JP6680659B2 - Information processing device, information processing method, and program - Google Patents

Description

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

  In recent years, there is known a technique of outputting an answer to an input question by using machine learning (for example, deep learning method) (for example, see Non-Patent Document 1). In the information processing apparatus using such a conventional technique, an answer estimated to be appropriate is selected and output from existing 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-described conventional information processing device, for example, when the question is a Non-Factoid type question that asks for an answer based on the explanation of the reason or phenomenon, the answer becomes a complicated long sentence, Since the prepared existing answer is output, it is difficult to generate a new answer. Therefore, in the above-described conventional information processing apparatus, an unnatural answer with a sense of discomfort may be output in response to a question.

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

  In order to solve the above problems, one embodiment of the present invention is a question acquisition unit that acquires an input input question sentence, a question sentence, and a plurality of answer sentences that are divided by a predetermined sentence path. Based on learning information having a plurality of pairs of correct and incorrect sentences corresponding to each partial item, based on the learning result machine-learned, the answer sentence to the input question sentence acquired by the question acquisition unit, An answer generation unit for generating, the learning result is a question feature generated based on the arrangement order of the words in the forward and backward bidirectional time series of the feature vector obtained by converting the question sentence for each word. A question intermediate vector generated by learning the word sequence bidirectionally based on a vector group and a correct answer intermediate vector corresponding to each of the plurality of partial items, wherein the correct answer sentence is Based on a correct feature vector group generated by converting the feature vector for each word based on the arrangement order of the bidirectional words, a correct intermediate vector generated by learning the arrangement of the words bidirectionally, An incorrect answer intermediate vector corresponding to each of the plurality of sub-items, wherein a feature vector obtained by converting the incorrect answer sentence into words is generated in an incorrect answer feature vector group based on the arrangement order of the bidirectional words. Based on the loss function calculated based on the combination of the plurality of partial items with the incorrect intermediate vector generated by learning the word sequence bidirectionally, and learning is performed. Is an information processing device.

  Further, according to an aspect of the present invention, in the information processing device, the learning result is learned by averaging a feature vector of a word included in the question sentence and a feature vector of a term related to the question sentence. It is characterized in that learning is performed based on the question intermediate vector, the correct answer intermediate vector, and the incorrect answer intermediate vector generated based on the additional vector.

  Further, according to an aspect of the present invention, in the information processing device, the question feature vector group is converted based on the question sentence and the additional vector, and the correct answer feature vector group is the correct answer sentence and the addition sentence. The incorrect answer feature vector group is converted based on the vector and the incorrect answer sentence and the additional vector.

  Further, according to an aspect of the present invention, in the information processing device, the learning result is based on the correct answer intermediate vector and the incorrect answer intermediate vector corresponding to a first partial item of the plurality of partial items. , And the correct answer feature vector group and the incorrect answer feature vector group corresponding to the second partial item different from the first partial item are updated and learned.

  Further, according to an aspect of the present invention, in the above information processing device, a learning processing unit that performs machine learning based on the learning information and generates the learning result is characterized.

  Further, according to one aspect of the present invention, the question acquisition unit obtains the input input question sentence, the question acquisition step, and the answer generation unit divides the question sentence and the answer sentence according to a predetermined sentence path. Based on learning information having a plurality of correct answer sentences and incorrect answer sentences corresponding to each of the plurality of partial items, based on the learning result machine-learned, the input question sentence acquired by the question acquisition step And an answer generation step of generating an answer sentence for the learning result, based on the arrangement sequence of the words in the forward and backward bidirectional time series of the feature vector obtained by converting the question sentence into words. A question intermediate vector generated by bidirectionally learning the sequence of words based on the generated question feature vector group, and a correct answer intermediate vector corresponding to each of the plurality of partial items. Based on a group of correct feature vectors generated from the correct answer sentence for each word based on the arrangement order of the bidirectional words, the arrangement of the words is learned bidirectionally. A correct answer intermediate vector generated by the above, and an incorrect answer intermediate vector corresponding to each of the plurality of partial items, wherein a characteristic vector obtained by converting the incorrect answer sentence for each word is based on the arrangement order of the bidirectional words. Optimal by the loss function calculated based on the combination of the plurality of partial items with the incorrect answer intermediate vector generated by learning the word sequence bidirectionally based on the incorrect answer feature vector group generated It is an information processing method characterized by being digitized and learned.

