JP6871809B2 - Information processing equipment, information processing methods, and programs - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

In recent years, there has been known a technique of outputting an answer to an input question by using machine learning (for example, a deep learning method) (see, for example, Non-Patent Document 1). In an information processing device using such a conventional technique, for example, an answer presumed to be appropriate is selected and output from known answers prepared in advance such as answers accumulated in the past.

Tan M, Xiang B, Zhou B, “LSTM-BASED DEEP LEARNING MODELS FOR NONFACTOID ANSWER SELECTION” 1511.04108v1, 12 Nov 2015

However, in the above-mentioned conventional information processing device, for example, when the question is a non-factoid type question that requires an answer based on the explanation of the reason or the event, the answer becomes a complicated long sentence, but the question is answered in advance. It is difficult to generate a new answer because the prepared known answer is output. Therefore, in the conventional information processing device described above, an unnatural answer with a sense of incongruity in the text may be output in response to the question.

The present invention has been made to solve the above problems, and an object of the present invention is to provide an information processing device, an information processing method, and a program capable of generating a natural written answer to a question with less discomfort. To provide.

In order to solve the above problem, one aspect of the present invention includes a question acquisition unit that acquires an input input question sentence, and a plurality of question sentences and answer sentences that are divided by a predetermined sentence path. A response sentence to the input question sentence acquired by the question acquisition unit is generated based on the learning result machine-learned based on the learning information having a plurality of pairs of known partial answer sentences corresponding to each partial item. The learning result is provided with an answer generation unit, and the question sentence is sequentially encoded word by word based on the order of the words to generate a context vector, and based on the generated context vector, for each of the plurality of sub-items. Input the set information including the encoder / decoder model for decoding and learning the known partial answer sentence, the partial answer sentence for each of the plurality of partial items decoded based on the encoder / decoder model, and the question sentence. As information, the sequence of the words is based on the question feature vector group in which the feature vector obtained by converting the question sentence for each word is generated based on the sequence order of the words in both the forward and reverse directions of the time series. The question intermediate vector generated by learning the above bidirectionally and the answer intermediate vector corresponding to each of the plurality of sub-items, and the feature vector obtained by converting the partial answer sentence for each word is the bidirectional. Based on the answer feature vector group generated based on the order of the words, the answer intermediate vector generated by learning the sequence of the words in both directions is learned based on the combination of the plurality of sub-items. It is an information processing device characterized in that it is optimized and learned by a loss function calculated by combining a sentence unit learning model.

Further, in one aspect of the present invention, in the above information processing apparatus, the sentence unit learning model uses the intermediate learning result of the encoder / decoder model and the answer sentence generated based on the context vector as the partial answer sentence. It is characterized in that the set information including the partial answer sentence and the question sentence is learned as the input information.

Further, in one aspect of the present invention, in the above information processing apparatus, the encoder / decoder model decodes the known partial answer sentence based on the topic information related to each word in the known partial answer sentence. It is characterized by learning.

Further, in one aspect of the present invention, in the above-mentioned information processing apparatus, the known answer sentence includes a correct answer sentence and an incorrect answer sentence for the question sentence, and the answer generation unit includes the question sentence. , The answer sentence is generated based on the learning result machine-learned based on the learning information having a plurality of pairs of the correct answer sentence and the incorrect answer sentence corresponding to each of the plurality of partial items, and the encoder. The decoder model decodes and learns the correct sentence and the incorrect answer sentence for each of the plurality of partial items based on the context vector, and the sentence unit learning model includes the question intermediate vector and the plurality of parts. The correct answer intermediate vector corresponding to each item, and the sequence of the words based on the correct answer feature vector group generated based on the sequence order of the words in both directions, which is a feature vector obtained by converting the correct sentence for each word. The correct intermediate vector generated by learning in both directions and the incorrect intermediate vector corresponding to each of the plurality of sub-items, and the feature vector obtained by converting the incorrect sentence word by word are used in both directions. Based on the incorrect answer feature vector group generated based on the order of the words, the incorrect intermediate vector generated by learning the sequence of the words in both directions, and the combination of the plurality of partial items. It is characterized by learning.

Further, in one aspect of the present invention, in the above information processing apparatus, the learning result is based on the correct answer intermediate vector and the incorrect answer intermediate vector corresponding to the first partial item among the plurality of partial items. , The correct answer feature vector group and the incorrect answer feature vector group corresponding to the second sub-item different from the first sub-item are updated and learned.

Further, one aspect of the present invention is characterized in that the information processing apparatus includes a learning processing unit that performs machine learning based on the learning information and generates the learning result.

Further, in one aspect of the present invention, the question acquisition unit divides the question acquisition step of acquiring the input input question sentence, and the answer generation unit divides the question sentence and the answer sentence by a predetermined sentence path. Answers to the input question sentence acquired by the question acquisition step based on the learning result machine-learned based on the learning information having a plurality of sets with known partial answer sentences corresponding to each of the plurality of partial items to be performed. The learning result includes the answer generation step of generating a sentence, and the question sentence is sequentially encoded word by word based on the order of the words to generate a context vector, and the plurality of the learning results are generated based on the generated context vector. Includes an encoder / decoder model that decodes and learns the known partial answer sentence for each partial item, a partial answer sentence for each of the plurality of partial items decoded based on the encoder / decoder model, and the question sentence. Using the set information as input information, the feature vector obtained by converting the question sentence for each word is generated based on the sequence order of the words in both the forward and reverse directions of the time series, based on the question feature vector group. A question intermediate vector generated by learning the sequence of words in both directions and an answer intermediate vector corresponding to each of the plurality of partial items, and a feature vector obtained by converting the partial answer sentence for each word. Based on the answer feature vector group generated based on the bidirectional order of the words, the combination of the plurality of sub-items of the answer intermediate vector generated by learning the bidirectional sequence of the words. It is an information processing method characterized in that it is optimized and learned by a loss function calculated by combining a sentence-based learning model that learns based on the sentence.

Further, one aspect of the present invention is a question acquisition step of acquiring an input question sentence input to a computer, and a plurality of partial items divided by a predetermined sentence path in the question sentence and the answer sentence, respectively. An answer generation step that generates an answer sentence to the input question sentence acquired by the question acquisition step based on a learning result machine-learned based on learning information having a plurality of pairs with a known partial answer sentence corresponding to. The learning result is a program for executing the above, and the learning result encodes the question sentence word by word sequentially based on the order of the words to generate a context vector, and based on the generated context vector, the plurality of words are executed. A set including an encoder / decoder model that decodes and learns the known partial answer sentence for each partial item, a partial answer sentence for each of the plurality of partial items decoded based on the encoder / decoder model, and the question sentence. Based on the question feature vector group generated based on the sequence order of the words in both the forward and reverse directions of the time series, the feature vector obtained by converting the question sentence for each word with the information as input information is described. The question intermediate vector generated by learning the sequence of words in both directions and the answer intermediate vector corresponding to each of the plurality of partial items, and the feature vector obtained by converting the partial answer sentence for each word are described above. Based on the answer feature vector group generated based on the bidirectional order of the words, and based on the combination of the plurality of sub-items of the answer intermediate vector generated by learning the bidirectional sequence of the words. It is a program characterized in that it is optimized and learned by a loss function calculated by combining a sentence-based learning model for learning.

According to the present invention, it is possible to generate a natural written answer to a question with less discomfort.

