JP2018206307A - Information processing device, information processing method, and program - Google Patents

Abstract

To generate an answer of a natural text obtained by reducing discomfort for a question.SOLUTION: An information processing device includes an answer generation part for generating an answer sentence to an input question sentence on the basis of a machine-learned learning result. The learning result generates a context vector by encoding a question sentence on the basis of an alignment sequence of sequential words by one word at a time, and is learned by being optimized by a loss function calculated by combining an encoder/decoder model for performing learning by decoding a known partial answer sentence in each of a plurality of partial items on the basis of the generated context vector and a sentence unit learning model for performing learning on the basis of a combination of a plurality of partial items of a question intermediate vector generated by learning on the basis of alignment sequence of bidirectional words of the question sentence with set information including a partial answer sentence and a question sentence decoded on the basis of the encoder/decoder model and an answer intermediate vector generated by learning on the basis of alignment sequence of bidirectional words of the partial answer sentence, respectively corresponding to each of the plurality of partial items.SELECTED DRAWING: Figure 1

Description

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

  In recent years, a technique for outputting an answer to an inputted question using machine learning (for example, deep learning technique) is known (for example, see Non-Patent Document 1). In such an information processing apparatus using the conventional technology, for example, an answer estimated to be appropriate is selected and output from previously prepared answers 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-described conventional information processing apparatus, for example, when the question is a non-factorial type question that asks for an answer based on the explanation of the reason or the event, the answer becomes a complicated long sentence. Since a prepared known answer is output, it is difficult to generate a new answer. For this reason, in the conventional information processing apparatus described above, an unnatural answer with an uncomfortable feeling may be output 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 apparatus, an information processing method, and a program capable of generating a natural sentence answer with reduced discomfort for a question. It is to provide.

  In order to solve the above-described problem, an aspect of the present invention provides a question acquisition unit that acquires an input question sentence that is input, a question sentence, and answer sentences that are divided by a predetermined sentence path. Generate an answer sentence for the input question sentence acquired by the question acquisition unit based on a learning result machine-learned based on learning information having a plurality of pairs with known partial answer sentences corresponding to the respective partial items. An answer generation unit, wherein the learning result is generated by encoding the question sentence one word at a time based on the word arrangement order, and generating a context vector, and for each of the plurality of partial items based on the generated context vector An encoder / decoder model that decodes and learns the known partial response sentence, and each of the plurality of partial items decoded based on the encoder / decoder model Based on the combination information including the partial answer sentence and the question sentence as input information, the feature vector obtained by converting the question sentence for each word is based on the arrangement order of the words in both the forward and reverse directions in time series. A question intermediate vector generated by bi-directional learning of the word sequence based on the question feature vector group generated in response, and an answer intermediate vector corresponding to each of the plurality of partial items, the partial answer An answer intermediate vector generated by learning the word sequence bidirectionally based on an answer feature vector group generated based on the bidirectional order of the words. And a sentence unit learning model that learns based on a combination of the plurality of sub-items, and is learned by being optimized by a loss function that is calculated in combination. It is an information processing apparatus.

  Further, according to one aspect of the present invention, in the information processing apparatus, the sentence unit learning model uses an answer sentence generated based on an intermediate learning result of the encoder decoder model and the context vector as the partial answer sentence, A group information including a partial answer sentence and the question sentence is learned as the input information.

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

  Further, according to one aspect of the present invention, in the information processing apparatus, the known answer sentence includes a correct answer sentence and an incorrect answer sentence with respect to the question sentence, and the answer generation unit includes the question sentence and Generating the answer sentence based on the learning result machine-learned based on the learning information having a plurality of sets 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 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, the plurality of partial parts Correct feature vectors corresponding to each item, wherein the correct feature vectors are generated based on the bi-directional order of the words. A correct answer intermediate vector generated by learning the word sequence bidirectionally and an incorrect answer intermediate vector corresponding to each of the plurality of partial items, wherein the incorrect answer sentence is converted for each word. A plurality of incorrect answer intermediate vectors generated by learning the word arrangement bidirectionally based on an incorrect answer feature vector group generated based on the bidirectional word arrangement order. Learning based on a combination of the partial items.

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

  Further, according to one embodiment of the present invention, the information processing apparatus includes a learning processing unit that performs machine learning based on the learning information and generates the learning result.

  Further, according to one aspect of the present invention, the question acquisition unit acquires the input question sentence that has been input, and the answer generation unit divides the question sentence and the answer sentence according to a predetermined sentence path Answer to the input question sentence acquired by the question acquisition step based on learning results machine-learned based on learning information having a plurality of pairs with known partial answer sentences corresponding to each of the plurality of partial items An answer generation step of generating a sentence, wherein the learning result is obtained by encoding the question sentence one word at a time based on the word arrangement order to generate a context vector, and based on the generated context vector, An encoder / decoder model that decodes and learns the known partial response sentence for each sub-item, and is decoded based on the encoder / decoder model. In addition, the group information including the partial answer sentence for each of the plurality of partial items and the question sentence is used as input information, and the feature vector obtained by converting the question sentence for each word is used in both the forward and reverse directions in time series. A question intermediate vector generated by learning the word arrangement bidirectionally based on a question feature vector group generated on the basis of the arrangement order of the words in the direction, and an intermediate answer corresponding to each of the plurality of partial items Learning the word sequence bidirectionally based on a response feature vector group, which is a vector and the feature vector obtained by converting the partial answer sentence into words is generated based on the bidirectional word sequence. And a sentence unit learning model that learns based on the combination of the plurality of partial items of the answer intermediate vector generated An information processing method characterized in that it is learned.