  Further, according to an aspect of the present invention, a question acquisition unit that causes a computer to acquire an input input question sentence, and an answer generation unit include a question sentence and a predetermined sentence in the answer sentence. Based on learning information having a plurality of pairs of correct and incorrect sentences corresponding to each of a plurality of partial items divided by a way, based on a learning result machine-learned, the acquired by the question acquisition step A program for executing an answer generation step of generating an answer sentence for an input question sentence, wherein the learning result is a feature vector obtained by converting the question sentence for each word in both forward and backward chronological order. Based on a question feature vector group generated based on the order of arrangement of the words, the question intermediate vector generated by learning the arrangement of the words bidirectionally, A correct answer intermediate vector corresponding to each of a number of sub-items, based on a correct answer feature vector group generated based on the arrangement order of the bidirectional words, the feature vector obtained by converting the correct answer sentence for each word, A correct answer intermediate vector generated by learning the sequence of words bidirectionally, and an incorrect answer intermediate vector corresponding to each of the plurality of partial items, wherein the incorrect answer sentence is converted into a feature vector for each word. A combination of the plurality of sub-items with an incorrect answer intermediate vector generated by learning the word sequence bidirectionally based on an incorrect answer feature vector group generated based on the bidirectional word arrangement order. The program is characterized by being optimized and learned by a loss function calculated based on.

  According to the present invention, it is possible to generate a natural 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 a learning processing part and learning processing in this embodiment. It is a figure explaining an example of the learning method of the word vector biased by the semantics in this embodiment. It is a flow chart which shows an example of processing which generates an answer sentence from a question sentence of an information processor in this embodiment. It is a figure explaining an example of the concept of generation of a reply sentence in this embodiment. It is a figure which shows comparison with the reply generation system in this embodiment, and a prior art.

  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 showing an example of an information processing system 100 according to this 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 the network NW1.
The information processing system 100 displays posted questions and answers on the terminal device 2 connected to the information processing device 1, and provides an information service such as a Q & A service for 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 addition, in the example shown in FIG. 1, for simplification of description, an example in which one terminal device 2 is connected to the information processing device 1 is shown, 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, for example, a server device that provides an information service such as a Q & A service. The information processing device 1 registers, for example, a question sentence received from a user via the terminal device 2 in the Q & A service so that the question sentence can be browsed, and an answer sentence received from another user via the terminal device 2 is registered. And make it viewable. Further, the information processing device 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 the question sentence can be browsed. Further, the information processing device 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 the Internet, for example, and communicates various information via the network NW1. The NW communication unit 11 connects to the terminal device 2 that has made a connection request, for example, via 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 question sentences and answer sentences 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. 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 as a whole. The control unit 13 provides, for example, a process of providing an information service such as the Q & A service described above, 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. Executes various processes such as the municipal processing of answer sentences to sentences. The control unit 13 also 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. The service providing unit 131, for example, causes the service storage unit 121 to store the received question text and answer text as post information from the terminal device 2 via the NW communication unit 11. Further, 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 desiring to browse the Q & A service, for example. , On the terminal device 2. Further, the service providing unit 131 causes the service storage unit 121 to store the answer sentence generated by the answer generating unit 134 described below. In the Q & A service, the service providing unit 131 is supposed to provide information to the user by category (classification) such as “love”, “family”, and “cooking”.

The learning processing unit 132 has a plurality of sets of a question sentence and a correct sentence and an incorrect sentence corresponding to each of a plurality of partial items in the answer sentence divided according to a predetermined sentence scenario (scenario). Machine learning is performed to generate a learning result. Here, for example, when the scenario (scenario) of the sentence of the answer sentence is set in the order of “conclusion” and “supplementary”, the partial items are “conclusion” and “supplementary”, and When the scenario (scenario) is defined in the order of “chat”, “case”, and “conclusion”, the partial items are “chat”, “case”, and “conclusion”.
In the following description, the case of “conclusion” and “supplement” will be described as an example of a plurality of partial items.

  In addition, the learning processing unit 132 generates a learning result by using the deep learning technology by using, for example, the Q & A service posting information (question sentence and answer sentence) stored in the service storage unit 121 as input information. To do. The learning processing unit 132 uses, as the learning information, a set of a question sentence, a correct answer sentence, and an incorrect answer sentence arbitrarily extracted from the answer sentence other than the correct sentence, as input information (learning information) for the learning process. . The learning processing unit 132 stores the generated learning result in the learning result storage unit 122. Here, the configuration and learning process of the learning processing unit 132 according to the present embodiment will be described with reference to FIG. 2.

FIG. 2 is a diagram illustrating an example of the learning processing unit 132 and learning processing according to this embodiment.
As shown in FIG. 2, the learning processing unit 132 includes a vector conversion unit 10, a QA-LSTM (Question Answering-Long Short-Term Memory) unit (20-1, 20-2), and a loss function generating unit 30. Is equipped with.