It is a functional block diagram which shows an example of the information processing system by this Embodiment. It is a figure explaining an example of the learning model in this embodiment. It is a figure explaining an example of the decoder model in this embodiment. It is a figure explaining an example of the learning processing part and learning processing in this embodiment. It is a figure explaining an example of the encoder-decoder model in this embodiment. It is a flowchart which shows an example of the learning process in this embodiment. It is a flowchart which shows an example of the process which generates the answer sentence from the question sentence of the information processing apparatus in this embodiment. It is a figure which shows the comparison between the answer generation method in this embodiment, and the prior art. It is a figure explaining an example of the answer sentence generated by the information processing apparatus in this embodiment.

Hereinafter, the information processing apparatus according to the embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic block diagram showing an example of the information processing system 100 according to the present embodiment.
As shown in FIG. 1, the information processing system 100 includes an information processing device 1 and a terminal device 2. The information processing device 1 and the terminal device 2 are connected via a network NW1.
The information processing system 100 provides an information service such as a Q & A service for displaying posted questions and answers on a terminal device 2 connected to the information processing device 1 and sharing information between users.

The terminal device 2 is a client terminal used by a user to use the information service provided by the information processing system 100. In the example shown in FIG. 1, for simplification of the description, one terminal device 2 is connected to the information processing device 1, but a plurality of terminal devices 2 are connected to the information processing device 1. May be connected to.

The information processing device 1 is a server device that provides an information service such as a Q & A service. For example, the information processing device 1 registers a question sentence received from a user via the terminal device 2 in a Q & A service so that it can be viewed, and registers an answer sentence received from another user via the terminal device 2. To make it viewable. Further, the information processing device 1 uses machine learning to generate an answer sentence to the registered question sentence, and registers the answer sentence in the Q & A service so that the answer sentence can be viewed. Further, the information processing device 1 includes a NW (network) communication unit 11, a storage unit 12, and a control unit 13.

The NW communication unit 11 connects to the network NW1 using, for example, the Internet, and communicates various information via the network NW1. The NW communication unit 11 connects to the terminal device 2 for which a connection request has been made via, for example, the network NW1 and communicates various information.

The storage unit 12 stores information used for various processes executed by the information processing device 1. The storage unit 12 includes, for example, a service storage unit 121 and a learning result storage unit 122.
The service storage unit 121 stores, for example, posted information such as a question sentence and an answer sentence posted from the terminal device 2 to the Q & A service by the user.
The learning result storage unit 122 stores the learning result machine-learned by the learning processing unit 132, which will be described later. The details of the learning result will be described later.

The control unit 13 is, for example, a processor including a CPU (Central Processing Unit) and the like, and controls the information processing device 1 in an integrated manner. The control unit 13 is, for example, a process of providing an information service such as the above-mentioned Q & A service, a process of generating a learning result stored in the learning result storage unit 122, and a question acquired by the information processing device 1 via the NW communication unit 11. Performs various processes such as generating answer sentences for sentences. Further, the control unit 13 includes a service providing unit 131, a learning processing unit 132, a question acquisition unit 133, and an answer generation unit 134.

The service providing unit 131 executes processing related to the information service provided by the information processing device 1. For example, the service providing unit 131 stores the received question text and answer text as posted information in the service storage unit 121 from the terminal device 2 via the NW communication unit 11. Further, the service providing unit 131 outputs the posted information stored in the service storage unit 121 to the terminal device 2 via the NW communication unit 11, for example, for the terminal device 2 wishing to browse the Q & A service. , Displayed on the terminal device 2. Further, the service providing unit 131 stores the answer sentence generated by the answer generating unit 134, which will be described later, in the service storage unit 121. In the Q & A service, the service providing unit 131 provides information to the user by dividing it into categories (classifications) such as "romance", "family", and "cooking".

The learning processing unit 132 provides learning information having a plurality of sets of a question sentence and a partial answer sentence corresponding to each of a plurality of partial items divided by a predetermined sentence path (scenario) in a known answer sentence. Based on this, machine learning is executed to generate learning results. Here, for example, when the scenario of the text of the answer sentence is defined in the order of "conclusion" and "supplement", the sub-items are "conclusion" and "supplement", and the text of the answer sentence When the scenario is defined in the order of "chat", "case", and "conclusion", the sub-items are "chat", "case", and "conclusion".
In the following description, the cases of "conclusion" and "supplement" will be described as an example of a plurality of sub-items.

Further, the learning processing unit 132 uses, for example, the posted information (known question text and known answer text) of the Q & A service stored in the service storage unit 121 as input information, and is a learning model using deep learning technology. Generates the learning result learned by. The known answer sentences include a correct answer sentence for the question sentence and an incorrect answer sentence. The learning processing unit 132 uses a set of a question sentence, a correct answer sentence, and an incorrect answer sentence arbitrarily extracted from the answer sentences other than the correct answer as learning information as input information (learning information) of the learning process. .. The learning processing unit 132 stores the generated learning result in the learning result storage unit 122.
The learning model of the present embodiment, the configuration of the learning processing unit 132, and the details of the learning processing will be described later.

The question acquisition unit 133 acquires the input question text input to the information processing device 1. The question acquisition unit 133 acquires an input question sentence from, for example, the question sentence stored in the service storage unit 121.

The answer generation unit 134 generates an answer sentence for the input question sentence acquired by the question acquisition unit 133 based on the learning result learned by the learning processing unit 132 described above. That is, the answer generation unit 134 generates an answer sentence generated by combining a plurality of sub-items based on the learning result stored in the learning result storage unit 122. Further, the answer generation unit 134 supplies the generated answer sentence to the service providing unit 131, and stores the answer sentence in the service storage unit 121 as a post of the answer of the input question sentence.

Next, the learning model in the present embodiment will be described with reference to FIGS. 2 and 3.
FIG. 2 is a diagram illustrating an example of the learning model M1 in the present embodiment.
As shown in FIG. 2, the learning model M1 in the present embodiment is a model using deep learning technology, and is a model in which an encoder / decoder model M11 and a sentence unit learning model M12 are combined.

The encoder / decoder model M11 encodes a question sentence word by word sequentially based on the order of the words to generate a context vector 10, and is known for each of a plurality of sub-items based on the generated context vector 10. Learn by decoding the partial answer sentence. The encoder / decoder model M11 is, for example, a neural encoder / decoder model. First, one word (token) of a question sentence is encoded at a time, and one word (token) of an answer sentence is decoded at a time. The encoder-decoder model M11 is a model for decoding and learning so as to generate a known answer sentence for each of a plurality of sub-items. The encoder / decoder model M11 is sequentially generated by adjusting the word (WORD) of the interrogative sentence by the input sentence X which is the interrogative sentence. The target sentence is Y = {y (1), ..., Y (T)}. In the encoder / decoder model M11, the loss function L is represented by the following equation (1).

The encoder / decoder model M11 includes an encoder model M111 and a decoder model M112.
The encoder model M111 uses biLSTM (bidirectional Long Short-Term Memory) encoded in both directions to capture the overall meaning of the question. The encoder model M111 generates the context vector 10 by a Max pooling process that extracts and accumulates the maximum value of each element of the output of biLSTM.

The decoder model M112 is a model that learns to output a partial answer sentence (conclusion answer sentence ac and supplementary answer sentence as) based on the context vector 10 generated by the encoder model M111, and is an LSTM (Long Short-Term). It is a model that predicts each word using Memory). Here, the decoder model M112 is a model that decodes the conclusion answer sentence ac and the supplementary answer sentence as based on the context vector 10.
The decoder model M112 is based on the context vector 10 encoded by the encoder model M111, for example, the question sentence q of "long distance", "love", "ha", "love", "o", ... Learn to decode known partial answer sentences (conclusion answer sentence ac) such as "long distance", "love", "ha", "truth", "love", and so on. Further, the decoder model M112 has known partial answer sentences (supplementary answer sentence as) such as "long distance", "love", "ha", "your", "love", ... Based on the context vector 10. ) To decode.