  Further, according to one aspect of the present invention, each of a plurality of partial items divided by a predetermined sentence path in a question acquisition step, a question sentence, and an answer sentence for acquiring an input question sentence input to a computer An answer generation step of generating an answer sentence for 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 known partial answer sentences corresponding to The learning result is obtained by encoding the question sentence one word at a time based on the word arrangement order to generate a context vector, and on the basis of the generated context vector, An encoder / decoder model for decoding and learning the known partial response sentence for each partial item, and an encoder / decoder model based on the encoder / decoder model The group information including the partial answer sentence for each of the plurality of partial items decoded and the question sentence is used as input information, and the feature vector obtained by converting the question sentence for each word is used as the forward and reverse time series. Corresponding to each of the plurality of partial items and the question intermediate vector generated by learning the word arrangement bidirectionally based on the question feature vector group generated based on the word arrangement order of the bidirectional directions An answer intermediate vector, wherein the partial answer sentence is converted for each word based on an answer feature vector group generated based on the bidirectional order of the word order, Loss function calculated by combining an answer intermediate vector generated by learning in a direction and a sentence unit learning model that learns based on a combination of the plurality of partial items Is a program characterized in that it is learned more optimized.

  ADVANTAGE OF THE INVENTION According to this invention, the reply of the natural sentence which reduced discomfort can be produced | generated with respect to a question.

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 process part in this embodiment, and a learning process. 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 produces | generates an answer sentence from the question sentence of the information processing apparatus in this embodiment. It is a figure which shows the comparison with the reply production | generation method in this embodiment, and a prior art. It is a figure explaining an example of the reply sentence which the information processor in this embodiment generated.

  An information processing apparatus according to an embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a schematic block diagram illustrating an example of an information processing system 100 according to the present embodiment.
As illustrated 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 that displays posted questions and answers on the terminal device 2 connected to the information processing device 1 and shares information between users.

  The terminal device 2 is a client terminal used by a user in order to use an information service provided by the information processing system 100. In the example illustrated in FIG. 1, for simplicity of explanation, an example in which one terminal device 2 is connected to the information processing device 1 is illustrated. However, a plurality of terminal devices 2 are connected to the information processing device 1. May be connected.

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

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

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

  The control unit 13 is a processor including, for example, a CPU (Central Processing Unit) and the like, and comprehensively controls the information processing apparatus 1. The control unit 13 includes, for example, a process for providing an information service such as the Q & A service described above, a generation process for a learning result stored in the learning result storage unit 122, and a question acquired by the information processing apparatus 1 via the NW communication unit 11. Various processes such as an answer sentence generation process for a sentence are executed. 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 apparatus 1. For example, the service providing unit 131 causes the service storage unit 121 to store the received question sentence and answer sentence from the terminal device 2 via the NW communication unit 11 as post information. For example, the service providing unit 131 outputs the posting information stored in the service storage unit 121 to the terminal device 2 via the NW communication unit 11 for the terminal device 2 that desires to browse the Q & A service. To be displayed on the terminal device 2. In addition, the service providing unit 131 causes the service storage unit 121 to store an answer sentence generated by an answer generation unit 134 described later. In the Q & A service, the service providing unit 131 provides information to the user in categories (classifications) such as “love”, “family”, and “cooking”, for example.

The learning processing unit 132 generates learning information having a plurality of pairs 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 a learning result. Here, for example, when the path of the sentence of the answer sentence is defined in the order of “conclusion” and “supplement”, the partial items are “conclusion” and “supplement”, and the sentence of the answer sentence When the path (scenario) is determined in the order of “chat”, “example”, and “conclusion”, the partial items are “chat”, “example”, and “conclusion”.
In the following description, a case of “conclusion” and “supplement” will be described as an example of a plurality of partial items.

The learning processing unit 132 also uses, for example, a learning model that uses deep learning technology with the post information (known question sentence and known answer sentence) of the Q & A service stored in the service storage unit 121 as input information. The learning result learned by is generated. The known answer sentence includes a correct answer sentence and an incorrect answer sentence for the question sentence. The learning processing unit 132 uses, as learning information, a combination of a question sentence, a correct answer sentence, and an incorrect answer sentence extracted from answer sentences other than the correct answer as input information (learning information) for the learning process. . The learning processing unit 132 causes the learning result storage unit 122 to store the generated learning result.
The details of the learning model, the configuration of the learning processing unit 132, and the learning processing of this embodiment will be described later.

  The question acquisition unit 133 acquires an input question sentence input to the information processing apparatus 1. For example, the question acquisition unit 133 acquires an input question sentence from the question sentences 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 partial items based on the learning result stored in the learning result storage unit 122. In addition, the answer generation unit 134 supplies the generated answer sentence to the service providing unit 131 and causes the service storage unit 121 to store the answer sentence as a post of an answer to the input question sentence.

Next, a learning model in the present embodiment will be described with reference to FIGS.
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 a deep learning technique, 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 generates a context vector 10 by encoding the question sentence one word at a time based on the word arrangement order. Based on the generated context vector 10, a known item for each of the plurality of partial items is generated. Decode and learn partial answer sentences. The encoder decoder model M11 is, for example, a neural encoder decoder model, and first encodes one word (token) of a question sentence at a time and decodes one word (token) at a time in an answer sentence. The encoder / decoder model M11 is a model that learns by decoding so as to generate a known answer sentence for each of a plurality of partial items. The encoder / decoder model M11 sequentially generates words by adjusting the word (WORD) of the question sentence with the input sentence X that is the question sentence. The target sentence is Y = {y (1),..., Y (T)}. In the encoder / decoder model M11, the loss function L is expressed 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 in order to capture the whole meaning of the question. The encoder model M111 generates the context vector 10 by Max pooling processing that extracts and accumulates the maximum value of each element of the biLSTM output.