  The vector conversion unit 10 converts each of the question sentence, the correct answer sentence, and the incorrect answer sentence included in the set of learning information into a feature vector. The correct sentence and the incorrect sentence are, for example, divided into “conclusion” and “supplementary” sub-items to generate feature vectors. In addition, the vector conversion unit 10 pre-learns a word vector by biasing it with semantics based on the learning information stored in the service storage unit 121, for example. For example, in the Q & A service of the “love” category, words (tokens) such as “distance”, “like”, “lover” are likely to be included, and learning to bias them is emphasized to emphasize these. To do. Specifically, the vector conversion unit 10 performs pre-learning by using a learning method as shown in FIG. 3, for example.

FIG. 3 is a diagram illustrating an example of a learning method of a word vector biased by the semantics in the present embodiment.
In the example shown in FIG. 3, when there is a target sentence arranged in the order of “word A”, “word B”, ... “Word F”, “word G”, the vector conversion unit 10 The category token and title token of the target sentence (for example, question sentence) and each of "word A", "word B", ... "Word F" are converted into word vectors and averaged, and the given context is given. In, learning is performed to predict the following word "word G". Here, the category token is a word (term) indicating the category (classification) of the target sentence, and is, for example, “love”, “family” or the like, which is the category of the Q & A service described above. The title token is defined by a noun extracted from the title assigned to the target sentence (for example, question sentence), and is, for example, “distance”, “like”, “lover”, etc. described above. Multiple title tokens may be extracted from the title while one category token is assigned to the question text, for example.

  The vector conversion unit 10 learns not only the word vector but also the category vector of the category token and the title vector of the category token. These additional vectors act as a semantic bias. In this way, the vector conversion unit 10 averages the feature vectors of the words included in the question sentence and the feature vectors of the terms related to the question sentence to obtain additional vectors (for example, a category vector and a title vector). Pre-learning is performed, and based on the additional vector, each of the question sentence, the correct answer sentence, and the incorrect answer sentence included in the set of learning information is converted into a feature vector for each word. Note that the vector conversion unit 10 stores the learning result that has been pre-learned in the learning result storage unit 122.

For example, in the example illustrated in FIG. 2, the vector conversion unit 10 converts the question sentence into a feature vector sequence W _q that is a set of feature vectors. Further, the vector conversion unit 10 converts the correct answer sentence of “conclusion” into a feature vector sequence W _{ac +} which is a set of feature vectors, and the incorrect answer sentence of “conclusion” is a feature vector sequence W which is a set of feature vectors. _Convert to _ac- . Further, the vector conversion unit 10 converts the “supplementary” correct answer sentence into a feature vector sequence W _{as +} that is a set of feature vectors, and the “supplementary” incorrect answer sentence is a feature vector sequence W that is a set of feature vectors. _Convert to _as- .

  The QA-LSTM units (20-1, 20-2) are bidirectional long short-term memories (biLSTMs) that are bidirectional learning neural networks. The QA-LSTM unit 20-1 is a biLSTM for “conclusion”, and the QA-LSTM unit 20-2 is a biLSTM for “supplementation”. Note that 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 no distinction is made, the QA-LSTM unit 20 will be described.

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 the sequence of forward and backward bidirectional words arranged in time series (a bidirectional vector sequence 24 (question feature vector group). ), The sequence of words is bidirectionally learned to generate the question embedding vector O _q . 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 extracts the maximum value of each element of the bidirectional vector sequence 24 and stores the max pooling (Max). The question embedding vector O _q (question intermediate vector) is generated by the (pooling) processing.

The correct answer embedded vector generation unit 22 generates a feature vector in which a correct sentence is converted for each word based on a time-series forward and backward bidirectional word arrangement order (a correct feature vector group 25). ), The word sequence is bidirectionally learned to generate the correct answer embedding vector O _{a +} . The correct-embedded vector generation unit 22 generates the bidirectional vector sequence 25 from the feature vector sequence W _{a + of} the correct sentence, extracts the maximum value of each element of the bidirectional vector sequence 25, and stores it by the max pooling process. The correct answer embedding vector O _{a +} (correct answer intermediate vector) is generated.

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

  Non-Patent Document 1 discloses the LSTM that is the basis of the QA-LSTM unit 20. In the basic LSTM, when learning, the time-series input is X = {x (1), x (2), ..., x (N)}, and x (t) is the t-th When the word feature vector 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 (1) every t hours. .

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. 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 learned network parameters.