As shown in FIG. 3, the decoder model M112 decodes and learns a known partial answer sentence based on topic information related to each word in the partial answer sentence, and learns C-LSTM (Contextual-Long Short-Term). It is a model that learns by Memory).

FIG. 3 is a diagram illustrating an example of a decoder model according to the present embodiment.
In the example shown in FIG. 3, the topic information corresponding to the word "information" is "IT" (information technology), and the topic information corresponding to the word "society" is "public". The topic information corresponding to the words "what", "," (comma) and "ga" is "connection", and the topic information corresponding to the word "resource" is "IT". As described above, the topic information is, for example, information indicating a related field of a word.
The decoder model M112 learns to generate an answer sentence (partial answer sentence) based on the word string of the answer sentence (partial answer sentence) as shown in FIG. 3 and the topic information corresponding to each word. ..

Returning to the explanation of FIG. 2, the sentence unit learning model M12 uses the output (partial answer sentence) of the encoder / decoder model M11 described above and the question sentence q as input information to learn sentence by sentence. -sentence) It is a model. The sentence unit learning model M12 uses biLSTM and max pooling processing for each of the question sentence q and the partial answer sentences of the plurality of partial items, and uses a loss function Lw that combines the encoder / decoder model M11 and the decoder model M112. Optimized for learning.

In the sentence unit learning model M12, the answer sentence generated based on the intermediate learning result of the encoder-decoder model M11 and the context vector 10 is used as a partial answer sentence, and the set information including the partial answer sentence and the question sentence is used as input information. You may try to learn. Further, the input information of the sentence unit learning model M12 is, for example, information (feature vector) obtained by vectorizing a set information including a partial answer sentence and a question sentence.
Further, in the sentence unit learning model M12, based on the output corresponding to the first sub-item of the plurality of sub-items, the biLSTM corresponding to the second sub-item different from the first sub-item is updated and learned. The attention mechanism to be used may be used.

Next, the learning processing unit 132 and the learning process in the present embodiment will be described with reference to FIGS. 4 and 5.
FIG. 4 is a diagram illustrating an example of the learning processing unit 132 and the learning processing in the present embodiment. In the example shown in FIG. 4, an example will be described in which a known answer sentence that is a part of the input information for learning includes a correct answer sentence and an incorrect answer sentence for the question sentence.

As shown in FIG. 4, the learning processing unit 132 includes an encoder / decoder model M11, a QA-LSTM (Question Answering-Long Short-Term Memory) unit (20-1, 20-2), and a loss function generation unit 30. It has.

The encoder / decoder model M11 is a model in which the encoder / decoder model M11 shown in FIG. 2 described above is associated with a correct answer sentence and an incorrect answer sentence, and is, for example, a model as shown in FIG. Here, the encoder / decoder model M11 includes the question sentence q, the correct answer sentence ac + of the "conclusion", the incorrect answer sentence ac- of the "conclusion", the correct answer sentence as + of the "supplement", and the "supplementary" included in the learning information set. Learn based on the incorrect sentence as-.

FIG. 5 is a diagram illustrating an example of the encoder / decoder model M11 in the present embodiment.
As shown in FIG. 5, the encoder / decoder model M11 includes an encoder model M111 and a decoder model M112, and includes the above-mentioned question sentence q, the correct answer sentence ac + of the “conclusion”, and the incorrect answer sentence ac− of the “conclusion”. , "Supplement" correct sentence as +, and "supplement" incorrect sentence as-.

Similar to FIG. 2 described above, the encoder model M111 encodes the question sentence q word by word based on the order of the words to generate the context vector 10.
Further, the decoder model M112 has the correct answer sentence ac + of the "conclusion", the incorrect answer sentence ac- of the "conclusion", the correct answer sentence as + of the "supplement", and the incorrect answer sentence aq of the "supplement" based on the context vector 10. Decode and learn to generate each of-.
In this way, the encoder / decoder model M11 decodes and learns the correct answer sentence and the incorrect answer sentence for each of a plurality of sub-items based on the context vector 10. Here, the correct and incorrect sentences of the "conclusion" generated by the encoder / decoder model M11 and the correct and incorrect sentences of the "supplement" are the correct and incorrect sentences of the "conclusion", the correct and incorrect sentences of the "conclusion", and the incorrect answers of the "conclusion". Sentence ac2-, correct sentence as2 + of "supplement", and incorrect answer sentence as2- of "supplement". Further, the correct answer sentence ac2 + of the "conclusion", the incorrect answer sentence ac2- of the "conclusion", the correct answer sentence as2 + of the "supplement", and the incorrect answer sentence as2- of the "supplementary" are output as feature vectors.

Returning to the explanation of FIG. 4, the learning processing unit 132 includes the question sentence q which is the output of the encoder / decoder model M11, the correct answer sentence ac2 + of the “conclusion”, the incorrect answer sentence ac2- of the “conclusion”, and the correct answer sentence as2 + of the “supplement”. , And the incorrect answer sentence as2- of "supplementary" are converted into each feature vector group. That is, the learning processing unit 132 converts the question sentence q into the feature vector sequence W _{q, which is a set of feature vectors.} Further, the learning processing unit 132 converts the correct answer sentence ac2 + of _{the "conclusion" into the feature vector sequence W ac +} which is a set of feature vectors, and the incorrect answer sentence ac2- of the "conclusion" is a feature which is a set of feature vectors. Convert to vector sequence W _ac−. Also, the learning processing unit 132, a correct sentence as2 + a "supplemental" into a feature vector sequence W _{the as +} a set of feature vectors, a set of non a correct sentence AS2-, feature vector "Supplement" feature Convert to vector sequence _Was−.

The QA-LSTM unit (20-1, 20-2) is a biLSTM, which is a neural network that learns in both directions. The QA-LSTM unit 20-1 is a biLSTM for "conclusion", and the QA-LSTM unit 20-2 is a biLSTM for "supplement". In the present embodiment, the QA-LSTM unit 20-1 and the QA-LSTM unit 20-2 have the same configuration, and indicate an arbitrary QA-LSTM unit included in the learning processing unit 132, or particularly. When no distinction is made, it will be described as QA-LSTM unit 20.

The QA-LSTM unit 20 includes a question embedding vector generation unit 21, a correct answer embedding vector generation unit 22, and an incorrect answer embedding vector generation unit 23.
The question embedding vector generation unit 21 generates a bidirectional vector sequence 24 (question feature vector group) in which a feature vector obtained by converting a question sentence for each word is generated based on the order of bidirectional words in the forward and reverse directions of the time series. ), The sequence of words is learned in both directions, and the question embedding vector _Oq is generated. Question embedding vector generation unit 21 generates, for example, a bi-directional vector sequence 24 from the feature vector sequence W _q question sentence, by Max pooling process and accumulates the extracted maximum value of each element of the bidirectional vector sequence 24 Generate the question embedding vector O _q (question intermediate vector).

The correct answer embedded vector generation unit 22 generates a bidirectional vector sequence 25 (correct answer feature vector group) in which a feature vector obtained by converting a correct sentence for each word is generated based on the order of bidirectional words in the forward and reverse directions of the time series. ), The sequence of words is learned in both directions, and the correct answer embedding vector O _{a +} is generated. The correct answer embedded vector generation unit 22 generates, for example, _{a bidirectional vector string 25 from the feature vector string Wa + of} the correct sentence, extracts the maximum value of each element of the bidirectional vector string 25, and accumulates it by max pooling processing. The correct answer embedding vector Oa ₊ (correct answer intermediate vector) is generated.

