based on the output of the normalization processing unit 206 and the label 210 (0 or 1) of the learning data, the two match. and a parameter updating unit 208 for updating each parameter of the question BERT 168 and the answer candidate BERT 170 in the direction using the error backpropagation method. The normalization by the normalization processing unit 206 can be performed using, for example, {cos (U, V)+1}/2. Note that in this embodiment, the mean square error is used as a loss function for updating in the parameter updating unit 208.

Query BERT 168 includes a BERT 220 and a pooling layer 222 for outputting a vector 200 by performing average pooling on each element of the final layer of BERT 220.

The answer candidate BERT 170 has exactly the same configuration as the question BERT 168. That is, the answer candidate BERT 170 includes a BERT 230 having the same configuration as the BERT 220, and a pooling layer 232 for outputting the vector 202 by performing average pooling on each element of the final layer of the BERT 230. At this stage, the parameter configurations of BERT 220 and BERT 230 are all common, and updates of their values are also reflected in each other.

As described above, the question BERT 168 and the answer candidate BERT 170 have the same configuration, and are updated by the parameter updating unit 208 so that the parameter values are also common. The BERT learning unit 166 repeats such learning for all learning data until a predetermined end condition is satisfied. As a result, if the answer candidate passage given to answer candidate BERT 170 includes an answer to the question given to question BERT 168, vector 200 and vector 202 are similar vectors, otherwise they are different vectors.

However, when the above learning is performed, the following problems may occur. Referring to FIG. 4, assume that a vector 254 is obtained as a result of transformation 256 by question BERT 168 for a certain question. On the other hand, it is assumed that the vector 260 is the result of converting the passage containing the correct answer to this question using the answer candidate BERT 170. It is also assumed that the vector 260 belongs to the cluster 250 and its centroid is the vector 270.

On the other hand, it is assumed that a cluster 252 different from the cluster 250 exists, and its centroid is the vector 272. Although FIG. 4 assumes that the vector is two-dimensional, it should be noted that the number of dimensions is actually much higher.

In this situation, the cosine similarity between vector 254 and vector 260 is greater than the cosine similarity between vector 254 and any vector included in cluster 252. However, considering the cosine similarity between vector 254 and vectors 270 and 272 of each centroid, the cosine similarity between vector 254 and vector 270 is 264 (cos ₁ ), and the cosine similarity between vector 254 and vector 272 is The relationship cos ₁ <cos ₂ holds true with the cosine similarity 262 (cos2). In such a situation, since the answer candidate search unit 92 shown in FIG. 1 selects clusters based on the cosine similarity between the question vector 82 and the centroid of each cluster, it may not be possible to perform correct processing. Therefore, the additional learning section 174 shown in FIG. 2 performs the following additional learning.

Referring to FIG. 5, in the additional learning, for example, the cosine similarity between the vector generated by the transformation by the question BERT 168 and the vector 270 may be made larger than the cosine similarity between the vector 272 and the vector 272. Therefore, by additional learning of the question BERT 168, the parameters of the question BERT 168 are updated so that a transformation 258, in which the obtained vector 280 becomes closer to the vector 270, is realized instead of the transformation 256 of the question BERT 168. That is, the vector 254 may be moved by an amount corresponding to the vector 282 so that it comes to the position of the vector 280. If the cosine similarity 286 between vectors 280 and 270 is expressed by cos' ₁ , and the cosine similarity 284 between vectors 280 and 272 is expressed by cos' ₂ , then cos' ₁ >cos' ₂ . Bye.

Therefore, in this embodiment, the additional learning section 174 has the following configuration. Referring to FIG. 6, the additional learning unit 174 inputs a certain question 60 to the question BERT 168 to output a vector 310, and on the other hand, the centroid vector of the cluster to which the passage corresponding to the question 60 mainly belongs is defined as a vector W. Then, the cosine similarity calculation unit 314 for calculating the cosine similarity between the vector 310 and the vector 312 consisting of the vector W, and the output of the cosine similarity calculation unit 314 (range [-1, 1]) are set to [0 . 318. The cosine similarity calculation unit 314 is the same as the cosine similarity calculation unit 204 shown in FIG. The normalization processing unit 316 is the same as the normalization processing unit 206 shown in FIG. The parameter update unit 318 is basically the same as the parameter update unit 208 shown in FIG. 3, but differs from the parameter update unit 208 in that it updates the parameters of only the BERT 220 of the question BERT 168. The additional learning unit 174 repeatedly executes this update until a predetermined termination condition is satisfied.

Label "1" indicates the correct answer. Such an update allows the parameters of BERT 220 to be updated such that the cosine similarity between each question's vector and the passage's vector containing its correct answer increases. In addition, in this additional learning, it is sufficient that the vector of the question is arranged close to the centroid of the cluster to which the vector of the correct passage belongs (so that the cosine similarity becomes large). Therefore, learning can be performed using only learning data related to correct passages, without using incorrect passages.

2. Operation The question answering device 50 and the learning device 150 having the configurations described above operate as follows. First, the learning device 150 during learning of the question BERT 80 will be explained, and then the operation of the question answering device 50 using the question BERT 80 will be explained.