QA-LSTM unit 20-1 is a biLSTM for "conclusion" feature vector sequence _{(W q, W ac +,} W ac-) based on the question embedded vector _{O qc,} correct embedded vector _{O ac +,} and not The correct embedded vector _Oac- is generated. The QA-LSTM section 20-1 includes a question embedded vector generation section 21-1, a correct answer embedded vector generation section 22-1, and an incorrect answer embedded vector generation section 23-1. The question embedding vector generation unit 21-1 generates the bidirectional vector sequence 24-1 from the feature vector sequence W _q, and generates the question embedding vector O _qc by the max pooling process. Further, the correct answer embedded vector generation unit 22-1 generates a bidirectional vector sequence 25-1 from the feature vector sequence W _{ac +} and generates a correct answer embedded vector O _{ac +} by the max pooling process. In addition, the incorrect solution embedded vector generation unit 23-1 generates the bidirectional vector sequence 26-1 from the feature vector sequence W _ac− and the correct answer embedded vector O _ac− by the max pooling process.

The QA-LSTM unit 20-2 is a biLSTM for “supplementation”, and based on the feature vector sequence (W _q , W _{as +} , W _as− ), the question embedding vector O _qs , the correct answer embedding vector O _{as +} , and the non- _correction vector. The correct embedded 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 the max pooling process. Further, the correct answer embedded vector generation unit 22-2 generates a bidirectional vector sequence 25-2 from the feature vector sequence W _{as +} and generates a correct answer embedded vector O _{as +} by the max pooling process. In addition, the incorrect solution embedded vector generation unit 23-2 generates the bidirectional vector sequence 26-2 from the feature vector sequence W _as− and generates the incorrect _solution embedded vector O _as− by the max pooling process.

Further, the QA-LSTM section 20-2 updates the bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2 by using the attention mechanism when learning. The QA-LSTM unit 20-2 corresponds to “supplement” based on, for example, the correct answer embedding vector O _{as +} and the incorrect answer embedding vector O _as− that correspond 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 calculates the bidirectional vector h _s (t), which is the internal vector of the bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2, by the following equation (2). Update.

Where t is the time step and W _sm , W _cm , and w _mb are attention parameters. Further, ^~ _hs (t) indicates the updated bidirectional vector. In addition, the superscript "~" in the text represents a symbol directly above the character.

The loss function generation unit 30 generates 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 generated by the QA-LSTM unit 20-2. The loss function L using the cosine similarity is generated based on the embedding vector O _qs , the correct embedding vector O _{as +} , and the incorrect answer embedding vector O _as− . The loss function generation unit 30 generates the loss function L based on a combination of “conclusion”, “supplement”, “correct answer”, and “incorrect answer”. The loss function generation unit 30 calculates the loss function L by, for example, the following equation (3). The loss function L is set so that the cosine value in each combination of the question and the answer during learning becomes maximum.

Here, [y, z] indicates a combination of 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 learning processing unit 132 optimizes the loss function L calculated using the configuration described above to generate a learning result, and stores the generated learning result in the learning result storage unit 122. The learning processing unit 132 uses the “conclusion” parameter set {W _i , W _f , W _o , W _c , U _i , U _f , U _o , U _c , b _i , b _f of the above-mentioned Expression (1). , B _o , b _c } _c and a parameter set for “supplementation” {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 an attention parameter {W _sm , W _cm , w _mb } are generated.

In the above example, an example in which the scenario of the answer sentence is composed of two partial items of “conclusion” and “supplement” has been described, but it may be composed of two or more partial 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 partial items, and each of the plurality of partial items. It is optimized and learned by the loss function L calculated based on the combination of the plurality of partial items with the incorrect intermediate vector O _a− .

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

  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. Here, as described above, the learning result includes a plurality of sets of a question sentence and a correct sentence and an incorrect sentence corresponding to each of a plurality of partial items in the answer sentence, which are divided by a predetermined passage of the sentence. It is a result of machine learning based on the learning information that it has, and is stored in the learning result storage unit 122. 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. 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 to the input question sentence.

Next, the operation of the information processing device 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.

  The learning processing unit 132 repeats the following procedures (1) to (6) N times to execute the learning process.

(1) The learning processing unit 132 first sets each set of learning information (question sentence q, correct sentence ac + of “conclusion”, incorrect sentence ac− of “conclusion”, correct sentence as + of “supplement”, “ Incorrect answer sentence "as-)" is acquired. Vector converter 10 of the learning processing unit 132, the assembled information feature vector sequence _{(W q, W ac +,} W ac-, W as +, W as-) into a.