The incorrect answer embedding vector generation unit 23 generates a bidirectional vector sequence 26 (incorrect answer) in which a feature vector obtained by converting an incorrect answer sentence for each word is generated based on the order of bidirectional words in the forward and reverse directions of the time series. Based on the feature vector group), the sequence of words is learned in both directions to generate the _{correct embedding vector Oa−.} The incorrect answer embedding vector generation unit 23 generates, for example, a _{bidirectional vector string 26 from the feature vector string Wa −} of the incorrect answer sentence, extracts the maximum value of each element of the bidirectional vector string 26, and accumulates the maximum value. Incorrect answer embedding vector _Oa- (incorrect answer intermediate vector) is generated by pooling processing.

The RSTM, which is the basis of the QA-LSTM unit 20, is disclosed in Non-Patent Document 1. In the basic LSTM, when learning, the input time series input X = {x (1), x (2), ..., X (N)}, and x (t) is the t-th. When the feature vector of a word is used, each bidirectional vector h (t), which is an internal vector of the bidirectional vector sequence (24, 25, 26), is updated by the following equation (2) every t time. ..

Here, the basic architecture of LSTM, 3 one gate _{(input i t, forget f t} , output o t) and, there is a cell memory vector _{c t.} Also, σ () is a sigmoid function. _{Further, W i, W f, W} o, W c, U i, U f, U o, U c, b i, b f, b o, the _{b c} a network parameter to be learned.

The QA-LSTM part 20-1 is a biLSTM for "conclusion", and based on the feature vector sequence (W _q , W _{ac +} , W _ac- ), the question embedding vector O _qc , the correct answer embedding vector O _{ac +} , and no Generate the correct embedding vector O _ac−. The QA-LSTM unit 20-1 includes a question embedding vector generation unit 21-1, a correct answer embedding vector generation unit 22-1, and an incorrect answer embedding vector generation unit 23-1. The question embedding vector generation unit 21-1 generates a _{bidirectional vector sequence 24-1 from the feature vector sequence W q,} and generates a question embedding vector O _qc by max pooling processing. Further, the correct answer embedded vector generation unit 22-1 generates the _{bidirectional vector sequence 25-1 from the feature vector sequence W ac +,} and generates the correct answer embedded vector O _{ac +} by the max pooling process. Further, the incorrect answer embedded vector generation unit 23-1 generates the _{bidirectional vector sequence 26-1 from the feature vector sequence W ac−,} and generates the correct answer embedded vector O _ac− by the max pooling process.

The QA-LSTM unit 20-2 is a biLSTM for "supplementation", and is based on the feature vector sequence (W _q , _{Was +} , Was-), and the question embedding vector O _qs , the correct embedding vector Os +, and the non- _{question embedding vector O as +} Generate the correct embedded vector O _as-. The QA-LSTM unit 20-2 includes a question embedding vector generation unit 21-2, a correct answer embedding vector generation unit 22-2, and an incorrect answer embedding vector generation unit 23-2. The question-embedded vector generation unit 21-2 generates the _{bidirectional vector sequence 24-2 from the feature vector sequence W q,} and generates the question-embedded vector O _qs by the max pooling process. Also, correct embedding vector generation unit 22-2 generates a bidirectional vector sequence 25-2 from the feature vector sequence _{W the as +,} generating the correct embedded vector _{O the as +} by Max pooling process. Also, incorrect embedding vector generation unit 23-2 generates a bidirectional vector sequence 26-2 from the feature vector sequence _{W as-,} generates an incorrect embedding vectors _{O as-by} Max pooling process.

Further, the QA-LSTM unit 20-2 updates the bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2 by utilizing the attention mechanism when learning. The QA-LSTM unit 20-2 corresponds to the "supplement" based on _{, for example, the correct answer embedding vector O as +} and the incorrect answer embedding vector O _as- corresponding to the "conclusion" generated by the QA-LSTM unit 20-1. The bidirectional vector sequence 25-2 (correct feature vector group) and the bidirectional vector sequence 26-2 (incorrect feature vector group) are updated. Specifically, the QA-LSTM unit 20-2 uses the following equation (3) to obtain a bidirectional vector h _s (t) which is an internal vector of the bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2. Update.

Here, t is a step of time, and W _sm , W _cm , and w _mb are attention parameters. Further, ^~ h _s (t) indicates an updated bidirectional vector. In addition, the above-mentioned ~ in the text represents the symbol attached directly above the character.

Also, questions embedding vectors _O QA-LSTM portion 20-1 has generated _qc, correct embedded vector _{O ac +,} and an incorrect embedding vectors _{O ac-question} embedding vectors _{O qs} to QA-LSTM portion 20-2 has generated, Based on the correct answer embedding vector Os ₊ and the incorrect embedding vector _Os− , the loss function Ls using the cosine similarity is expressed by, for example, the following equation (4). That is, the loss function Ls corresponding to the sentence unit learning model M12 described above is expressed by the equation (4). In the sentence unit learning model M12, the loss function Ls is calculated based on the combination of "conclusion", "supplement", "correct answer", and "incorrect answer".

Here, [y, z] indicates the connection between the vector y and the vector z. O _q is [O _qc , O _qs ]. Further, M indicates a constant, and k (0 <k <1) is a parameter for controlling the margin.

The loss function generation unit 30 calculates the loss function Lw by combining the encoder / decoder model M11 and the sentence unit learning model M12. That is, the loss function generation unit 30 combines the loss function L of the above equation (1) and the loss function Ls of the equation (4) to calculate the loss function Lw by the following equation (5).

Here, α is a parameter that controls the weighting with two loss functions.
The loss function Lw is set so that the cosine value in each combination of the question and the answer being learned is maximized.

The learning processing unit 132 optimizes by the loss function Lw calculated using the above configuration to generate a learning result, and stores the generated learning result in the learning result storage unit 122. Learning processing unit 132, a parameter set _{{W i} for "conclusion" of the above-mentioned formula _{(2), W f, W} o, W c, U i, U f, U o, U c, b i, b f _{, b} o, _{b c}} c and the parameter set for "supplementary" _{{W i, W f, W} o, W c, U i, U f, U o, U c, b i, b f, b o , B _c } _s and the attention parameters {W _sm , W _cm , w _mb } to generate a learning result.

In the above-mentioned example, the example in which the scenario of the answer sentence is composed of two sub-items of "conclusion" and "supplement" has been described, but it may be composed of two or more sub-items. In that case, the learning result learned by the learning processing unit 132 _{corresponds to the question embedding vector O q} generated from the question sentence, the correct answer embedding vector O _{a +} corresponding to each of the plurality of sub-items, and each of the plurality of sub-items. It is learned is optimized by loss function Lw calculated based on a combination of a plurality of portions items of the the incorrect intermediate vector O _a-.

Next, the operation of the information processing apparatus 1 according to the present embodiment will be described with reference to the drawings.
<Learning process>
First, the learning process of the learning processing unit 132 in the information processing device 1 will be described with reference to FIG.

FIG. 6 is a flowchart showing an example of the learning process in the present embodiment.
As shown in FIG. 6, the learning processing unit 132 first substitutes “1” for the variable n (step S101). The variable n counts the number of times the learning is repeated. In this example, a case where learning is repeated NN times will be described.

Next, the learning processing unit 132 acquires a set of a question sentence and an answer sentence as input information (step S102). As shown in FIG. 4, the learning processing unit 132 includes, for example, a question sentence q, a conclusion answer sentence ac, and a supplementary answer sentence as (“Conclusion”) from among a plurality of existing question sentences and existing answer sentences. "Correct sentence ac +," conclusion "incorrect sentence ac-," supplement "correct answer sentence as +, and" supplement "incorrect sentence as-).