The decoder model M112 is a model that learns to output a partial answer sentence (a conclusion answer sentence ac and a supplemental answer sentence as) based on the context vector 10 generated by the encoder model M111, and is a LSTM (Long Short-Term This 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 also the supplementary answer sentence as based on the context vector 10.
The decoder model M112, for example, is based on the context vector 10 obtained by encoding the question sentence q “Long distance”, “Love”, “Ha”, “Love”, “O”,... By the encoder model M111. Learning is performed so as to decode known partial answer sentences (conclusion answer sentences ac) such as “long distance”, “love”, “ha”, “true”, “love”,. Also, the decoder model M112 is based on the context vector 10 and includes known partial answer sentences (supplementary answer sentences as, “long distance”, “love”, “ha”, “your”, “love”,...). ) To decode.

  As shown in FIG. 3, the decoder model M112 decodes 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). This is a model that learns by (Memory).

FIG. 3 is a diagram for explaining an example of the decoder model in 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 “to”, “,” (reading marks), and “ga” is “connection”, and the topic information corresponding to the word “resource” is “IT”. Thus, topic information is information which shows the related field of a word, for example.
The decoder model M112 learns to generate an answer sentence (partial answer sentence) based on a word string of an answer sentence (partial answer sentence) as shown in FIG. 3 and topic information corresponding to each word. .

  Returning to the description of FIG. 2, the sentence-by-sentence learning model M12 uses a sentence-by-sentence (Sentence-by sentence) for learning for each sentence using the output (partial answer sentence) of the encoder decoder model M11 and the question sentence q as input information. -sentence) model. The sentence unit learning model M12 uses a biLSTM and a max pooling process for each of the question sentence q and the partial answer sentences of a plurality of partial items, and combines a loss function Lw that combines the encoder decoder model M11 and the decoder model M112. Optimized for learning.

The sentence unit learning model M12 uses the answer sentence generated based on the learning result of the encoder decoder model M11 and the context vector 10 as a partial answer sentence, and sets information including the partial answer sentence and the question sentence as input information. You may make it learn. The input information of the sentence unit learning model M12 is, for example, information (feature vector) obtained by vectorizing group information including a partial answer sentence and a question sentence.
In addition, the sentence unit learning model M12 learns by updating the biLSTM corresponding to the second partial item different from the first partial item based on the output corresponding to the first partial item among the plurality of partial items. An attention mechanism may be used.

Next, the learning processing unit 132 and the learning process in the present embodiment will be described with reference to FIGS.
FIG. 4 is a diagram illustrating an example of the learning processing unit 132 and learning processing in the present embodiment. In the example illustrated in FIG. 4, an example will be described in which a known answer sentence that is a part of learning input information 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), 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, for example, a model as shown in FIG. Here, the encoder / decoder model M11 includes the question sentence q, the correct sentence ac + of the “conclusion”, the incorrect sentence ac− of the “conclusion”, the correct sentence as + of the “supplement”, and the “supplement” included in the set of learning information. "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. The question sentence q, the correct sentence ac + of the “conclusion”, and the incorrect answer sentence ac− of the “conclusion”. , Learning based on the set information of the correct sentence as + of “supplement” and the incorrect sentence as− of “supplement”.

Similarly to FIG. 2 described above, the encoder model M111 generates the context vector 10 by encoding the question sentence q one word at a time based on the word arrangement order.
Also, the decoder model M112 determines that the correct sentence ac + of “conclusion”, the incorrect sentence ac− of “conclusion”, the correct sentence as + of “supplement”, and the incorrect sentence aq of “supplement” based on the context vector 10. Decode and learn to generate each of-.
As described above, the encoder / decoder model M11 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 10. Here, the correct sentence and incorrect answer sentence of “conclusion” generated by the encoder / decoder model M11, the correct answer sentence and incorrect answer sentence of “supplement”, the correct answer ac2 + of “conclusion”, and the incorrect answer of “conclusion” The sentence ac2-, the correct sentence as2 + of "supplement", and the incorrect answer sentence as2- of "supplement". In addition, the correct sentence ac2 + of “conclusion”, the incorrect answer sentence ac2- of “conclusion”, the correct sentence as2 + of “supplement”, and the incorrect answer sentence as2- of “supplement” are output as feature vectors.

Returning to the description of FIG. 4, the learning processing unit 132 outputs the question sentence q, which is the output of the encoder decoder model M11, the correct sentence ac2 + of “conclusion”, the incorrect answer sentence ac2- of “conclusion”, and the correct sentence as2 + of “supplement”. And “supplement” incorrect answer sentence as2- are converted into respective feature vector groups. That is, the learning processing unit 132 converts the question sentence q into a feature vector sequence W _q that is a set of feature vectors. Further, the learning processing unit 132 converts the correct sentence ac2 + of “conclusion” into a feature vector string W _{ac +} that is a set of feature vectors, and the incorrect answer sentence ac2 of “conclusion” is a feature that is a set of feature vectors. Convert to a vector sequence W _ac- . Further, the learning processing unit 132 converts the “supplementary” correct sentence as2 + into a feature vector string W _{as +} that is a set of feature vectors, and the “supplementary” incorrect answer sentence as2− is a feature that is a set of feature vectors. Convert to vector sequence W _as- .

  The QA-LSTM unit (20-1, 20-2) is 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 any QA-LSTM unit included in the learning processing unit 132, or particularly When not distinguished, it demonstrates as the QA-LSTM part 20. FIG.