2-1. Learning of Question BERT 80 Referring to FIG. 2, learning device 150 operates as follows when learning Question BERT 80. First, the question/passage collection unit 160 crawls the Internet 62, collects question sentences, and stores them in the question DB 162. The question/passage collecting unit 160 further collects passages from the Internet 62 that are considered to include an answer to each question stored in the question DB 162. The question/passage collection unit 160 stores the collected passages in the passage DB 164 in association with the corresponding questions. These processes by the question/passage collection unit 160 can be realized by conventional techniques.

Next, the BERT learning unit 166 generates learning data consisting of positive examples and negative examples from each of the questions stored in the question DB 162 and the passages stored in the passage DB 164 for these questions. The BERT learning unit 166 uses these learning data to simultaneously learn the question BERT 168 and the answer candidate BERT 170 using the Siamese BERT network.

More specifically, referring to FIG. 3, the teacher data for this learning includes a question, a passage, and a label indicating whether or not the passage includes an answer to the question. The label is, for example, 1 when the passage contains an answer to a question, and 0 otherwise.

The BERT learning unit 166 inputs the question and answer candidate passage of the learning data into the question BERT 168 and the answer candidate BERT 170, respectively. BERT 220 and Pooling layer 222 of query BERT 168 transform the query into a vector 200(U). The BERT 230 and pooling layer 232 of the answer candidate BERT 170 convert the answer candidate passage into a vector 202(V). The cosine similarity calculation unit 204 calculates the cosine similarity between the vector U and the vector V, and inputs the calculated cosine similarity to the normalization processing unit 206. This value is in the range [-1,1]. The normalization processing unit 206 normalizes the value of the cosine similarity calculation unit 204 to the range [0, 1] and inputs the normalized value to the parameter updating unit 208. Based on the output of the normalization processing unit 206 and the label 210 (0 or 1) of the learning data, the parameter updating unit 208 uses the error backpropagation method to change the parameters of the question BERT 168 and the answer candidate BERT 170 in the direction that the two match. Update. The question BERT 168 and the answer candidate BERT 170 have the same configuration, and are updated by the parameter updating unit 208 so that the parameter values are also common.

The BERT learning unit 166 repeats such learning for all learning data until a predetermined end condition is satisfied. As a result, if the answer candidate passage given to answer candidate BERT 170 includes an answer to the question given to question BERT 168, vector 200 and vector 202 are similar vectors, otherwise they are different vectors.

The passage clustering unit 172 converts all the passages stored in the passage DB 164 into passage vectors using the answer candidate BERT 170 that has been trained in this way. The passage clustering unit 172 further classifies the passage vectors into a predetermined number of clusters using the k-means method. Each of these clusters is assigned an identifier. The passage clustering unit 172 further calculates the centroid vector of each cluster and assigns the cluster identifier to the centroid identifier. For each centroid thus obtained, the passage clustering unit 172 registers the cluster identifier and the vector of the centroid as a pair in the centroid DB 84.

On the other hand, the passage clustering unit 172 registers a record in the answer candidate DB 90 by combining the URL of the passage corresponding to the passage vector belonging to each cluster and the identifier of the cluster to which the passage belongs. The passage clustering unit 172 also adds, to each question record stored in the question DB 162, the cluster identifier of the cluster containing the most answers to that question.

When the registration of the centroid in the centroid DB 84 is completed, the additional learning unit 174 reads out the centroid vector corresponding to each question stored in the question DB 162 from the centroid DB 84. The additional learning unit 174 performs additional learning of the question BERT 80 for each question using the centroid vector read from the centroid DB 84 as correct answer data. Learning of the question BERT 80 ends when a predetermined end condition is met.

More specifically, referring to FIG. 6, the learning data is a question, a centroid vector W of a cluster to which a passage vector of a passage including a correct answer to the question belongs, and a label. In this case, the value of the label is always "1".

The question is given to question BERT168. BERT 220 and Pooling layer 222 of query BERT 168 process the query and output vector 310. Vector 310 is given to the first input of cosine similarity calculating section 314. On the other hand, the vector W is given to the second input of the cosine similarity calculation unit 314. The cosine similarity calculation unit 314 calculates the similarity between the first input vector and the second input vector and provides it to the normalization processing unit 316. The normalization processing unit 316 normalizes the output of the cosine similarity calculation unit 314 to the range [0, 1] and provides it to the parameter updating unit 318. The parameter update unit 318 updates the parameters of the BERT 220 using the error backpropagation method so that this value approaches label (1).

The additional learning unit 174 repeatedly performs such additional learning until a predetermined termination condition is met.

This additional learning reduces the possibility that the situation shown in Figure 4 will occur, and as shown in Figure 5, the centroid with the highest cosine similarity to the question vector obtained by vectorizing the question using the Question BERT80 will be assigned to the correct cluster. The probability that the vector 270 of the centroid 250 will be obtained can be increased.

2-2. Generation of Answers Referring to FIG. 1, the answer candidate search unit 64 having the question BERT 80 after additional learning and the question answering device 50 including the answer candidate search unit 64 operate as follows. Note that the answer candidate DB 90 and centroid DB 84 have already been obtained through learning.

Assume that some question 60 has been input by the conversation partner. Question 60 is given to question BERT80. The question 60 is also given to the answer generation section 68 at the same time.