Next, the learning processing unit 132 encodes the question sentence q (step S103). The learning processing unit 132 encodes the acquired question sentence q by the encoder model M111 of the encoder-decoder model M11 described above. That is, the learning processing unit 132 encodes the acquired question sentence q by biLSTM and max pooling processing to generate the context vector 10. The context vector 10 here is a question vector.

Next, the learning processing unit 132 decodes the conclusion answer sentence ac and the supplementary answer sentence as (step S104). The learning processing unit 132 uses, for example, two type identification vectors (sub-item vectors) prepared in advance in order to learn the conclusion answer sentence ac and the supplementary answer sentence as. The learning processing unit 132 uses the type identification vector as input information together with the context vector 10 generated by the encoder model M111 in the decoder model M112 of the encoder decoder model M11, and answers the target type (partial item) (conclusion answer). Decode so that it becomes sentence ac and supplementary answer sentence as).

Here, the output string (word string) of the decoder model M112 constitutes one word (token) at a time. The type identification vector is also similar in that it is a context vector in the C-LSTM model. As a result, the learning processing unit 132 can generate a target sequence (answer sentence) according to the input answer type identification information (“conclusion” and “supplement”) by a single encoder / decoder network.
As shown in FIG. 5, the learning processing unit 132 decodes the correct answer sentence and the incorrect answer sentence for each of the “conclusion” and the “supplement”. Further, in step S104, the learning processing unit 132 learns the answer sentences of "conclusion" and "supplement" one character at a time (one word).

Next, the learning processing unit 132 generates the conclusion answer sentence ac2 and the supplementary answer sentence as2 (step S105). In the above-mentioned encoder-decoder learning process learned for the set information of one set of question sentence q, conclusion answer sentence ac, and supplementary answer sentence as, the learning processing unit 132 serves as an input for predicting the output next. , The predicted output words are simply supplied to generate the conclusion answer sentence ac2 and the supplementary answer sentence as2. In this process, when the learning in the process of step S104 described above progresses to EOS (end of sequence; the last character of the conclusion answer sentence ac or the supplementary answer sentence as), the update parameter (up to that point) is performed. This is a process of temporarily generating the conclusion answer sentence ac2 and the supplementary answer sentence as2 based on the intermediate learning result). That is, the learning processing unit 132 generates the conclusion answer sentence ac2 and the supplementary answer sentence as2 based on the intermediate learning result of the encoder / decoder model M11 and the context vector 10. Specifically, as shown in FIG. 4, the learning processing unit 132 has the correct answer sentence ac2 + of the “conclusion”, the incorrect answer sentence ac2- of the “conclusion”, the correct answer sentence as2 + of the “supplement”, and the incorrect answer of the “supplement”. Generate sentence as2-.

Next, the learning processing unit 132 inputs the question sentence q, the conclusion answer sentence ac2, and the supplementary answer sentence as2 into the sentence unit learning model M12 to generate the loss function Lw (step S106). The learning processing unit 132 generates the loss function Lw by, for example, the following processes (a) to (f).
(A) First, as shown in FIG. 4, the learning processing unit 132 first asks the question sentence q, the correct answer sentence ac2 + of the “conclusion”, the incorrect answer sentence ac2- of the “conclusion”, the correct answer sentence as2 + of the “supplement”, and “supplement”. The set information of the incorrect answer sentence as2- is converted into a feature vector sequence (W _q , W _{ac +} , W _{ac −, Was +} , _{Was −} ).

(B) Next, QA-LSTM 20-1 of the learning processing unit 132, feature vector sequence _{(W q, W ac +,} W ac-) based on the question embedded vector _{O qc,} correct embedded vector _{O ac} +, _And generate the incorrect embedding vector O ac−. Further, the QA-LSTM unit 20-1 is the t-th bidirectional vector (h _qc (t), h _{ac +} (t), h) of the bidirectional vector sequence (24-1, 25-1, 26-1). _{Each of ac−} (t)) is max-pooled to generate a _{question embedding vector O qc} , a correct embedding vector O _{ac +} , and an incorrect embedding vector O _ac−.

(C) Next, the QA-LSTM unit 20-2 of the learning processing unit 132 first generates the _{question embedding vector O qs.} The QA-LSTM unit 20-2 max-pools each t-th bidirectional vector h _qs (t) of the bidirectional vector sequence 24-2 to generate a question embedding vector O _qs.

(D) Next, the QA-LSTM unit 20-2 is the t-th bidirectional vector (has ₊ (t), _has- (t)) of the bidirectional vector sequence (25-2, 26-2). _Is updated using _{the correct embedding vector O ac +} and the incorrect embedding vector O ac −. That is, the QA-LSTM unit 20-2 updates each bidirectional vector (has ₊ (t), _has- (t)) using the above equation (2), and updates the update vector ( ^~ _{has +).} (T), ^~ h _as- (t)) is generated.

(E) Next, the QA-LSTM unit 20-2 max-pools the update vectors ( ^~ h _{as +} (t), ^~ h _as- (t)) to obtain the correct embedding vector O _{as +} and the incorrect answer. Generate the embedded vector O _as-.
(F) The loss function generation unit 30 of the learning processing unit 132, the generated embedding vectors _{(O qc, O ac +,} O ac-, O qs, O as +, O as-), and the formula (5) Based on this, the loss function Lw is generated.

Next, the learning processing unit 132 optimizes with the generated loss function Lw (step S107). The learning processing unit 132 optimizes each parameter by the calculated loss function Lw. Learning processing unit 132, for example, parameter set _{{W i} for _{"conclusion", W f, W o, W} c, U i, U f, U o, U c, b i, b f, b o, b _{c} c} and the parameter set for "supplementary" _{{W i, W f, W} o, W c, U i, U f, U o, U c, b i, b f, b o, b c} s And the attention parameters {W _sm , W _cm , w _mb }. The learning processing unit 132 stores each optimized parameter as a learning result in the learning result storage unit 122.

Next, the learning processing unit 132 determines whether or not the input information is completed (step S108). The learning processing unit 132 includes a question sentence q, which is input information, a conclusion answer sentence ac, and a supplementary answer sentence as (correct answer sentence ac + of "conclusion", incorrect answer sentence ac- of "conclusion", correct answer sentence as + of "supplement", And, it is determined whether or not there is the following set information with the incorrect answer sentence as-) of "Supplement". When the input information is completed (step S108: YES), the learning processing unit 132 advances the processing to step S109. Further, the learning processing unit 132 returns the processing to step S102 when the input information is not completed (there is the next input information) (step S108: YES).

In step S109, the learning processing unit 132 determines whether or not the variable n is equal to or greater than the number of repetitions NN (n ≧ NN). That is, the learning processing unit 132 determines whether or not the processing from step S102 to step S108 has been repeated NN times for learning. The learning processing unit 132 ends this learning processing when the variable n is the number of repetitions NN or more (step S109: YES). Further, the learning processing unit 132 advances the processing to step S110 when the variable n is less than the number of repetitions NN (step S109: NO).

In step S110, the learning processing unit 132 assigns “n + 1” to the variable n and returns the processing to step S102. That is, the learning processing unit 132 adds "1" to the value of the variable n and returns the processing to step S102.
In this way, the learning processing unit 132 repeats the processes from step S102 to step S108 NN times to learn, and stores the learning result in the learning result storage unit 122.

<Answer sentence generation process>
Next, a process of generating an answer sentence from the question sentence of the information processing apparatus 1 in the present embodiment will be described with reference to the drawings.
FIG. 7 is a flowchart showing an example of a process of generating an answer sentence from the question sentence of the information processing apparatus 1 in the present embodiment.