(2) Next, the QA-LSTM unit 20-1 uses the feature vector sequence (W _q , W _{ac +} , W _ac− ) to query embedded vector O _qc , correct embedded vector O _{ac +} , and incorrect embedded vector. _Oac- is generated. The QA-LSTM unit 20-1 includes the t-th bidirectional vectors (h _qc (t), h _{ac +} (t), h _ac- of the bidirectional vector sequence (24-1, _25-1 , _26-1 ). (T)) is subjected to max pooling processing to _generate a question embedding vector O _qc , a correct answer embedding vector O _{ac +} , and an incorrect answer embedding vector O _ac− .

(3) Next, the QA-LSTM section 20-2 first generates the 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 to generate a question embedding vector O _qs .

(4) Next, the QA-LSTM unit 20-2 causes the t-th bidirectional vector (h _{as +} (t), h _as- (t)) of the bidirectional vector sequence (25-2, _26-2 ). _Are updated using the correct answer embedding vector O _{ac +} and the incorrect answer embedding vector O _ac− . That is, the QA-LSTM unit 20-2 updates each bidirectional vector (has ₊ (t), _has- (t)) using the above-described equation (2), and updates the updated vector ( ^~ has _+). (T), ^~ _has- (t)) are generated.

(5) Next, the QA-LSTM unit 20-2 performs max pooling processing on each of the update vectors ( ^{~ has} ₊ (t), ^~ _has- (t)) to obtain the correct answer embedding vector _{Oas +} and the incorrect answer. Generate the embedding vector O _as- .

(6) Next, the learning processing unit 132 calculates the loss function L based on the combination of the generated embedded vectors (O _qc , O _{ac +} , O _ac− , O _qs , O _{as +} , O _as− ). The loss function generation unit 30 of the learning processing unit 132 calculates the loss function L using, for example, Expression (3) described above. Then, the learning processing unit 132 optimizes each parameter by the calculated loss function L. The learning processing unit 132 may use, for example, the “conclusion” 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 } _c and the parameter set for “supplementation” {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 parameter {W _sm , W _cm , w _mb }.

  The learning processing unit 132 repeats the procedure of (1) to (6) described above N times to perform learning, and stores the learning result in the learning result storage unit 122.

<Response sentence generation process>
Next, with reference to the drawings, a process of generating an answer sentence from a question sentence of the information processing apparatus 1 according to the present embodiment will be described.
FIG. 4 is a flowchart showing an example of processing of generating an answer sentence from a question sentence of the information processing device 1 according to the present embodiment.

  As shown in FIG. 4, the information processing apparatus 1 first acquires a question sentence from the service storage unit 121 (step S101). The question acquisition unit 133 of the information processing device 1 acquires the input question text from the question texts 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 question sentence and the learning result stored in the learning result storage unit 122 (step S102). The answer generation unit 134 generates an answer sentence from the question sentence based on the learning result learned by the learning process shown in FIGS. 2 and 3, for example.

  Note that the answer generation unit 134 does not simply select an existing answer sentence, but generates a new answer sentence by appropriately combining partial items in consideration of the passage of the sentence. For example, the example illustrated in FIG. 5 is an example of generating an answer sentence when the partial items are “chat”, “case”, and “conclusion”.

  In FIG. 5, the answer generation unit 134 considers the appropriateness of the combination of “chat”, “case”, and “conclusion” (adequacy of connection of partial items) based on the learning result, for example, “chat” , Select "There are various things", and select "Because people repeat encounters and parting." Further, the answer generation unit 134 selects “time is resolved.” As the “conclusion”. Then, the answer generation unit 134 generates an answer sentence "There are various things. People repeat encounters and parting. Time is resolved." Since the answer generation unit 134 considers the appropriateness of the connection of the partial items in the learning result, the optimum one is not selected as each item of the partial items, and an answer sentence with no discomfort is generated. As described above, each item of the partial items is selected to generate the answer sentence.

  Next, the answer generation unit 134 of the information processing device 1 stores the answer sentence in the service storage unit 121 (step S103). 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 to the input question sentence. This allows the terminal device 2 to connect to the information processing device 1 via the network NW1 and browse the answer sentence generated by the answer generation unit 134 with respect to the question sentence. After the processing of step S103, the information processing device 1 ends the processing of generating the answer sentence.

  Next, with reference to FIG. 6, an evaluation result of the answer sentence generated by the information processing apparatus 1 according to the present embodiment will be described.