Question BERT 80 outputs a question vector 82 representing question 60 in response to question 60 being input. The answer candidate cluster specifying unit 86 searches the centroid DB 84 for the centroid vector closest to the question vector 82 and outputs its cluster identifier 88.

The answer candidate search unit 92 searches the answer candidate DB 90, extracts all passages (answer candidates) having this cluster identifier 88, and outputs a group of answer candidate URLs 94. The passage search unit 96 receives the answer candidate URL group 94, downloads a passage from each URL on the Internet 62, and stores it in the passage DB 66.

The answer generation unit 68 selects the most appropriate answer to the question 60 from the passage group stored in the passage DB 66 and outputs it as an answer 70. The selection of answers by the answer generation unit 68 is exactly the same as the conventional method.

3. Effects According to the first embodiment, by using the answer candidate DB 90 and the centroid DB 84 created in advance in an offline state, and by using the question BERT 80 learned in advance in an offline state, answer candidates for the question 60 are created in a passage. Collect to DB66. The number of answer candidates is limited to answer candidates belonging to the cluster selected by the answer candidate cluster specifying unit 86 shown in FIG. When answering a question, there is no need to select an answer by applying a neural network to a large number of answer candidates. Therefore, the required storage capacity and computational resources are much smaller. In particular, as in the example shown in FIG. 1, the answer candidate DB 90 stores not the passage itself but the URL where the passage exists, so the storage capacity is smaller compared to the case where the passage itself is stored, and the processing is faster. It has the effect of being lightweight. Furthermore, downloading of each passage from the Internet by the passage search unit 96 can be processed in parallel, which has the effect of shortening the time required to generate the answer 70 to the question 60.

In addition to the above configuration, the question history (past questions) of the person asking the question 60 is used as reference information to control the answering operation (for example, answering for beginners and answering for experts). It is also possible to configure the answer level (to change the answer level).

Second Second Embodiment 1. Configuration The first embodiment relates to a question answering device that provides answers to questions. However, the present invention can be applied not only to such embodiments but also to so-called dialogue devices. The input that the system receives in the dialog device is not necessarily a question. However, as a method for searching for a response to an input, a method similar to that for answering questions can be used.

A dialog device differs from a question-answer device in that when generating a response to the other party's utterance, it is desirable to make a response that is related to the history of previous dialogs. This second embodiment relates to such a dialogue system.

A. Dialogue Device FIG. 7 shows the configuration of a dialogue device 350 according to a second embodiment of the present invention. This dialog device 350 also uses a question BERT 380 having the same configuration as the question BERT 80 of the first embodiment. Learning of the question BERT 380 will be described later with reference to FIG. 8 and subsequent figures.

Referring to FIG. 7, the dialogue device 350 includes a response candidate search device 360 that searches the Internet 62 for appropriate response candidates as a response to the utterance 362 in response to the other party's utterance 362; a passage DB 363 for storing response candidates stored in the passage DB 363; a response generation unit 364 for selecting an appropriate response to the utterance 362 from among the response candidates stored in the passage DB 363, and generating and outputting a response 366; It includes a response shaping section 368 that formats the output of the response generation section 364 into a form suitable for dialogue and outputs it. The passage DB 363 has the same configuration as that used in the first embodiment. However, since the passage selected by the response candidate search device 360 is different from the first embodiment, the storage contents of the passage DB 363 are also different from the first embodiment. Dialogue has a speaking style that is appropriate for dialogue. Therefore, in this embodiment, the response shaping section 368 formats and outputs the output of the response generation section 364.

The response candidate search device 360 includes a dialogue history management unit 370 for managing the history of dialogue between the dialogue device 350 and the other party, a dialogue history DB 372 that stores the dialogue history managed by the dialogue history management unit 370, and a dialogue history DB 372 that stores a plurality of contents. and a topic model 374 for outputting information indicating a conversation topic including the content word in response to input of the content word. It is assumed that the topic model 374 is, for example, a statistical model and has been trained in advance. For example, the method described in Japanese Patent Application Publication No. 2015-045915 can be used to learn the topic model. It is also possible to construct topic models using neural networks. The output of topic model 374 is, for example, one or more words representing the topic.

The response candidate search device 360 further responds to the utterance 362 and uses the information stored in the dialogue history DB 372 and the topic model 374 to search for a topic for giving the utterance 362 information indicating the topic of the dialogue with the other party. It includes an imparting section 376. Specifically, the topic assigning unit 376 reads the history of the conversation with the other party from the conversation history DB 372 and extracts the content words. The topic providing unit 376 provides these content words to the topic model 374 and receives one or more words representing the topic output from the topic model 374. The topic adding unit 376 further adds one or more words received from the topic adding unit 376 to the end of the utterance 362 and outputs the added word.

The response candidate search device 360 further has a configuration similar to the question BERT 80 in the first embodiment for outputting a question vector 382 similar to the question vector in the first embodiment upon receiving the output of the topic assigning unit 376. In response to the question BERT 380, the centroid DB 378 obtained in the same manner as the centroid DB 84 in the first embodiment, and the question vector 382, a predetermined number of centroid vectors having the largest cosine similarity with the question vector 382 are generated. It also includes a response candidate cluster identification unit 386 for searching in the centroid DB 378 and outputting cluster identifiers 388 thereof. The configuration of the centroid DB 378 itself is the same as the centroid DB 84 of the first embodiment. However, in this second embodiment, learning of the centroid DB 378 is performed by a method slightly different from that in the first embodiment. Therefore, the centroid DB378 is described here as being different from the centroid DB84.