As shown in FIG. 7, the information processing apparatus 1 first acquires a question sentence from the service storage unit 121 (step S201). The question acquisition unit 133 of the information processing device 1 acquires an input question sentence from the question sentences stored in the service storage unit 121.

Next, the answer generation unit 134 of the information processing device 1 generates an answer sentence based on the input question sentence and the learning result stored in the learning result storage unit 122 (step S202). The learning result here includes an encoder / decoder model M11 that learns in character units (word units) described above by the learning processing unit 132, and a sentence unit learning model M12 that is a sentence-by-sentence model that learns sentence by sentence. It is learned in combination. That is, the answer generation unit 134 generates an answer sentence from the input question sentence based on the learning result, for example, which was learned by the learning process shown in FIG.

The answer generation unit 134 does not simply select an existing answer sentence, but generates a new answer sentence by appropriately combining the partial answer sentences of the partial items in consideration of the course of the sentence. Further, the answer generation unit 134 does not simply select the "conclusion" and "supplementary" partial answer sentences from the existing answer sentences by combining the encoder / decoder model M11 that learns in character units (word units). , Select a partial answer sentence word by word to generate a new answer sentence.

Next, the answer generation unit 134 stores the answer sentence in the service storage unit 121 (step S203). That is, the answer generation unit 134 supplies the generated answer sentence to the service providing unit 131, and stores the answer sentence in the service storage unit 121 as a post of the answer of the input question sentence. As a result, the information processing device 1 is connected to the information processing device 1 via the network NW1, and the answer sentence generated by the answer generation unit 134 can be viewed from the terminal device 2 for the question sentence. After the process of step S203, the information processing device 1 ends the process of generating the answer sentence.

Next, with reference to FIG. 8, the evaluation result of the response sentence generated by the information processing apparatus 1 according to the present embodiment will be described.

FIG. 8 is a diagram showing a comparison between the method in the present embodiment and the prior art.
In FIG. 8, “Seq2seq” shows the evaluation result when only the loss function Ls according to the equation (4) by the sentence unit learning model M12, which is a sentence-by-sentence model, is used for comparison. Further, "C-LSTM" is an evaluation result when the loss function L according to the equation (1) by the above-mentioned C-LSTM model is used.

Further, the "method of the present embodiment" is a method when the loss function Ls according to the equation (5) by the learning processing unit 132 of the present embodiment described above is used.
The evaluation results "ROUGE-1" (uni-gram), "ROUGE-2" (bigram), and "ROUGE-L" (Longest common subsequence) are from the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) series. It is an evaluation index. "ROUGE-1", "ROUGE-2", and "ROUGE-L" are all shown in the range of "0" to "1.0", and the larger the value, the higher the evaluation. There is. For more information on ROUGE, see, for example, the technical literature (Chin-Yew Lin, “ROUGE: A Package for Automatic Evaluation of Summaries”, In Text Summarization Branches Out: Proceedings of the ACL-04 Workshop, pages 74-81, 2004). .)It is described in.

In addition, the learning information used in the evaluation is 5000 sets of question sentences accumulated in 16 categories including "love consultation", "travel", "medical care", etc. in the Q & A service "Teach me goo". You are using the answer text. In addition, there are two sub-items, "conclusion" and "supplement".

In addition, as test data for evaluation, two sets of 100 sets of question sentence q, conclusion answer sentence ac, and supplementary answer sentence as are prepared, and the evaluation result shows an average value. Further, in the evaluation, the number of embedded characters is set to "300", α in the formula (5) is set to "0.5", and the number of repetitions NN is set to "500".

As shown in FIG. 8, the evaluation results of the "method according to the present embodiment" are "ROUGE-1" of "0.4546", "ROUGE-2" of "0.2030", and "ROUGE-". "L" is "0.3311". The evaluation result of the "method according to the present embodiment" is higher than that of the conventional "Seq2seq" and "C-LSTM" for all of "ROUGE-1", "ROUGE-2", and "ROUGE-L". is there.

Further, FIG. 9 is a diagram illustrating an example of a response sentence generated by the information processing device 1 in the present embodiment.
In addition, in FIG. 9, "answer sentence generated by C-LSTM" shows the answer sentence when the above-mentioned C-LSTM model is used for comparison.
As shown in Fig. 9, in the "answer sentence generated by C-LSTM", in response to the question sentence, "I think that there is a person who cannot think of the other person. It is better to wait for the size. I think it's good. ”Is an unnatural answer.

On the other hand, the answer sentence generated by the information processing device 1 according to this embodiment is "I think you should take courage and convey your feelings. You can see how indecisive it is." It is possible to generate a natural written answer with less discomfort.

As described above, the information processing apparatus 1 according to the present embodiment includes a question acquisition unit 133 and an answer generation unit 134. The question acquisition unit 133 acquires the input input question sentence. The answer generation unit 134 is a machine based on learning information having a plurality of sets of a question sentence and a known partial answer sentence corresponding to each of a plurality of partial items divided by a predetermined sentence path in the answer sentence. Based on the learned learning result, the answer sentence to the input question sentence acquired by the question acquisition unit 133 is generated. Here, the learning result is optimized and learned by the loss function Lw calculated by combining the encoder / decoder model M11 and the sentence unit learning model M12. Further, the encoder / decoder model M11 generates a context vector 10 by sequentially encoding a question sentence word by word based on the order of the words, and based on the generated context vector 10, known partial answers for each of a plurality of partial items. Decode and learn sentences.

Further, the sentence unit learning model M12 uses set information including a partial answer sentence for each of a plurality of partial items decoded based on the encoder / decoder model M11 and a question sentence as input information. The sentence unit learning model M12 includes a question embedding vector O _q (question intermediate vector), an answer intermediate vector corresponding to each of a plurality of sub-items (correct answer embedding vector (O _{ac +} , Os ₊ ), and an incorrect answer embedding vector (O _ac−). , O _as- )) and, learning based on the combination of multiple sub-items. Here, the question embedding vector O _q (question intermediate vector) generates a feature vector in which the question sentence of the input information is converted for each word based on the order of the words in both the forward and reverse directions of the time series. It is generated by learning the sequence of words in both directions based on the bidirectional vector sequence 24 (question feature vector group). In addition, the answer intermediate vector (correct answer embedding vector (O _{ac +} , O _{as +} ) and incorrect answer embedding vector (O _ac- , O _as- )) is a bidirectional word that is a feature vector in which a partial answer sentence is converted word by word. Based on the answer feature vector group (bidirectional vector string 25 and bidirectional vector sequence 26) generated based on the order of the words, the word sequence is learned in both directions and generated.

As a result, the information processing device 1 according to the present embodiment generates an answer sentence based on the learning result obtained by combining the encoder / decoder model M11 and the sentence unit learning model M12 and collectively learning, so that the answer sentence is generated in character (word) units. In addition to being able to generate an answer sentence with, it is possible to create a new answer sentence by optimizing the connection of the answers of each sub-item and combining the answer sentences of each selected sub-item. Therefore, the information processing device 1 according to the present embodiment can generate a natural written answer to the question with less discomfort. That is, the information processing device 1 according to the present embodiment has a natural word (word) or sentence order, and can generate a more complicated answer than a short sentence answer by a conventional chatbot. As described above, the information processing apparatus 1 according to the present embodiment can generate one character and can be applied to answers to long sentences, complicated, and various non-factoid type questions.