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 feature vector obtained by converting a question sentence for each word based on a bi-directional vector sequence 24 (question feature vector group) based on a time-series forward and reverse word arrangement order. ) To learn the word sequence bidirectionally to generate a question embedding vector _Oq . The question embedding vector generation unit 21 generates, for example, a bidirectional vector sequence 24 from the feature vector sequence W _q of the question sentence, and performs max pooling processing that extracts and accumulates the maximum value of each element of the bidirectional vector sequence 24. A question embedding vector O _q (question intermediate vector) is generated.

The correct embedded vector generation unit 22 generates a feature vector obtained by converting a correct sentence for each word based on a time-series forward and reverse bidirectional word arrangement order (a correct feature vector group 25). ) To learn the word sequence bidirectionally to generate the correct embedded vector O _{a +} . The correct embedded vector generation unit 22 generates, for example, _a bidirectional vector sequence 25 from the feature vector sequence W _{a + of} the correct sentence, and extracts and accumulates the maximum value of each element of the bidirectional vector sequence 25 by a max pooling process. A correct answer embedding vector O _{a +} (correct answer intermediate vector) is generated.

The incorrect answer embedding vector generating unit 23 generates a feature vector obtained by converting the incorrect answer sentence for each word based on the time-series forward and reverse bidirectional word arrangement order (an incorrect vector string 26 (incorrect answer). Based on the feature vector group), the word sequence is learned bidirectionally to generate the correct embedded vector O _a− . The incorrect answer embedded vector generation unit 23 generates, for example, a bidirectional vector string 26 from the feature vector string W _a− of the incorrect answer sentence, and extracts and accumulates the maximum value of each element of the bidirectional vector string 26. An incorrect solution embedding vector O _a− (incorrect solution intermediate vector) is generated by the pooling process.

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

Here, in the basic LSTM architecture, there are three gates (input i _t , forget f _t , output o _t ) and a cell memory vector c _t . Σ () is a sigmoid function. W _i , W _f , W _o , W _c , U _i , U _f , U _o , U _c , b _i , b _f , b _o , and b _c are network parameters to be learned.

The QA-LSTM unit 20-1 is a biLSTM for “conclusion”, and based on a feature vector sequence (W _q , W _{ac +} , W _ac− ), a question embedding vector O _qc , a correct embedding vector O _{ac +} , A correct answer embedding vector _Oac- is generated. 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. Question embedding vector generation unit 21-1 generates a bidirectional vector sequence 24-1 from the feature vector sequence _{W q,} generates the vector _{O qc} embedded questions by Max pooling process. In addition, the correct embedded vector generation unit 22-1 generates a bidirectional vector sequence 25-1 from the feature vector sequence W _{ac +} and generates a correct embedded vector O _{ac +} by a max pooling process. Also, incorrect embedding vector generation unit 23-1 generates a bidirectional vector sequence 26-1 from the feature vector sequence _{W ac-,} generates the correct embedded vector _{O ac-by} Max pooling process.

The QA-LSTM unit 20-2 is a biLSTM for “supplement”, and based on a feature vector sequence (W _q , W _{as +} , W _as− ), a question embedding vector O _qs , a correct embedding vector O _{as +} , A correct answer embedding vector O _as− is generated. 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 embedding vector generation unit 21-2 generates a bidirectional vector sequence 24-2 from the feature vector sequence W _q, and generates a question embedding vector O _qs by a max pooling process. The correct embedded vector generation unit 22-2 generates a bidirectional vector sequence 25-2 from the feature vector sequence W _{as +} and generates a correct embedded vector O _{as +} by the max pooling process. The incorrect answer embedded vector generation unit 23-2 generates a bidirectional vector string 26-2 from the feature vector string W _as− and generates an incorrect answer embedded vector O _as− by the max pooling process.

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

Here, t is a time step, and W _sm , W _cm , and w _mb are attention parameters. In ^addition, ~ _h s (t) shows a bi-directional vector after update. Note that the superscript “˜” in the text represents a symbol attached immediately above the character.

Further, the question embedding vector O _qc , the correct answer embedding vector O _{ac +} and the incorrect answer embedding vector O _ac− generated by the QA-LSTM unit 20-1, and the question embedding vector O _qs generated by the QA-LSTM unit 20-2, Based on the correct answer embedding vector O _{as +} and the incorrect answer embedding vector O _as− , the loss function Ls using the cosine similarity is expressed by, for example, the following formula (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 a combination of “conclusion”, “supplement”, “correct answer”, and “incorrect answer”.

Here, [y, z] indicates a combination of the vector y and the vector z. O _q is [O _qc , O _qs ]. M represents 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 calculates the loss function Lw by the following equation (5) by combining the loss function L of equation (1) and the loss function Ls of equation (4).

Here, α is a parameter for controlling weighting by 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 generates a learning result by optimization using the loss function Lw calculated using the above-described configuration, and stores the generated learning result in the learning result storage unit 122. The learning processing unit 132 sets the “conclusion” parameter set {W _i , W _f , W _o , W _c , U _i , U _f , U _o , U _c , b _i , b _{f in} Equation (2) described above. , B _o , b _c } _c and “supplement” parameter set {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 attention parameters {W _sm , W _cm , w _mb } are generated.

In the above-described example, the example in which the scenario of the answer sentence is configured by the two partial items “conclusion” and “supplement” has been described. However, the scenario may be configured by two or more partial items. In this 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 embedded vector O _{a +} corresponding to each of the plurality of partial items, and each of the plurality of partial items. The learning is optimized by the loss function Lw calculated based on the combination of the plurality of partial items with the incorrect answer 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 apparatus 1 will be described with reference to FIG.