The response candidate search device 360 further searches the response candidate DB 390 in response to the response candidate DB 390 having the same configuration as the response candidate DB 90 in the first embodiment and the cluster identifier 388 from the response candidate cluster specifying unit 386, and searches the response candidate DB 390 to determine the cluster. It includes a response candidate search unit 392 having the same configuration as that of the first embodiment, for reading out all records of response candidates having cluster identifiers that match the identifier 388 and outputting them as a response candidate URL group 394.

The response candidate search device 360 further receives the response candidate URL group 394, searches the Internet 62, downloads response candidate passages from the URLs stored in the records of each response candidate, and stores them in the passage DB 363. , this also includes a passage search section 396 similar to the passage search section 96 of the first embodiment.

B. Learning Device Referring to FIG. 8, a learning device 400 for learning the question BERT 380 of the response candidate search device 360 crawls the Internet 62 and analyzes the question by searching several sentences before and after it (hereinafter, like this). A question/passage collection unit 410 downloads a question and a collection of sentences before and after the question (called a question passage), and further downloads an appropriate passage as a response to each question, and the question passage downloaded by the question/passage collection unit 410. and a passage DB 164 having a configuration similar to that shown in FIG. 2, for storing passages downloaded by the question/passage collection unit 410 in association with questions.

The learning device 400 further uses learning data created using the topic model 374, each question passage stored in the question passage DB 412, the passage stored in the passage DB 164, and the words related to the topic obtained from the topic model 374. , a BERT learning unit 414 for performing learning of the question BERT 380 and response candidate BERT 418 using the Siamese BERT network. Similar to the question BERT 168 and answer candidate BERT 170 in the first embodiment, the question BERT 380 and response candidate BERT 418 have the same configuration, and updating of one parameter during learning is reflected in updating of the other parameter. In reality, the configurations of the question BERT 380 and the response candidate BERT 418 are the same as the question BERT 168 and the answer candidate BERT 170 of the first embodiment, and their internal parameters differ as a result of learning.

The learning device 400 further uses the response candidate BERT 418 that has been trained by the BERT learning unit 414 to cluster the passages stored in the passage DB 164 into a predetermined number of clusters, and assigns a cluster identifier to each cluster. Clustering unit 172; a response candidate DB 390 that stores each response candidate clustered by the passage clustering unit 172 with a cluster identifier of the cluster to which it belongs; and a response candidate DB 390 that stores each response candidate clustered by the passage clustering unit 172; It includes a centroid DB 378 for storing, for each cluster, a record including a centroid vector and a cluster identifier of the cluster it represents. Note that for each question stored in the question passage DB 412, the passage clustering unit 172 assigns to each question record in the question passage DB 412 the cluster identifier of the cluster to which the largest number of passages related to the question belong.

The learning device 400 further includes an additional learning unit 422 for performing additional learning of the question BERT 380 using the questions stored in the question passage DB 412 and information regarding the centroid of each cluster stored in the centroid DB 378. . By performing additional learning on the question BERT 380 by the additional learning unit 422, the question BERT 380 shown in FIG. 7 is obtained.

FIG. 9 shows the functional configuration of the BERT learning section 414 in block diagram form. Referring to FIG. 9, the BERT learning unit 414 receives each question and the question passage including the question stored in the question passage DB 412 shown in FIG. A topic adding unit 450 for adding one or more words indicating a topic to a question and a topic model 374 are used for each response candidate passage stored in the passage DB 164 to indicate the topic of the response candidate passage. and a topic adding unit 452 for adding one or more words to the response candidate passage. Learning data for the question BERT 380 is generated by the output of the topic assignment unit 450, the output of the topic assignment unit 452, and a label that is 1 if the response candidate passage provides an answer to the question, and 0 otherwise. Ru.

The query BERT 416 includes a BERT 480 and a pooling layer 482 for performing average pooling on each element of the final layer of the BERT 480 and outputting a vector 454(U). Similarly, the response candidate BERT 418 includes a BERT 490 and a pooling layer 492 for performing average pooling on each element of the final layer of the BERT 484 and outputting a vector 456(V).

The BERT learning unit 414 further provides the output of the topic assignment unit 450 to the question BERT 416 and the output of the topic assignment unit 452 to the response candidate BERT 418, and calculates the difference between the vector 454 output from the question BERT 416 and the vector 456 output from the response candidate BERT 418. A cosine similarity calculation unit 204 for calculating cosine similarity, and a cosine similarity calculation unit 204 for normalizing the value output from the cosine similarity calculation unit 204 (range of [-1, 1]) to the range of [0, 1]. It includes a normalization processing unit 206 and a parameter updating unit 208 for updating each parameter of the BERT 480 and 490 in a direction in which the cosine similarity after normalization obtained by the normalization processing unit 206 becomes equal to the label 210. .