Further, in the present embodiment, the sentence unit learning model M12 is an answer sentence generated based on the intermediate learning result of the encoder / decoder model M11 and the context vector 10 (for example, the correct answer sentence ac2 + of the “conclusion” and the incorrect answer of the “conclusion””. Sentence ac2-, correct answer sentence as2 + of "supplement", incorrect answer sentence as2- of "supplement", etc.) are used as partial answer sentences. The sentence unit learning model M12 learns the set information including the partial answer sentence and the question sentence as input information.
As a result, the information processing device 1 according to the present embodiment learns the answer sentence, which is a more natural sentence regenerated based on the intermediate learning result of the encoder / decoder model M11, as input information. It is possible to generate a natural written answer with reduced.

Further, in the present embodiment, the encoder / decoder model M11 decodes and learns the known partial answer sentence based on the topic information related to each word in the known partial answer sentence. That is, the encoder / decoder model M11 is, for example, a model learned by C-LSTM.
As a result, the information processing device 1 according to the present embodiment can reduce the selection of characters (words) having low relevance based on the topic information, so that it is natural to further reduce the sense of discomfort on a character (word) basis. Can generate written answers.

Further, in the present embodiment, the known answer sentence includes a correct answer sentence for the question sentence and an incorrect answer sentence. The answer generation unit 134 generates an answer sentence based on the learning result machine-learned based on the learning information having a plurality of pairs of the question sentence and the correct answer sentence and the incorrect answer sentence corresponding to each of the plurality of sub-items. .. The encoder / decoder model M11 decodes and learns a correct answer sentence and an incorrect answer sentence for each of a plurality of sub-items based on the context vector 10. The sentence unit learning model M12 has a question embedding vector O _q (question intermediate vector), a correct answer embedding vector (O _{ac +} , O _{as +} ) (correct answer intermediate vector) corresponding to each of a plurality of sub-items, and a plurality of sub-items. Learning is based on a combination of a plurality of sub-items of the corresponding incorrect answer embedding vector (O _ac- , O _{as-) (incorrect answer intermediate vector).} Here, the correct answer embedding vector (O _{ac +} , O _{as +} ) is converted into a bidirectional vector sequence 25 (correct answer feature vector group) generated based on the bidirectional word arrangement order of the feature vector obtained by converting the correct sentence for each word. Based on this, it is generated by learning the sequence of words in both directions. In addition, the incorrect answer embedding vector (O _ac- , O _as- ) is a bidirectional vector sequence 26 (incorrect answer feature) in which an incorrect answer sentence is converted for each word and a feature vector is generated based on the bidirectional word arrangement order. It is generated by learning the sequence of words in both directions based on the vector group).

As a result, the information processing device 1 according to the present embodiment generates an answer sentence based on the learning result learned by using the correct answer sentence and the incorrect answer sentence. Can be generated.

_{Further, in the present embodiment, the learning result includes a correct answer embedding vector (O ac +} , O _{as +} ) and an incorrect answer embedding vector (O) corresponding to the first sub-item (for example, “conclusion”) of the plurality of sub-items. Bidirectional vector sequence 25 and bidirectional vector sequence 26 (incorrect feature vector group) corresponding to a second sub-item (for example, "supplement") different from the first sub-item based on _ac- , O _as-). ) Is updated and learned. That is, the learning processing unit 132 updates each bidirectional vector (h (t)) of the bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2 by the attention mechanism using the above equation (3). To learn.
As a result, the information processing apparatus 1 according to the present embodiment can perform learning by optimizing the relationship between the sub-items (for example, the connection of the sub-items). Therefore, the information processing apparatus 1 according to the present embodiment can generate a more natural answer sentence by combining the partial items.

Further, the information processing device 1 according to the present embodiment includes a learning processing unit 132 that performs machine learning based on learning information and generates a learning result.
As a result, the information processing device 1 according to the present embodiment can learn by its own device and generate a learning result. Further, the information processing apparatus 1 according to the present embodiment can, for example, relearn to improve the answer to the question.

Further, in the present embodiment, for example, the loss function Lw is calculated based on the above equation (5). Since the loss function Lw simultaneously optimizes learning in character (word) units and combinations of each sub-item, the information processing apparatus 1 according to the present embodiment generates sub-items suitable for generating an answer sentence. can do.

The learning processing unit 132 may relearn the learning result periodically or based on a predetermined condition (for example, the evaluation value (for example, ROUGE) is lowered to a predetermined value or less). ..
As a result, the information processing apparatus 1 according to the present embodiment can improve the answer to the question in response to the change in time.

Further, the information processing method according to the present embodiment includes a question acquisition step and an answer generation step. In the question acquisition step, the question acquisition unit 133 acquires the input input question sentence. In the answer generation step, the answer generation unit 134 has a plurality of sets of a question sentence and a correct answer sentence and an incorrect answer sentence corresponding to each of a plurality of sub-items divided by a predetermined sentence path in the answer sentence. Based on the learning information, the answer sentence to the input question sentence acquired by the question acquisition step is generated based on the above-mentioned learning result obtained by machine learning.
As a result, the information processing method according to the present embodiment has the same effect as that of the information processing device 1 described above, and can generate a natural text answer to the question with less discomfort.

The present invention is not limited to the above embodiment, and can be modified without departing from the spirit of the present invention.
For example, in the above embodiment, the information processing device 1 includes the learning processing unit 132, but the present invention is not limited to this, and the learning processing unit 132 may be used as long as the learning result can be obtained. You don't have to prepare.

Further, although the information processing device 1 has described an example including the service storage unit 121 and the learning result storage unit 122, one or both of the service storage unit 121 and the learning result storage unit 122 may be used, for example. It may be provided by an external server device. Further, the information processing device 1 may include a part of the functional unit included in the control unit 13 in the external server device.
In the above embodiment, the information processing device 1 is configured by one server device, but the information processing device 1 may be composed of a plurality of devices.

Further, in the above embodiment, the information processing apparatus 1 has described an example in which the answer sentence is composed of two sub-items of "conclusion" and "supplement", but the present invention is not limited to this, and 3 It may correspond to one or more sub-items.
Further, in the sentence unit learning model M12 of the above embodiment, an example in which the information processing apparatus 1 applies a method including a QA-LSTM unit 20 for each sub-item and a method based on an attention mechanism has been described. It is not limited. The information processing device 1 may be in a form in which, for example, a part of these methods is not applied in the sentence unit learning model M12.

Further, in the above embodiment, an example in which the loss function Ls of the sentence unit learning model M12 is calculated by the above-mentioned equation (4) has been described, but the present invention is not limited to this, and the following equation (6) is used. It may be calculated by.

Further, in the learning process of the above embodiment, the answer sentences (for example, conclusion answer sentence ac2, supplementary answer sentence as2, etc.) regenerated based on the intermediate learning result of the encoder decoder model M11 and the context vector 10 are used as a sentence unit learning model. An example of using the input information of M12 has been described, but the present invention is not limited to this. The learning processing unit 132 may use the answer sentence before regeneration used for the learning of decoding as the input information of the sentence unit learning model M12.

Each configuration included in the information processing device 1 described above has a computer system inside. Then, a program for realizing the functions of each configuration included in the above-mentioned information processing apparatus 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed. Therefore, the processing in each configuration included in the information processing apparatus 1 described above may be performed. Here, "loading and executing a program recorded on a recording medium into a computer system" includes installing the program in the computer system. The term "computer system" as used herein includes hardware such as an OS and peripheral devices.
Further, the "computer system" may include a plurality of computer devices connected via a network including a communication line such as the Internet, WAN, LAN, and a dedicated line. Further, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built in a computer system. As described above, the recording medium in which the program is stored may be a non-transient recording medium such as a CD-ROM.