FIG. 6 is a diagram showing a comparison between the answer generation method according to the present embodiment and the conventional technique.
In FIG. 6, “QA-LSTM” indicates an evaluation result when the technique described in Non-Patent Document 1 is used for comparison. In "QA-LSTM", the answer sentence is not generated as in the present embodiment, and thus the selected partial items are simply combined and evaluated.
Further, “Sematic-LSTM” is an evaluation result when the bias is applied to “QA-LSTM” with the semantics according to the present embodiment. In this case as well, as in the case of "QA-LSTM", since the answer sentence is not generated, the selected partial items are simply combined and evaluated.

  Further, the “response generation method of the present embodiment” is a method of the learning processing unit 132, and “not attention” is used when the attention mechanism of the above-mentioned expression (2) is not used and “attention is used” when the attention mechanism is used. Is the evaluation result. In the “response generation method of the present embodiment”, the processing of applying the bias with the semantics according to the present embodiment described above is applied.

  Further, the evaluation result shows the average precision rate AP for the top K. The average precision rate AP is calculated by the following equation (4).

Here, N _j indicates the number of correct answers in the top j when the rank is K, and D indicates the number of all correct answers (paired with the question).

  In addition, the learning information used at the time of evaluation is the question and answer sentences accumulated in 16 categories including “love consultation”, “travel”, “medical care”, etc. in the Q & A service “Teach me google”. I'm using it. Further, the partial items are two cases of "conclusion" and "supplement".

As shown in FIG. 6, in both cases of “Sematic-LSTM” using the method according to the present embodiment and “response generation method of the present embodiment” (“no attention”, “with attention”), the conventional technique is used. The average precision rate AP is higher than that of "QA-LSTM". For example, the value of the average precision rate AP (K = 1) of the “response generation method of the present embodiment” (“attention”) is “0.3901”, and the average precision rate AP of the “QA-LSTM” ( It is about 20% higher than "0.3262" which is the value of K = 1).
Further, in the love consultation of the Q & A service “Teach me goo” described above, when the answer sentence generated by the information processing apparatus 1 according to the present embodiment is applied to 123 questions, 21 of the 123 questions are applied. Was selected by the questioner as the best answer.

As described above, the information processing device 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 input question text. The answer generation unit 134 is based on the learning information having a plurality of sets of question sentences and correct and incorrect sentences corresponding to each of a plurality of partial items in the answer sentence, which are divided according to a predetermined sentence path. An answer sentence to the input question sentence acquired by the question acquisition unit 133 is generated based on the learning result of the machine learning. Here, the learning result includes the question embedding vector O _q (question intermediate vector), the correct answer embedding vectors (O _{ac +} , O _{as +} ) (correct answer intermediate vector) corresponding to each of the plurality of partial items, and each of the plurality of partial items. Learning is performed by optimization with a loss function L calculated based on a combination of a plurality of partial items of the corresponding incorrect answer embedding vector (O _ac− , O _as− ) (incorrect answer intermediate vector). The question embedding vector O _q is 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 a sequence of forward and backward bidirectional words arranged in time series. Based on, it is generated by learning the sequence of words in both directions. The correct answer embedding vector (O _{ac +} , O _{as +} ) is based on a bidirectional vector sequence 25 (correct answer feature vector group) generated based on a bidirectional word arrangement sequence of feature vectors obtained by converting correct sentences for each word. It is generated by learning the sequence of words in both directions. The incorrect answer embedding vector (O _ac− , O _as− ) is a bidirectional vector sequence 26 (incorrect answer characteristic vector group) generated based on the bidirectional word arrangement order of feature vectors obtained by converting an incorrect answer sentence for each word. ), It is generated by learning the sequence of words in both directions.

  As a result, the information processing apparatus 1 according to the present embodiment creates a new answer sentence by combining the answer sentences of the selected partial items by optimizing the connection of the responses of the partial items based on the learning result. can do. Therefore, the information processing apparatus 1 according to the present embodiment can generate a natural text response with a reduced sense of discomfort for a question.

  Further, in the present embodiment, the loss function L is calculated based on, for example, the above equation (3). Since the loss function L optimizes the combination of the respective sub-items at the same time, the information processing apparatus 1 according to the present embodiment can select the sub-items suitable for generating the answer sentence.