The BERT learning unit 414 has a function of repeatedly executing such update processing using the above-mentioned learning data until a predetermined termination condition is satisfied.

The configuration of the additional learning section 422 is substantially the same as the additional learning section 174 shown in FIG. 6 in the first embodiment. However, this example differs from the additional learning section 174 in that the topic of the question passage is attached to the question.

2. Operation Also in this second embodiment, first, the learning device 400 shown in FIG. 8 learns the question BERT 416 and generates the centroid DB 378 and response candidate DB 390. Further, the additional learning unit 422 performs additional learning on the question BERT 416. As a result, the question BERT 380 is obtained, and the processing of the dialog device 350 shown in FIG. 7 becomes possible.

2-1. Learning of Question BERT380 The learning flow of Question BERT380 is roughly the same as the learning flow of Question BERT80 in the first embodiment.

Referring to FIG. 8, learning device 150 operates as follows when learning question BERT380. First, the question/passage collection unit 410 crawls the Internet 62, collects question sentences and question passages including the parts before and after the question sentences, and stores them in the question passage DB 412. The question/passage collection unit 410 further collects from the Internet 62, for each question text stored in the question passage DB 412, a plurality of passages that are considered to include a response to that question. The question/passage collection unit 410 stores the collected passages in the passage DB 164 in association with the questions stored in the question passage DB 412. These processes by the question/passage collection unit 410 can be realized using conventional techniques.

Next, the BERT learning unit 414 uses the topic model 374 to correct Generate learning data consisting of examples and negative examples. The BERT learning unit 414 uses these learning data to simultaneously learn the question BERT 416 and the response candidate BERT 418 using the Siamese BERT network.

The passage clustering unit 172 converts all the passages stored in the passage DB 164 into passage vectors using the response candidate BERT 418 that has been trained in this way. The passage clustering unit 172 further classifies the passage vectors into a predetermined number of clusters using the k-means method. Each of these clusters is assigned an identifier. The passage clustering unit 172 further calculates the centroid vector of each cluster and assigns the cluster identifier to the centroid identifier. For each centroid thus obtained, the passage clustering unit 172 registers the cluster identifier and the vector of the centroid as a pair in the centroid DB 378.

On the other hand, the passage clustering unit 172 registers in the response candidate DB 390 a record that is a set of the URL of the passage corresponding to the passage vector belonging to each cluster and the identifier of the cluster to which it belongs. The passage clustering unit 172 also adds, to each question record stored in the question passage DB 412, the cluster identifier of the cluster containing the most responses to the question.

When the registration of the centroid in the centroid DB 378 is completed, the additional learning unit 422 reads out the centroid vector corresponding to each question stored in the question passage DB 412 from the centroid DB 378. The additional learning unit 422 performs additional learning of the question BERT 380 using the centroid vector read from the centroid DB 378 as correct answer data for each question. Learning of the question BERT 380 ends when a predetermined end condition is met.

2-2. Generation of Responses Referring to FIG. 7, response candidate search device 360 having question BERT 380 after additional learning and interaction device 350 including response candidate search device 360 operate as follows. Note that the response candidate DB 390 and centroid DB 378 have already been obtained through the above-described learning. Further, as the topic model 374, the same one used for learning is used.

Assume that some kind of utterance 362 has been input from the conversation partner. The utterance 362 is provided to a topic adding section 376, a dialogue history management section 370, and a response generation section 364. The dialogue history management unit 370 stores the history of past dialogues received over a certain period of time in the dialogue history DB 372 for each dialogue partner. On the other hand, the topic assigning unit 376 searches the dialogue partner's past utterance history in the dialogue history DB 372 and extracts content words from those utterances. The topic assigning unit 376 further inputs these content words into the topic model 374 and receives one or more words representing each topic as an output of the topic model 374. The topic adding unit 376 adds these one or more words to the end of the utterance 362 and inputs the result to the question BERT 380.

Question BERT 380 outputs a question vector in response to topic-added utterance 362. This question vector is given to the response candidate cluster identifying section 386.

In response to the question vector 382, the response candidate cluster specifying unit 386 selects a centroid having the largest cosine similarity between the vector and the question vector 382 among the centroids stored in the centroid DB 378. The selected number of cluster identifiers are given to the response candidate search unit 392 as cluster identifiers 88.

The response candidate search unit 392 searches the response candidate DB 390, extracts all records of passages (response candidates) having a cluster identifier equal to one of the cluster identifiers 388, and outputs the URLs together as a response candidate URL group 394. . The passage search unit 396 receives the response candidate URL group 394, accesses the Internet 62 using each URL included in the response candidate URL group 394, downloads the response candidate passage, and stores it in the passage DB 363. .

The response generation unit 364 selects the most appropriate response to the utterance 362 from among the passage group stored in the passage D3B 66, and provides it to the response shaping unit 368. The response formatter 368 formats the input into a format suitable for dialogue and outputs it as a response 366. The response selection and generation method by the response generation unit 364 can be the same as the conventional method.