FIG. 6 is a flowchart illustrating 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). Note that the variable n counts the number of learning repetitions. In this example, a case where learning is repeated NN times will be described.

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

  Next, the learning processing unit 132 encodes the question sentence q (step S103). The learning processing unit 132 encodes the obtained 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 the biLSTM and the max pooling process 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 (partial item vectors) prepared in advance to learn the conclusion answer sentence ac and the supplementary answer sentence as. In the decoder model M112 of the encoder decoder model M11, 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, and the target type (partial item) answer sentence (conclusion answer) Decoding so as to become a sentence ac and a supplemental 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 the same in that it is a context vector in the C-LSTM model. Thereby, 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 “conclusion” and “supplement”. In step S104, the learning processing unit 132 learns the answer sentences “conclusion” and “supplement” one character (one word) at a time.

  Next, the learning processing unit 132 generates a conclusion answer sentence ac2 and a supplementary answer sentence as2 (step S105). As the input for predicting the next output in the above-described encoder / decoder learning process in which the learning processing unit 132 has learned for each pair information of the question sentence q, the conclusion answer sentence ac, and the supplementary answer sentence as. The predicted output word is 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 has advanced to EOS (end of sequence; the last character of the conclusion answer sentence ac or the supplementary answer sentence as), the update parameter ( This is a process of temporarily generating a conclusion answer sentence ac2 and a supplementary answer sentence as2 according to the learning result on the way. 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 determines that the correct sentence ac2 + for “conclusion”, the incorrect answer sentence ac2- for “conclusion”, the correct sentence as2 + for “supplement”, and the incorrect answer for “supplement”. A sentence as2- is generated.

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

(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 an incorrect answer embedding vector _Oac- is generated. In addition, the QA-LSTM unit 20-1 includes the t-th bidirectional vectors (h _qc (t), h _{ac +} (t), h of the bidirectional vector sequence (24-1, 25-1, 26-1). _ac- (t)) is subjected to a max pooling process to _generate a question embedding vector O _qc , a correct answer embedding vector O _{ac +} , and an incorrect answer embedding vector O _ac- .

(C) Next, the QA-LSTM unit 20-2 of the learning processing unit 132 first generates a question embedding vector O _qs . The QA-LSTM unit 20-2 performs max pooling processing on each t-th bidirectional vector h _qs (t) of the bidirectional vector sequence 24-2, and generates a question embedded vector O _qs .

(D) Next, the QA-LSTM unit 20-2 calculates each of the t-th bidirectional vectors (h _{as +} (t), h _as- (t)) of the bidirectional vector sequence (25-2, _26-2 ). _Is updated using the correct answer embedding vector O _{ac +} and the incorrect answer embedding vector O _ac− . That, QA-LSTM section 20-2, using Equation (2) above, each bidirectional vector _{(h as + (t),} h as- (t)) to update the update vector ^(~ _{h the as +} (T), ^~ _has- (t)).

(E) Next, QA-LSTM unit 20-2, the update vector ^{_{(~ h as + (t)}} , ~ h as- (t)) of each processing Max pooling, correct embedded vector _{O the as +,} and incorrect answers An embedding vector O _as- is generated.
(F) Then, the loss function generation unit 30 of the learning processing unit 132 _applies the generated embedding vectors (O _qc , O _{ac +} , O _ac− , O _qs , O _{as +} , O _as− ) and Expression (5). Based on this, a 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 with the calculated loss function Lw. For example, the learning processing unit 132 sets a parameter set for “conclusion” {W _i , W _f , W _o , W _c , U _i , U _f , U _o , U _c , b _i , b _f , b _o , b _c } _c and a “supplementary” parameter set {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 the optimized parameters in the learning result storage unit 122 as learning results.

  Next, the learning processing unit 132 determines whether or not the input information is complete (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 (a correct sentence ac + of “conclusion”, an incorrect sentence ac− of “conclusion”, a correct sentence as + of “supplement”, And it is determined whether there is next set information with an incorrect sentence as-) of “supplement”. When the input information is the end (step S108: YES), the learning processing unit 132 proceeds with the process to step S109. If the input information is not finished (there is next input information) (step S108: YES), the learning processing unit 132 returns the process to step S102.

  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 learning has been repeated NN times from step S102 to step S108. The learning processing unit 132 ends the learning process when the variable n is equal to or greater than the repetition count NN (step S109: YES). Moreover, the learning process part 132 advances a process to step S110, when the variable n is less than the repetition frequency NN (step S109: NO).

In step S110, the learning processing unit 132 assigns “n + 1” to the variable n and returns the process to step S102. That is, the learning processing unit 132 adds “1” to the value of the variable n, and returns the process to step S102.
As described above, the learning processing unit 132 repeatedly learns the processing from step S102 to step S108 NN times, and stores the learning result in the learning result storage unit 122.

<Answer generation process>
Next, processing for generating an answer sentence from a question sentence of the information processing apparatus 1 according to the present embodiment will be described with reference to the drawings.
FIG. 7 is a flowchart illustrating an example of processing for generating an answer sentence from a question sentence of the information processing apparatus 1 according to 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 apparatus 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 is obtained by using the above-described encoder decoder model M11 that learns in character units (word units) and a sentence unit learning model M12 that is a sentence bi-sentence model that learns for each sentence. Learned in combination. That is, for example, the answer generation unit 134 generates an answer sentence from the input question sentence based on the learning result learned by the learning process shown in FIG.

  Note that 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 sentence path. Further, the answer generation unit 134 does not simply select the “conclusion” and “supplement” partial answer sentences from the existing answer sentences by combining the encoder decoder model M11 that learns in character units (word units). The partial answer sentence is selected in units of words, and a new answer sentence is generated.