The recording medium also includes an internal or external recording medium that can be accessed from the distribution server to distribute the program. It should be noted that the program may be divided into a plurality of units, downloaded at different timings, and then combined with each configuration provided in the information processing device 1, or the distribution server for distributing each of the divided programs may be different. Furthermore, a "computer-readable recording medium" is a volatile memory (RAM) inside a computer system that serves as a server or client when a program is transmitted via a network, and holds the program for a certain period of time. It shall also include things. Further, the above program may be for realizing a part of the above-mentioned functions. Further, a so-called difference file (difference program) may be used, which can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Further, a part or all of the above-mentioned functions may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each of the above-mentioned functions may be made into a processor individually, or a part or all of them may be integrated into a processor. Further, the method of making an integrated circuit is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. Further, when an integrated circuit technology that replaces an LSI appears due to advances in semiconductor technology, an integrated circuit based on this technology may be used.

1 Information processing device 2 Terminal device 10 Context vector 11 NW communication unit 12 Storage unit 13 Control unit 20, 20-1, 20-2 QA-LSTM unit 21, 21-1, 21-2 Question embedded vector generation unit 22, 22 -1,22-2 Correct embedded vector generator 23,23-1,23-2 Incorrect embedded vector generator 24,24-1,24-2,25,25-1,25-2,26,26- 1, 26-2 Bidirectional vector sequence 30 Loss function generation unit 100 Information processing system 121 Service storage unit 122 Learning result storage unit 131 Service provision unit 132 Learning processing unit 133 Question acquisition unit 134 Answer generation unit M1 Learning model M11 Encoder decoder model M12 sentence unit learning model M111 encoder model M112 decoder model NW1 network

Claims (8)

The question acquisition section that acquires the input question text, and
Machine-learned learning results based on learning information that has multiple pairs of question sentences and known partial answer sentences corresponding to each of a plurality of partial items divided by a predetermined sentence path in the answer sentence. Based on this, it is provided with an answer generation unit that generates an answer sentence to the input question sentence acquired by the question acquisition unit.
The learning result is
The question sentence is sequentially encoded word by word based on the order of the words to generate a context vector, and based on the generated context vector, the known partial answer sentence for each of the plurality of partial items is decoded and learned. Encoder Decoder model and
Using the set information including the partial answer sentence for each of the plurality of partial items decoded based on the encoder-decoder model and the question sentence as input information, the feature vector obtained by converting the question sentence for each word is used. The question intermediate vector generated by learning the sequence of words in both directions based on the question feature vector group generated based on the sequence order of the words in both the forward and reverse directions of the series, and the plurality of questions. It is an answer intermediate vector corresponding to each of the sub-items of the above, and is based on the answer feature vector group generated by converting the partial answer sentence for each word based on the order of the words in both directions. A sentence-based learning model that learns based on the combination of the plurality of sub-items of the answer intermediate vector generated by learning the sequence of words in both directions.
An information processing device characterized in that it is optimized and learned by a loss function calculated by combining. In the sentence unit learning model, the intermediate learning result of the encoder / decoder model and the answer sentence generated based on the context vector are used as the partial answer sentence, and the set information including the partial answer sentence and the question sentence is the input information. The information processing apparatus according to claim 1, wherein the information processing apparatus is characterized by learning as. The encoder / decoder model according to claim 1 or 2, wherein the encoder / decoder model decodes and learns the known partial answer sentence based on topic information related to each word in the known partial answer sentence. Information processing equipment. The known answer sentence includes a correct answer sentence and an incorrect answer sentence for the question sentence.
The answer generation unit is based on the learning result machine-learned based on the learning information having a plurality of pairs of the question sentence and the correct answer sentence and the incorrect answer sentence corresponding to each of the plurality of sub-items. , Generate the answer sentence,
The encoder / decoder model decodes and learns the correct answer sentence and the incorrect answer sentence for each of the plurality of partial items based on the context vector, and learns the encoder / decoder model.
The sentence-based learning model is a question intermediate vector and a correct answer intermediate vector corresponding to each of the plurality of sub-items, and a feature vector obtained by converting the correct answer sentence for each word is a sequence of the words in both directions. A correct intermediate vector generated by learning the sequence of words in both directions based on a group of correct feature vectors generated based on the order, and an incorrect intermediate vector corresponding to each of the plurality of sub-items. It is generated by learning the sequence of words in both directions based on the incorrect feature vector group generated by converting the incorrect sentence for each word based on the sequence order of the words in both directions. The information processing apparatus according to any one of claims 1 to 3, wherein learning is performed based on a combination of the incorrect answer intermediate vector and the plurality of sub-items. The learning result is
A group of correct answer feature vectors corresponding to a second sub-item different from the first sub-item based on the correct intermediate vector corresponding to the first sub-item and the incorrect intermediate vector among the plurality of sub-items. The information processing apparatus according to claim 4, wherein the incorrect answer feature vector group is updated and learned. The information processing apparatus according to any one of claims 1 to 5, further comprising a learning processing unit that performs machine learning based on the learning information and generates the learning result. The question acquisition step, in which the question acquisition department acquires the input input question text,
Machine learning based on learning information in which the answer generation unit has a plurality of sets of a question sentence and a known partial answer sentence corresponding to each of a plurality of partial items divided by a predetermined sentence path in the answer sentence. Includes an answer generation step that generates an answer sentence to the input question sentence acquired by the question acquisition step based on the obtained learning result.
The learning result is
The question sentence is sequentially encoded word by word based on the order of the words to generate a context vector, and based on the generated context vector, the known partial answer sentence for each of the plurality of partial items is decoded and learned. Encoder Decoder model and
Using the set information including the partial answer sentence for each of the plurality of partial items decoded based on the encoder-decoder model and the question sentence as input information, the feature vector obtained by converting the question sentence for each word is used. The question intermediate vector generated by learning the sequence of words in both directions based on the question feature vector group generated based on the sequence order of the words in both the forward and reverse directions of the series, and the plurality of questions. It is an answer intermediate vector corresponding to each of the sub-items of the above, and is based on the answer feature vector group generated by converting the partial answer sentence for each word based on the order of the words in both directions. A sentence-based learning model that learns based on the combination of the plurality of sub-items of the answer intermediate vector generated by learning the sequence of words in both directions.
An information processing method characterized in that it is optimized and learned by a loss function calculated by combining. On the computer
The question acquisition step to acquire the entered input question text, and
Machine-learned learning results based on learning information that has multiple pairs of question sentences and known partial answer sentences corresponding to each of a plurality of partial items divided by a predetermined sentence path in the answer sentence. Based on this, it is a program for executing the answer generation step of generating the answer sentence to the input question sentence acquired by the question acquisition step.
The learning result is
The question sentence is sequentially encoded word by word based on the order of the words to generate a context vector, and based on the generated context vector, the known partial answer sentence for each of the plurality of partial items is decoded and learned. Encoder Decoder model and
Using the set information including the partial answer sentence for each of the plurality of partial items decoded based on the encoder-decoder model and the question sentence as input information, the feature vector obtained by converting the question sentence for each word is used. The question intermediate vector generated by learning the sequence of words in both directions based on the question feature vector group generated based on the sequence order of the words in both the forward and reverse directions of the series, and the plurality of questions. It is an answer intermediate vector corresponding to each of the sub-items of the above, and is based on the answer feature vector group generated by converting the partial answer sentence for each word based on the order of the words in both directions. A sentence-based learning model that learns based on the combination of the plurality of sub-items of the answer intermediate vector generated by learning the sequence of words in both directions.
A program characterized in that it is optimized and learned by a loss function calculated by combining.

Images