3. Effects According to the second embodiment, the passage to be processed to generate a response to an input utterance is a cluster corresponding to one or more railroad tracks selected by the response candidate cluster identification unit 386 shown in FIG. is limited to those corresponding to vectors belonging to . Therefore, the optimal response does not need to be selected from a plurality of response candidates generated based on a plurality of answers obtained from a question answering system as in the past, so when the response generation unit 364 generates a response, The amount of processing required is significantly reduced compared to conventional methods. Additionally, the computational resources required for processing are also significantly reduced. Also, unlike the first embodiment, the topic of the utterance is used as information when searching for response candidate passages. Therefore, the responses obtained will be appropriate to the topic of the conversation. As a result, it is possible to provide an interactive device that can operate at high speed and with sufficient accuracy using fewer resources than before.

Third Realization by Computer FIG. 10 is an external view of a computer system that operates as, for example, the question answering device 50 shown in FIG. 1. FIG. 11 is a hardware block diagram of the computer system shown in FIG. 10. The learning device 150 shown in FIG. 2, the dialogue device 350 shown in FIG. 7, and the learning device 400 shown in FIG. 8 can each be realized by a computer system having substantially the same configuration as the question answering device 50. Here, only the configuration of the computer system that operates as the question answering device 50 will be described, and details of the configuration of the computer system that implements other devices will not be described.

Referring to FIG. 10, this computer system 950 includes a computer 970 having a DVD (Digital Versatile Disc) drive 1002, a keyboard 974, a mouse 976, and a monitor for interacting with the user, all connected to the computer 970. 972. Of course, these are examples of configurations for when user interaction is required, and any general hardware and software (e.g. touch panel, voice input, general pointing device) that can be used for user interaction can be used. Also available.

Referring to FIG. 11, in addition to a DVD drive 1002, a computer 970 includes a CPU (Central Processing Unit) 990, a GPU (Graphics Processing Unit) 992, and a bus 101 connected to the CPU 990, GPU 992, and DVD drive 1002. 0 and including. The computer 970 further includes a ROM (Read-Only Memory) 996 that is connected to the bus 1010 and stores boot-up programs for the computer 970, etc. A RAM (Random Access Memory) 998 for storing , and an SSD (Solid State Drive) 1000 which is a nonvolatile memory connected to a bus 1010 are included. The SSD 1000 is for storing programs executed by the CPU 990 and GPU 992, data used by the programs executed by the CPU 990 and GPU 992, and the like. The computer 970 further includes a network I/F (Interface) 1008 that provides connection to a network 986 (response generation unit 364 shown in FIG. 7) that enables communication with other terminals, and a USB (Universal Serial Bus) memory 984. is removable and includes a USB memory 984 and a USB port 1006 that provides communication with various parts within the computer 970.

The computer 970 is further connected to a microphone 982, a speaker 980, and a bus 1010, reads audio signals, video signals, and text data generated by the CPU 990 and stored in the RAM 998 or the SSD 1000 according to instructions from the CPU 990, and performs analog conversion and amplification processing. It includes an audio I/F 1004 having a function of driving a speaker 980, digitizing an analog audio signal from a microphone 982, and storing it in an arbitrary address specified by the CPU 990 in the RAM 998 or the SSD 1000.

In the above embodiment, the programs for realizing each function of the question answering device 50 shown in FIG. 1, the learning device 150 shown in FIG. 2, the dialog device 350 shown in FIG. 7, and the learning device 400 shown in FIG. It is also stored in, for example, the SSD 1000, RAM 998, DVD 978, or USB memory 984 shown in FIG. 11, or a storage medium of an external device (not shown) connected via the network I/F 1008 and network 986. Typically, these data and parameters are written into the SSD 1000 from the outside, for example, and loaded into the RAM 998 when the computer 970 is executed.

This computer system is operated to realize the functions of the question answering device 50 shown in FIG. 1, the learning device 150 shown in FIG. 2, the dialogue device 350 shown in FIG. 7, and the learning device 400 shown in FIG. 8, and their respective components. A computer program for doing this is stored on the DVD 978 installed in the DVD drive 1002 and transferred from the DVD drive 1002 to the SSD 1000. Alternatively, these programs are stored in the USB memory 984, the USB memory 984 is attached to the USB port 1006, and the programs are transferred to the SSD 1000. Alternatively, this program may be transmitted to computer 970 via network 986 and stored on SSD 1000.

The program is loaded into RAM 998 during execution. Of course, a source program may be input using the keyboard 974, monitor 972, and mouse 976, and the compiled object program may be stored in the SSD 1000. If the program is written in a script language, the script input using the keyboard 974 or the like may be stored in the SSD 1000. In the case of a program that operates on a virtual machine, it is necessary to install the program that functions as a virtual machine on the computer 970 in advance. Neural networks are used for speech recognition, speech synthesis, etc. In the question answering device 50 and the dialogue device 350, a trained neural network may be used, or the question answering device 50 and the dialogue device 350 may be used as the learning devices 150 and 400, respectively, to perform neural network learning. Good too.

The CPU 990 reads the program from the RAM 998 according to the address indicated by an internal register called a program counter (not shown), interprets the instruction, and stores the data necessary for executing the instruction in the RAM 998 and the SSD 1000 according to the address specified by the instruction. Or read it from other devices and execute the process specified by the command. The CPU 990 stores the data of the execution result at an address specified by the program, such as the RAM 998, the SSD 1000, or a register within the CPU 990. Some addresses are output externally from the computer. At this time, the value of the program counter is also updated by the program. Computer programs may be loaded directly into RAM 998 from DVD 978, from USB memory 984, or via network 986. Note that in the program executed by the CPU 990, some tasks (mainly numerical calculations) are dispatched to the GPU 992 according to instructions included in the program or according to an analysis result when the CPU 990 executes the instructions.