  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 causes the service storage unit 121 to store the answer sentence as a post of an answer to the input question sentence. Thereby, it is possible to view the answer sentence generated by the answer generation unit 134 with respect to the question sentence from the terminal device 2 connected to the information processing apparatus 1 via the network NW1. After the process of step S203, the information processing apparatus 1 ends the process of generating an answer sentence.

  Next, with reference to FIG. 8, the evaluation result of the answer 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” indicates an evaluation result when only the loss function Ls according to the expression (4) by the sentence unit learning model M12 which is a sentence bi-sentence model is used for comparison. “C-LSTM” is an evaluation result when the loss function L according to the equation (1) based on the C-LSTM model described above is used.

The “method of this embodiment” is a method in the case of using the loss function Ls according to the equation (5) by the learning processing unit 132 of this embodiment described above.
The evaluation results “ROUGE-1” (uni-gram), “ROUGE-2” (bigram), and “ROUGE-L” (Longest common subsequence) are 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. Yes. For details 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 includes 5000 sets of question texts accumulated in 16 categories including “Love Consultation”, “Travel”, “Medical”, etc. in the Q & A service “Teach me goo” Answer text is used. The partial items are two cases of “conclusion” and “supplement”.

  As test data for evaluation, two sets of 100 sets of a question sentence q, a conclusion answer sentence ac, and a supplementary answer sentence as are prepared, and the evaluation result shows an average value. In this evaluation, the number of embedded characters is “300”, α in equation (5) is “0.5”, and the number of repetitions NN is “500”.

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

FIG. 9 is a diagram illustrating an example of an answer sentence generated by the information processing apparatus 1 according to the present embodiment.
In FIG. 9, “An answer sentence generated by C-LSTM” indicates an answer sentence when the above-described C-LSTM model is used for comparison.
As shown in Fig. 9, in "An answer sentence generated by C-LSTM", I think that there is a person that "I can't think of the other party" with respect to the question sentence. It is an unnatural answer like “I think it ’s good.”

  On the other hand, the response sentence generated by the information processing apparatus 1 according to the present embodiment is “I think it's enough to give your courage and convey your feelings. It is possible to generate a natural sentence answer with reduced discomfort.

  As described above, the information processing apparatus 1 according to the present embodiment includes the question acquisition unit 133 and the answer generation unit 134. The question acquisition unit 133 acquires the input question sentence that has been input. The answer generation unit 134 uses a machine based on learning information having a plurality of pairs 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 result, an answer sentence for the input question sentence acquired by the question acquisition unit 133 is generated. Here, the learning result is optimized and learned by a 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 encoding the question sentence one word at a time based on the word arrangement order, and based on the generated context vector 10, a known partial answer for each of a plurality of partial items is generated. Decode and learn sentences.

The sentence unit learning model M12 uses, as input information, combination 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. The sentence unit learning model M12 includes a question embedding vector O _q (question intermediate vector), an answer intermediate vector (correct answer embedding vector (O _{ac +} , O _{as +} ) and an incorrect answer embedding vector (O _ac− ) corresponding to each of a plurality of partial items. , O _as− )), and learning based on a combination of a plurality of partial items. Here, the question embedding vector O _q (question intermediate vector) is a feature vector obtained by converting the question sentence of the input information for each word based on the time-series forward and reverse bidirectional word arrangement order. Based on the bidirectional vector sequence 24 (question feature vector group), it is generated by learning the word sequence bidirectionally. The answer intermediate vectors (correct answer embedding vectors (O _{ac +} , O _{as +} ) and incorrect answer embedding vectors (O _ac− , O _as− )) are bi-directional words obtained by converting partial answer sentences for each word. Based on the answer feature vector group (bidirectional vector string 25 and bidirectional vector string 26) generated based on the arrangement order of the words, it is generated by bi-directionally learning the word arrangement.

  Thereby, the information processing apparatus 1 according to the present embodiment generates an answer sentence based on a learning result obtained by collectively learning the encoder / decoder model M11 and the sentence unit learning model M12. Can generate an answer sentence, and the answer sentence of each partial item selected by optimizing the connection of the answers of the partial items can be combined to create a new answer sentence. Therefore, the information processing apparatus 1 according to the present embodiment can generate a natural sentence answer with reduced discomfort with respect to the question. That is, the information processing apparatus 1 according to the present embodiment has a natural word (word) or sentence order, and can generate a more complex 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, complicated, and various non-factoroid questions.

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

In this embodiment, the encoder / decoder model M11 decodes and learns a known partial answer sentence based on topic information related to each word in the known partial answer sentence. That is, the encoder decoder model M11 is a model that learns by C-LSTM, for example.
Thereby, since the information processing apparatus 1 according to the present embodiment can reduce the selection of a less relevant character (word) based on the topic information, it is natural to further reduce the sense of discomfort for each character (word). A written answer can be generated.

In this embodiment, the known answer sentence includes a correct answer sentence and an incorrect answer sentence for the question sentence. The answer generation unit 134 generates an answer sentence based on a learning result that is machine-learned based on learning information that includes a plurality of pairs of a question sentence and a correct answer sentence and an incorrect answer sentence corresponding to each of the plurality of partial items. . Based on the context vector 10, the encoder / decoder model M11 decodes and learns the correct and incorrect answer sentences for each of the partial items. The sentence unit learning model M12 includes 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 the plurality of partial items, and a plurality of partial items. Learning is performed based on a combination of a plurality of partial items of a corresponding incorrect answer embedding vector (O _ac− , O _as− ) (incorrect answer intermediate vector). Here, the correct embedded vector (O _{ac +} , O _{as +} ) is a bidirectional vector sequence 25 (correct feature vector group) generated based on a bidirectional word arrangement order based on a feature vector obtained by converting a correct sentence for each word. Based on this, it is generated by bi-directional learning of word sequences. The incorrect answer embedding vector (O _ac− , O _as− ) is a bidirectional vector string 26 (incorrect answer feature) generated based on a bidirectional word arrangement order from a feature vector obtained by converting an incorrect answer sentence for each word. Based on the vector group), the word sequence is generated by bidirectional learning.

