JP2018045630A - Information processing device, information processing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing device, information processing method, and program that improve the satisfaction of a user with a response message.SOLUTION: An information processing device includes a reception unit, a first processing unit, a second processing unit, and a generation unit. The reception unit receives a message of a user. The first processing unit performs selection processing for selecting a response message candidate corresponding to the message from a response message candidate group. The second processing unit performs evaluation processing for evaluating words that follow the message on the basis of probability values in order in which words appear. The generation unit generates a response message corresponding to the message, on the basis of the selection processing and evaluation processing.SELECTED DRAWING: Figure 1

Description

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

  Conventionally, information processing that generates response message candidates for a message such as a question sentence using a ranking model that has been machine-learned based on a feature amount obtained for a pair of a question sentence, a correct sentence, and a sentence including an incorrect sentence An apparatus is known (see Patent Document 1).

JP2013-254420A

  However, the response message generated from the response message candidate generated using only the model as described above has low accuracy as a response message to the user's message and may not satisfy the user's satisfaction sufficiently. There is room for.

  The present application has been made in view of the above, and an object thereof is to provide an information processing apparatus, an information processing method, and a program that improve user satisfaction with response messages.

  An information processing apparatus according to the present application includes a reception unit, a first processing unit, a second processing unit, and a generation unit. The receiving unit receives a user message. A 1st process part performs the selection process which selects the response message candidate with respect to a message from a response message candidate group. A 2nd process part performs the evaluation process which evaluates the word following the said message based on the probability value of the order that a word appears. The generation unit generates a response message for the message based on the selection process and the evaluation process.

  According to one aspect of the embodiment, it is possible to provide an information processing apparatus, an information processing method, and a program that improve user satisfaction with response messages.

FIG. 1 is an explanatory diagram of information processing according to the first embodiment. FIG. 2 is a diagram illustrating a configuration example of the information processing system. FIG. 3 is a diagram illustrating a configuration example of the information processing apparatus according to the first embodiment. FIG. 4 is a diagram illustrating an example of scores for response message candidates. FIG. 5 is a flowchart illustrating an example of a response message generation process according to the embodiment. FIG. 6 is a diagram illustrating a configuration example of the information processing apparatus according to the second embodiment. FIG. 7 is a flowchart illustrating an example of a response message generation process according to the second embodiment. FIG. 8 is a diagram illustrating an example of scores for response message candidates in the modification. FIG. 9 is a hardware configuration diagram illustrating an example of a computer that implements the functions of the information processing apparatus.

  Hereinafter, an information processing apparatus, an information processing method, and a form for implementing a program (hereinafter referred to as “embodiment”) according to the present application will be described in detail with reference to the drawings. Note that the information processing apparatus, the information processing method, and the program according to the present application are not limited by this embodiment.

(First embodiment)
[1. Information processing]
An example of information processing according to the first embodiment will be described. FIG. 1 is an explanatory diagram of information processing according to the first embodiment.

  The information processing device 1 receives a message based on the user's utterance from the user's terminal device 2 (step S1). For example, the information processing apparatus 1 receives a message based on an utterance “What is the weather today?” From the user.

  The information processing apparatus 1 selects a response message candidate for the message using the matching model for the message (step S2).

  The matching model is a model for selecting a response message candidate corresponding to an input message using a distributed expression indicating the input message as a multi-dimensional vector.

  The matching model uses a dialogue sentence obtained from the web, Twitter (registered trademark), etc. as learning data, and a distributed expression in the input message and a distributed expression in the corresponding message that is a suitable pair for the input message. Are learned to exist close together in the distributed representation space. The matching model is generated by learning using a technique of DSSM (Deep Structured Sematic Model) that uses a recurrent neural network (RNN) such as LSTM (Long Short-Term Memory) for distributed expression generation.

  For example, if the input message is "What's the weather today?" And there is a corresponding message such as "Sunny", "That's right" or "Good morning" for this message The matching model is such that distributed expressions such as "Is sunny", "It is cloudy", and "I don't know" exist near the distributed expression of the message "What's the weather today?" To be learned. In addition, distributed expressions such as “Return” that are not appropriate as a response to “What is the weather today?” Exist far away from the distributed expression “What is the weather today?” In the distributed expression space. The matching model is learned.

  The information processing apparatus 1 performs matching between the distributed representation in the input message and the distributed representation of the corresponding message in the distributed representation space.

  For example, the information processing apparatus 1 calculates the cosine similarity between the distributed representation in the input message and the distributed representation of the corresponding message in the distributed representation space (hereinafter, may be referred to as a response message candidate group). Then, a predetermined number of corresponding messages are selected as response message candidates from the distributed messages corresponding to a high cosine similarity. Note that the information processing apparatus 1 may select a response message candidate using another scale such as the reciprocal of the Euclidean distance instead of the cosine similarity.

  That is, the information processing apparatus 1 performs a selection process of selecting a response message candidate for a message from the response message candidate group. The predetermined number is a preset value. The information processing apparatus 1 selects a predetermined number of corresponding messages in descending order of cosine similarity as response message candidates.

  The information processing apparatus 1 calculates a score indicating the likelihood of the response message candidate for the message, using the translation model for the selected response message candidate (step S3).

  The translation model is a model for evaluating and generating a word (message) following an input message by using a distributed expression indicating the input message by a multi-dimensional vector. The response message evaluated and generated using the translation model is a response message suitable for the input message.