A program for realizing the functions of each unit according to the above-described embodiments by the computer 970 includes a plurality of instructions written and arranged to cause the computer 970 to operate to realize those functions. Some of the basic functions required to execute this instruction are provided by the operating system (OS) running on the computer 970, third party programs, modules of various toolkits installed on the computer 970, or the program execution environment. In some cases, it may be provided. Therefore, this program does not necessarily include all the functions necessary to implement the system and method of this embodiment. This program is constructed by statically linking appropriate functions or modules in a controlled manner at compile time or by calling them dynamically at run time in a controlled manner to achieve the desired results. It is sufficient to include only the instructions for executing the operations of each device and its constituent elements. The manner in which computer 970 operates for this purpose is well known. Therefore, the description of how computer 970 operates will not be repeated here.

Note that the GPU 992 can perform parallel processing, and can execute a large amount of calculations associated with machine learning simultaneously in parallel or in a pipeline manner. For example, parallel computing elements found in a program when the program is compiled, or parallel computing elements discovered when the program is executed are dispatched from the CPU 990 to the GPU 992 and executed, and the results are sent directly or to the RAM 998. is returned to the CPU 990 via a predetermined address, and is substituted into a predetermined variable in the program.

Fourth Modified Example In the above embodiment, only a predetermined termination condition is mentioned as the termination condition for updating BERT. The termination condition in this case is, for example, that the BERT parameter update has been completed after all learning data have been used a predetermined number of times. Another condition may be, for example, when the gradient of each parameter becomes approximately equal to zero during parameter updating. In addition, various other termination conditions can be considered.

In the above embodiment, it is assumed that BERT large is used as BERT. However, the invention is not limited to BERT large. A BERT base may also be used. Also, similar to BERT, a model that can perform sentence-based conversion by arranging transformer encoders or similar elements may be used.

Furthermore, in the above embodiment, average pooling is performed in order to vectorize the output of BERT in the query BERT 80 shown in FIG. 1 and the like. However, the invention is not limited to such embodiments. Maximum pooling may be used instead of average pooling.

In the above embodiment, the mean square error is used as the loss function for BERT learning. However, the invention is not limited to such embodiments. Instead of cosine similarity, softmax may be used as the measure, and cross-entropy error may be used as the loss function.

Further, in the above embodiment, the k-means method is used for clustering passage vectors during learning. However, the invention is not limited to such embodiments. Clustering based on Related Minimum Variance criteria, clustering based on scatter plot criteria, etc. may also be used.

Further, in the above embodiment, the response candidate cluster identifying unit 386 is given the question vector of the utterance to which the topic has been added, and obtains an appropriate centroid cluster identifier from the centroids in the centroid DB 378. However, the present invention is not limited to this, and the response candidate cluster identification unit 386 may be provided with a question vector based on utterances and a past dialogue history expressed as a vector, determine a centroid from similarity (or distance), and obtain its cluster identifier. You can also do this.

Furthermore, in the embodiment described above, the URL of the response candidate is stored in the response candidate DB 90 shown in FIG. 1 and the response candidate DB 390 shown in FIG. However, the invention is not limited to such embodiments. The response candidate DB 90 and the response candidate DB 390 may store the passage itself instead of the URL of the response candidate. In that case, what the answer candidate search unit 92 and the response candidate search unit 392 output is not the URL group of the response candidates but the response candidate passage itself. Therefore, the passage search section 96 shown in FIG. 1 and the passage search section 396 shown in FIG. 7 are no longer necessary. Since there is no need to download passages from the Internet 62 when answering or responding to a question, the amount of processing is less and the response is likely to be faster than in the first and second embodiments.

Further, in the above embodiment, the passage DB 164 shown in FIGS. 2 and 8 is unnecessary and can be discarded after BERT learning is completed. However, if the passage DB 164 has enough storage capacity to store a large number of passages, the passage DB 164 may be left alone. In this case, the answer candidate DB 90 and the response candidate DB 390 may store information necessary to search for the relevant passage from the passage DB 164 instead of the response candidate passage. The information necessary for the search is, for example, the identifier of the record containing the passage, or the URL where the passage existed. In particular, if you store the URL, if the passage exists in the passage DB 164, you can retrieve the passage from the passage DB 164, and if the passage does not exist in the passage DB 164, access the URL and download the passage. can. In this case, the answer candidate search unit 92 and the response candidate search unit 392 output a group of response candidate URLs, but normally the passage can be read from the locally existing passage DB 164 without accessing the Internet. . Therefore, the response is likely to be faster than in the first and second embodiments. Further, even if some passages are deleted from the passage DB 164, since the URL is known, the passages can be downloaded from the Internet.