  Thereby, since the information processing apparatus 1 according to the present embodiment generates an answer sentence based on the learning result learned using the correct answer sentence and the incorrect answer sentence, a natural sentence answer with further reduced discomfort can be obtained. Can be generated.

Further, in the present embodiment, the learning result includes correct answer embedding vectors (O _{ac +} , O _{as +} ) and incorrect answer embedding vectors (O O ₊ ) corresponding to a first partial item (for example, “conclusion”) among a plurality of partial items. _ac− , O _as− ), the bidirectional vector sequence 25 and the bidirectional vector sequence 26 (incorrect feature vector group) corresponding to a second partial item (for example, “supplement”) different from the first partial item. ) 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-described equation (3). To learn.
Thereby, the information processing apparatus 1 according to the present embodiment can perform learning that optimizes the relationship between the partial items (for example, the connection of the partial items). Therefore, the information processing apparatus 1 according to the present embodiment can generate a more natural answer sentence by combining partial items.

Further, the information processing apparatus 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.
Thereby, the information processing apparatus 1 according to the present embodiment can learn by itself and generate a learning result. In addition, the information processing apparatus 1 according to the present embodiment can improve the answer to the question, for example, by re-learning.

  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 partial items, the information processing apparatus 1 according to the present embodiment generates partial items suitable for generating an answer sentence. can do.

Note that the learning processing unit 132 may re-learn the learning result based on a predetermined condition (for example, periodically or based on an evaluation value (for example, ROUGE) being reduced to a predetermined value or less). .
Thereby, the information processing apparatus 1 according to the present embodiment can improve the answer to the question in response to a change in time.

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 sentence that has been input. In the answer generation step, the answer generation unit 134 has a plurality of combinations of the question sentence and the correct answer sentence and the incorrect answer sentence corresponding to each of the plurality of partial items divided in the answer sentence according to a predetermined sentence path. Based on the learning information, an answer sentence to the input question sentence acquired by the question acquisition step is generated based on the above-described learning result that has been machine-learned.
Thereby, the information processing method according to the present embodiment produces the same effect as that of the information processing apparatus 1 described above, and can generate a natural sentence answer with reduced discomfort for the question.

In addition, this invention is not limited to said embodiment, It can change in the range which does not deviate from the meaning of this invention.
For example, in the above-described embodiment, the information processing apparatus 1 has been described with an example including the learning processing unit 132. However, the present invention is not limited to this, and the learning processing unit 132 is not limited to this, as long as the learning result can be acquired. It does not have to be provided.

Moreover, although the information processing apparatus 1 demonstrated the example provided with the service memory | storage part 121 and the learning result memory | storage part 122, either one or both of the service memory | storage part 121 and the learning result memory | storage part 122, for example, An external server device may be provided. Further, in the information processing apparatus 1, an external server device may include a part of the functional units included in the control unit 13.
In the above embodiment, the example in which the information processing apparatus 1 is configured by one server apparatus has been described. However, the information processing apparatus 1 may be configured by a plurality of apparatuses.

Further, in the above-described embodiment, the information processing apparatus 1 has described an example in which the answer sentence is configured by two partial items of “conclusion” and “supplement”, but is not limited thereto. You may make it respond | correspond to two or more partial items.
Moreover, in the sentence unit learning model M12 of the above-described embodiment, the information processing apparatus 1 has described the example in which the method including the QA-LSTM unit 20 for each partial item and the method based on the attention mechanism are applied. It is not limited. In the sentence unit learning model M12, for example, the information processing apparatus 1 may have a form in which some of these methods are not applied.

  In the above embodiment, the example has been described in which the loss function Ls of the sentence unit learning model M12 is calculated by the above-described equation (4). However, the present invention is not limited to this, and the following equation (6) May be calculated.

  In the learning process of the above embodiment, the reply sentence (for example, the conclusion answer sentence ac2, the supplementary answer sentence as2, etc.) regenerated based on the intermediate learning result of the encoder decoder model M11 and the context vector 10 is used as the sentence unit learning model. Although the example which uses M12 as input information has been described, the present invention is not limited to this. The learning processing unit 132 may use, as input information of the sentence unit learning model M12, the answer sentence before re-generation used for decoding learning.

Each configuration included in the information processing apparatus 1 described above has a computer system therein. And the program for implement | achieving the function of each structure with which the information processing apparatus 1 mentioned above is provided is recorded on a computer-readable recording medium, and the program recorded on this recording medium is read into a computer system and executed. 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 “computer system” here includes an OS and hardware such as 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 dedicated line. The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. As described above, the recording medium storing the program may be a non-transitory recording medium such as a CD-ROM.

  The recording medium also includes a recording medium provided inside or outside that is accessible from the distribution server in order to distribute the program. It should be noted that the program may be divided into a plurality of parts and downloaded at different timings, and the structure combined with each component included in the information processing apparatus 1 or the distribution server that distributes each of the divided programs may be different. Furthermore, the “computer-readable recording medium” holds a program for a certain period of time, such as a volatile memory (RAM) inside a computer system that becomes a server or a client when the program is transmitted via a network. Including things. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  Moreover, you may implement | achieve part or all of the function mentioned above as integrated circuits, such as LSI (Large Scale Integration). Each function described above may be individually made into a processor, or a part or all of them may be integrated into a processor. Further, the method of circuit integration is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. In addition, when an integrated circuit technology that replaces LSI appears due to the advancement of semiconductor technology, an integrated circuit based on the technology may be used.