Further, in the present embodiment, the learning result is an additional vector learned by averaging the feature vector of the word included in the question sentence and the feature vector of the term related to the question sentence by the learning processing unit 132 (for example, Learning is performed based on the question embedding vector O _q , the correct answer embedding vector (O _{ac +} , O _{as +} ), and the incorrect answer embedding vector (O _ac− , O _as− ) generated based on the category vector and the title vector). . For example, the bidirectional vector sequence 24 is converted based on the question sentence and the additional vector, the bidirectional vector sequence 25 is converted based on the correct answer sentence and the additional vector, and the bidirectional vector sequence 26 is converted into the incorrect answer sentence. And the additional vector. That is, the learning result is learned by the learning processing unit 132 while biasing the semantics according to the category.
Thereby, the information processing apparatus 1 according to the present embodiment can generate, for example, a category-specific answer to a question. Therefore, the information processing apparatus 1 according to the present embodiment can generate a more optimal answer sentence for a question.

Further, in the present embodiment, the learning result is obtained by the learning processing unit 132 as the correct answer embedding vector (O _{ac +} ) and the incorrect answer embedding corresponding to the first partial item (for example, “conclusion”) of the plurality of partial items. The bidirectional vector sequence 25-2 and the bidirectional vector sequence 26-2 corresponding to the second partial item (for example, “correction”) different from the first partial item are updated based on the vector (O _ac− ). Be 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 (2). To learn.
As a result, the information processing apparatus 1 according to the present embodiment can perform learning that optimizes the relationship between partial items (for example, the connection of partial items). Therefore, the information processing apparatus 1 according to the present embodiment can combine partial items to generate a more natural answer sentence.

The information processing apparatus 1 according to the present embodiment also 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. Further, the information processing apparatus 1 according to the present embodiment can improve the answer to the question by re-learning, for example.

Note that the learning processing unit 132 may re-learn the learning result based on a predetermined condition (for example, periodically or when the average matching rate AP drops below a predetermined value).
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 text. In the answer generation step, the answer generation unit 134 has a plurality of sets of a question sentence and a correct sentence and an incorrect sentence corresponding to each of a plurality of partial items in the answer sentence divided by a predetermined sentence path. Based on the learning information, the answer sentence to the input question sentence acquired in the question obtaining step is generated based on the above-described learning result that is machine-learned.
As a result, the information processing method according to the present embodiment has the same effect as that of the information processing apparatus 1 described above, and can generate a natural text answer with a reduced sense of discomfort to a question.

It should be noted that the present invention is not limited to the above-described embodiment, and can be modified within a range not departing from the spirit of the present invention.
For example, although the information processing apparatus 1 includes the learning processing unit 132 in the above-described embodiment, the information processing apparatus 1 is not limited to this. It does not have to be provided.

Further, although the information processing apparatus 1 has been described with reference to the 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, for example, It may be provided in an external server device. Further, in the information processing device 1, an external server device may include a part of the functional units included in the control unit 13.
In the above embodiment, the information processing device 1 has been described as an example including one server device, but may include 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 partial items of “conclusion” and “supplement”, but the present invention is not limited to this. You may correspond to one or more partial items.
Further, in the above-described embodiment, the information processing apparatus 1 applies the method including the QA-LSTM section 20 for each partial item, the method in which the semantics by category is biased to learn, and the method by the attention mechanism. Although an example has been described, the present invention is not limited to this. The information processing device 1 may be, for example, a mode in which some of these methods are not applied, or may be in a mode in which one of these methods is applied.

Each configuration included in the above-described information processing device 1 has a computer system inside. Then, a program for realizing the function of each configuration included in the above-described information processing apparatus 1 is recorded in a computer-readable recording medium, and the program recorded in 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 according to. Here, “reading and executing a program recorded in a recording medium on 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 the Internet, WAN, LAN, and a communication line such as a dedicated line. The "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built 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 accessible from the distribution server for distributing the program. It should be noted that a configuration in which the programs are divided into a plurality of pieces and downloaded at different timings and then combined in the configurations provided in the information processing apparatus 1 or a distribution server that distributes each of the divided programs may be different. Further, the "computer-readable recording medium" holds a program for a certain period of time, such as a volatile memory (RAM) inside a computer system which is a server or a client when the program is transmitted via a network. It also includes things. Further, the program may be for realizing some of the functions described above. Further, it may be a so-called difference file (difference program) that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

  Further, some or all of the above-mentioned functions may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each of the functions described above may be individually implemented as a processor, or a part or all of the functions may be integrated and implemented as 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 a technology for forming an integrated circuit that replaces LSI appears due to the progress of semiconductor technology, an integrated circuit according to the technology may be used.