In the second embodiment described above, the Internet 62 is searched for passages that are thought to include response candidates to the input to the dialogue device 350, and the response generation unit 364 generates a response using these passages. However, the invention is not limited to such embodiments. For example, if an appropriate response candidate can be prepared in advance as a response to the utterance 362, such a response candidate can be stored in the response candidate DB 390 instead of a passage. Various utterances 362 are prepared, and once appropriate response candidates are determined, they are stored in the response candidate DB 390. If a large number of these response candidates are obtained, the dialog device 350 can be constructed using them using the method described in the second embodiment. In this case, rather than generating a response from a passage that is thought to include response candidates, appropriate responses to the utterance are collected in advance and a response is generated from them. Therefore, the response becomes more appropriate than that described in the second embodiment.

The following methods can be considered as methods for collecting appropriate responses to utterances in advance. First, a large number of nouns (for example, 1 million words) that may constitute an utterance are selected. Multiple questions are generated from each of these nouns using existing techniques. These questions are input into the question answering device 50 of the first embodiment. Based on the answers from the question answering device 50, a response suitable for the question including the original noun is created, and the question and response are stored in the question passage DB 412 and the passage DB 164 in FIG. After that, the question BERT 380 shown in FIGS. 7 and 8 may be trained using the same method as in the second embodiment. Dialogue device 350 can be constructed using response candidates suitable as responses to utterances, rather than passages that are considered to include response candidates to utterances. In such a modified example, response candidates considered to be most appropriate for the utterance are generated and selected in advance, so that heavy computational work in the dialogue device can be completed in advance. Therefore, it is possible to speed up the inference processing and reduce the weight of computational resources.

Note that in the second embodiment shown in FIG. 7 and below, the topic model 374 is used in both training and inference. However, it is not essential to use the topic model 374 when assigning topics in dialogue. For example, a word that satisfies some condition or the dialogue history itself may be extracted from the dialogue history and added to the utterance 362 by the topic adding unit 376. Rather than specifying a topic using the conversation history, the topic may be determined based on settings made in advance. For example, the speaker may explicitly specify the topic. Further, the topic adding section 376 itself may be omitted.

The embodiment disclosed this time is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim, with reference to the description of the detailed description of the invention, and all changes within the scope and meaning equivalent to the words described therein are defined. include.

50 Question answering device 64, 92 Answer candidate search unit 66, 164, 363 Passage DB
68 Answer generation unit 80, 168, 380, 416 Question BERT
82, 382 Question vector 84, 378 Centroid DB
86 Answer candidate cluster identification unit 88, 388 Cluster identifier 90 Answer candidate DB
94 Answer candidate URL group 96, 396 Passage search unit 150, 400 Learning device 160, 410 Question/passage collection unit 162 Question DB
166, 414 BERT learning section 170 Answer candidate BERT
172 Passage clustering unit 174, 422 Additional learning unit 204, 314 Cosine similarity calculation unit 206, 316 Normalization processing unit 208, 318 Parameter update unit 210 Label 220, 230, 480, 484, 490 BERT
222, 232, 482, 492 Pooling layer 250, 252 Cluster 350 Dialogue device 360 Response candidate search device 362 Utterance 364 Response generation section 368 Response shaping section 370 Dialogue history management section 372 Dialogue history DB
374 Topic models 376, 450, 452 Topic assignment unit 386 Response candidate cluster identification unit 390 Response candidate DB
392 Response candidate search unit 394 Response candidate URL group 412 Question passage DB
418 Response candidate BERT

Claims (6)

a first neural network provided with a first input and outputting a vector representation of the first input;
a second neural network provided with a second input and outputting a vector representation of the second input;
At least, when the vector representation of the first input and the vector representation of the second input have the predetermined relationship using learning data of the first and second inputs that have a predetermined relationship, learning the first and second neural networks to be located close to each other in vector space;
The vector representation that is the output of the second neural network that has been trained is clustered based on the position on the vector space, and based on the vector representation of the first input, it is possible to search and extract the cluster in advance. further including the constructed database,
An inference device that infers an output having the predetermined relationship based on information on the clusters searched and extracted from the database based on a vector representation of the input by the first neural network with respect to an input to the device. 2. The inference device according to claim 1, wherein the database is searched and extracted using a centroid of a vector representation of an output included in each cluster. The first neural network has undergone additional learning based on the first input and a vector representation related to a cluster to which the second input, which has the predetermined relationship with the first input, belongs. The inference device according to claim 1 or claim 2, characterized in that: A question answering device including an inference device according to any one of claims 1 to 3, wherein the predetermined relationship includes a question and an answer to the question. 4. An interaction device including an inference device according to claim 1, wherein the predetermined relationship includes an utterance and a response to the utterance. a first neural network provided with a first input and outputting a vector representation of the first input; and a second neural network provided with a second input and outputting a vector representation of the second input. and a step of preparing the
At least, when the vector representation of the first input and the vector representation of the second input have the predetermined relationship using learning data of the first and second inputs that have a predetermined relationship, training the first and second neural networks to be located close to each other in vector space;
The vector representation that is the output of the second neural network that has been trained is clustered based on the position on the vector space, and based on the vector representation of the first input, it is possible to search and extract the cluster in advance. a step of building a database;
and inferring an output having the predetermined relationship based on the cluster information searched and extracted from the database based on the vector representation of the input by the first neural network. .

Images