DESCRIPTION OF SYMBOLS 1 Information processing apparatus 2 Terminal apparatus 10 Context vector 11 NW communication part 12 Storage part 13 Control part 20, 20-1, 20-2 QA-LSTM part 21, 21-1, 21-2 Question embedding vector generation part 22, 22 -1, 22-2 Correct answer embedded vector generator 23, 23-1, 23-2 Incorrect answer 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 providing 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 The

Claims (8)

A question acquisition unit for acquiring the input question text,
In a learning result obtained by machine learning based on learning information having a plurality of pairs 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 And an answer generation unit that generates an answer sentence for the input question sentence acquired by the question acquisition part,
The learning result is
The question sentence is sequentially encoded word by word based on the word arrangement order 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. An encoder / decoder model to
Using as input information set information including a partial answer sentence for each of the plurality of partial items decoded based on the encoder decoder model and the question sentence, a feature vector obtained by converting the question sentence into words is sometimes used. A plurality of question intermediate vectors generated by learning the word sequence bidirectionally based on a question feature vector group generated based on the sequence of the words in both the forward and reverse directions of the sequence; An answer intermediate vector corresponding to each of the partial items, and based on the answer feature vector group generated based on the bidirectional order of the words, the feature vector obtained by converting the partial answer sentence for each word, A sentence unit learning model for learning based on a combination of the plurality of partial items of an answer intermediate vector generated by learning the word sequence in the bidirectional direction;
An information processing apparatus characterized by being optimized and learned by a loss function calculated by combining. The sentence unit learning model uses an answer sentence generated based on an intermediate learning result of the encoder decoder model and the context vector as the partial answer sentence, and sets information including the partial answer sentence and the question sentence as the input information The information processing apparatus according to claim 1, wherein learning is performed as follows. The said encoder decoder model decodes and learns the said known partial answer sentence based on the topic information relevant for every word in the said known partial answer sentence. The claim 1 or Claim 2 characterized by the above-mentioned. Information processing device. The known answer sentence includes a correct answer sentence and an incorrect answer sentence for the question sentence,
The answer generator is based on the learning result machine-learned based on the learning information including a plurality of sets of the question sentence and the correct answer sentence and the incorrect answer sentence corresponding to each of the plurality of partial 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,
The sentence-by-sentence learning model includes the question intermediate vector and a correct answer intermediate vector corresponding to each of the plurality of partial items, and a feature vector obtained by converting the correct answer sentence for each word is used as the bidirectional arrangement of the words. A correct answer intermediate vector generated by learning the word sequence bidirectionally based on a correct answer feature vector group generated based on the order, and an incorrect answer intermediate vector corresponding to each of the plurality of partial items, The incorrect answer sentence is generated by learning the word sequence bidirectionally based on the incorrect feature vector group generated based on the bidirectional order of the word order. The information processing according to any one of claims 1 to 3, wherein learning is performed based on a combination of the plurality of partial items with an incorrect answer intermediate vector. Science device. The learning result is
Correct feature vector group corresponding to a second partial item different from the first partial item based on the correct intermediate vector and the incorrect intermediate vector corresponding to the first partial item of the plurality of partial 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. A question acquisition step in which the question acquisition unit acquires the inputted input question sentence;
The answer generation unit performs machine learning based on learning information including a plurality of pairs 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. An answer generation step of generating an answer sentence for the input question sentence acquired by the question acquisition step based on the learned result,
The learning result is
The question sentence is sequentially encoded word by word based on the word arrangement order 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. An encoder / decoder model to
Using as input information set information including a partial answer sentence for each of the plurality of partial items decoded based on the encoder decoder model and the question sentence, a feature vector obtained by converting the question sentence into words is sometimes used. A plurality of question intermediate vectors generated by learning the word sequence bidirectionally based on a question feature vector group generated based on the sequence of the words in both the forward and reverse directions of the sequence; An answer intermediate vector corresponding to each of the partial items, and based on the answer feature vector group generated based on the bidirectional order of the words, the feature vector obtained by converting the partial answer sentence for each word, A sentence unit learning model for learning based on a combination of the plurality of partial items of an answer intermediate vector generated by learning the word sequence in the bidirectional direction;
An information processing method characterized by being optimized and learned by a loss function calculated by combining. On the computer,
A question acquisition step for acquiring the input question sentence,
In a learning result obtained by machine learning based on learning information having a plurality of pairs 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 And an answer generation step of generating an answer sentence for 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 word arrangement order 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. An encoder / decoder model to
Using as input information set information including a partial answer sentence for each of the plurality of partial items decoded based on the encoder decoder model and the question sentence, a feature vector obtained by converting the question sentence into words is sometimes used. A plurality of question intermediate vectors generated by learning the word sequence bidirectionally based on a question feature vector group generated based on the sequence of the words in both the forward and reverse directions of the sequence; An answer intermediate vector corresponding to each of the partial items, and based on the answer feature vector group generated based on the bidirectional order of the words, the feature vector obtained by converting the partial answer sentence for each word, A sentence unit learning model for learning based on a combination of the plurality of partial items of an answer intermediate vector generated by learning the word sequence in the bidirectional direction;
A program characterized by being optimized and learned by a loss function calculated by combining.

Images