DESCRIPTION OF SYMBOLS 1 Information processing apparatus 2 Terminal device 10 Vector conversion part 11 NW communication part 12 Storage part 13 Control part 20, 20-1, 20-2 QA-LSTM part 21, 21-1, 21-2 Question embedded vector generation part 22, 22-1, 22-2 Correct embedded vector generation unit 23, 23-1, 23-2 Incorrect embedded vector generation unit 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 NW1 network

Claims (7)

A question acquisition unit that acquires the input input question text,
Machine-learned learning based on learning information having a plurality of pairs of correct and incorrect sentences corresponding to a plurality of sub-items divided by a predetermined sentence path in a question sentence and an answer sentence An answer generation unit that generates an answer sentence to the input question sentence acquired by the question acquisition unit based on the result,
The learning result is
Based on a question feature vector group generated by converting the question sentence for each word into a sequence of forward and backward bidirectional sequences of words in a time series, the sequence of the words is defined as the bidirectional sequence. The question intermediate vector generated by learning
A correct intermediate vector corresponding to each of the plurality of sub-items, based on a correct characteristic vector group generated based on the arrangement order of the bidirectional words of the characteristic vector converted the correct sentence for each word, A correct answer intermediate vector generated by learning the sequence of the words in the two directions,
An incorrect answer intermediate vector corresponding to each of the plurality of sub-items, wherein a feature vector obtained by converting the incorrect answer sentence into words is generated in an incorrect answer feature vector group based on the arrangement order of the bidirectional words. On the basis of the loss function calculated based on the combination of the plurality of partial items with the incorrect intermediate vector generated by learning the word sequence bidirectionally, the learning is performed by being optimized. Information processing device. The learning result is
A feature vector of a word included in the question sentence and a feature vector of a term related to the question sentence are generated based on an additional vector learned by averaging, the question intermediate vector, the correct intermediate vector, and The information processing apparatus according to claim 1, wherein learning is performed based on an incorrect answer intermediate vector. The question feature vector group is converted based on the question sentence and the additional vector,
The correct answer feature vector group is converted based on the correct answer sentence and the additional vector,
The information processing device according to claim 2, wherein the incorrect solution feature vector group is converted based on the incorrect solution sentence and the additional vector. The learning result is
A correct answer 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 incorrect information feature vector group is updated and learned. The information processing device according to claim 1. The information processing apparatus according to any one of claims 1 to 4, 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 input input question text,
The answer generation unit, based on the learning information having a plurality of pairs of correct sentence and incorrect sentence corresponding to each of a plurality of sub-items divided by a predetermined sentence course in the question sentence and the answer sentence, An answer generation step of generating an answer sentence to the input question sentence obtained by the question obtaining step based on a learning result machine-learned,
The learning result is
Based on a question feature vector group generated by converting the question sentence for each word into a sequence of forward and backward bidirectional sequences of words in a time series, the sequence of the words is defined as the bidirectional sequence. The question intermediate vector generated by learning
A correct intermediate vector corresponding to each of the plurality of sub-items, based on a correct characteristic vector group generated based on the arrangement order of the bidirectional words of the characteristic vector converted the correct sentence for each word, A correct answer intermediate vector generated by learning the sequence of the words in the two directions,
An incorrect answer intermediate vector corresponding to each of the plurality of sub-items, wherein a feature vector obtained by converting the incorrect answer sentence into words is generated in an incorrect answer feature vector group based on the arrangement order of the bidirectional words. On the basis of the loss function calculated based on the combination of the plurality of partial items with the incorrect intermediate vector generated by learning the word sequence bidirectionally, the learning is performed by being optimized. Information processing method. On the computer,
A question acquisition step in which the question acquisition unit acquires the input input question text,
The answer generation unit, based on the learning information having a plurality of pairs of correct sentence and incorrect sentence corresponding to each of a plurality of sub-items divided by a predetermined sentence course in the question sentence and the answer sentence, A program for executing an answer generation step of generating an answer sentence to the input question sentence acquired by the question acquisition step, based on a learning result machine-learned,
The learning result is
Based on a question feature vector group generated by converting the question sentence for each word into a sequence of forward and backward bidirectional sequences of words in a time series, the sequence of the words is defined as the bidirectional sequence. The question intermediate vector generated by learning
A correct intermediate vector corresponding to each of the plurality of sub-items, based on a correct characteristic vector group generated based on the arrangement order of the bidirectional words of the characteristic vector converted the correct sentence for each word, A correct answer intermediate vector generated by learning the sequence of the words in the two directions,
An incorrect answer intermediate vector corresponding to each of the plurality of sub-items, wherein a feature vector obtained by converting the incorrect answer sentence into words is generated in an incorrect answer feature vector group based on the arrangement order of the bidirectional words. On the basis of the loss function calculated based on the combination of the plurality of partial items with the incorrect intermediate vector generated by learning the word sequence bidirectionally, the learning is performed by being optimized. And the program.

Images