  The translation model is a model in which dialogue sentences obtained from the web or the like are used as learning data and the probability that words of corresponding messages corresponding to a certain input message are chained.

  When the input message is given to the translation model, a probability table (probability of occurrence of each word in the vocabulary) for the first word of the corresponding message is obtained, and further, which word comes to the first word When you decide, you get a probability table for the second word that follows. By repeating this, a probability table can be obtained in order for the word sequence of the opposite message.

  The information processing apparatus 1 can generate response message candidates by sampling words sequentially from the first word based on the probability table. In addition, if there is a response message candidate, the information processing apparatus 1 measures the probability that the word included in the response message candidate has a probability table obtained by the translation model, so that The likelihood of the response message candidate can be evaluated.

  The translation model is learned and generated using an RNN (Recurrent Neural Network) technique such as LSTM (Long Short-Term Memory).

  For example, by using a translation model, if the message is “What's the weather today?” Followed by the message “Is it sunny?” It is possible to calculate a probability value in which “deo” continues after “sunny”, and the like.

  Using the translation model, the information processing apparatus 1 calculates the probability value of the order in which words included in the response message candidate appear for the message, and performs an evaluation process for evaluating the word following the message.

  The information processing device 1 calculates a probability value that the first word of each response message candidate appears for the input message using the translation model. Then, the information processing device 1 calculates a probability value that the second word appears for the first word of each response message candidate using the translation model. In this way, the information processing device 1 calculates probability values in the order of appearance for each word included in each response message candidate, and obtains a score obtained by combining the probability values, for example, the product of the probability values. Calculate the score.

  When the probability value calculated in the order in which each word appears in the response message candidate is large and the score is large for the input message, the response message candidate can be evaluated as likely as a response message. .

  The information processing apparatus 1 selects a response message candidate having the highest score calculated using the translation model as a response message, and generates a response message (step S4).

  The information processing device 1 transmits the generated response message to the user terminal device 2 (step S5).

  The information processing apparatus 1 selects a response message using a translation model after selecting a response message candidate using a matching model for the message.

  In the evaluation using the matching model, it is necessary to generate a distributed representation of the response message candidates without knowing what message is input, and it may be difficult to perform an accurate evaluation of the response message candidates. . On the other hand, in the evaluation using the translation model, it is possible to evaluate whether the content corresponding to the message is included in the response message candidate after grasping the input message, so that a more appropriate evaluation can be performed. .

  On the other hand, in the matching model, the response message candidates are already distributed, and when a distributed representation of the message is generated, evaluation can be performed by matching the distributed representations, and the processing load is small. On the other hand, in the translation model, if the probability of each word of a large number of response message candidates is evaluated, the processing load is large.

  In this way, between the matching model and the translation model, there is a trade-off relationship between the precision of evaluation and the load during processing.

  Therefore, the information processing apparatus 1 narrows down from a large number of corresponding messages to a predetermined number of response message candidates using a matching model with a small processing load, and uses only a translation model for the narrowed response message candidates. Make an evaluation. Thereby, the precision as the whole system can be improved, the precision of the response message with respect to a message can be improved, and a user's satisfaction can be improved.

[2. Configuration of information processing system 5]
FIG. 2 is a diagram illustrating a configuration example of the information processing system 5. As illustrated in FIG. 2, the information processing system 5 according to the first embodiment includes an information processing device 1, a terminal device 2, a speech recognition server 3, and a speech synthesis server 4.

  The terminal device 2, the speech recognition server 3, the speech synthesis server 4, and the information processing device 1 are connected to be communicable with each other wirelessly or via a network N. The network N is, for example, a LAN (Local Area Network) or a WAN (Wide Area Network) such as the Internet.

  The terminal device 2 is realized by a smartphone, a tablet terminal, a desktop PC (Personal Computer), a notebook PC, a PDA (Personal Digital Assistant), or the like.

  The speech recognition server 3 is a device that performs natural language processing on speech information and converts speech data into text data. When the speech recognition server 3 receives speech speech data from the terminal device 2, the speech recognition server 3 converts the speech data into text data. The voice recognition server 3 transmits text data obtained by converting the voice data to the information processing apparatus 1.

  The voice synthesis server 4 converts the text data of the response message generated by the information processing apparatus 1 into voice data. The voice synthesis server 4 transmits the voice data obtained by converting the text data to the terminal device 2.

  The information processing device 1 generates text data of a response message based on text data transmitted from the terminal device 2 or text data obtained by converting the speech data via the speech recognition server 3. The information processing device 1 transmits the text data of the generated response message to the speech synthesis server 4 and the terminal device 2.

  Note that the speech recognition server 3 and the speech synthesis server 4 may be configured integrally with the information processing apparatus 1. When the terminal device 2 has a voice recognition function or a voice synthesis function, the voice data and the text data may be converted using these functions.

[3. Configuration of information processing apparatus 1 according to the first embodiment]
Next, the information processing apparatus 1 according to the first embodiment will be described with reference to FIG. FIG. 3 is a diagram illustrating a configuration example of the information processing apparatus 1 according to the first embodiment.

  Here, an example in which voice data based on a user's utterance is transmitted from the terminal device 2 and voice data is transmitted to the terminal device 2, that is, a voice dialogue is described as an example, but a dialogue using text data may be used. .

  The information processing device 1 is a response generation device that generates a response message for a message transmitted from the terminal device 2 (see FIG. 2) via the voice recognition server 3 (see FIG. 2) when the user speaks. is there. The information processing apparatus 1 includes a reception unit 10, a transmission unit 20, a storage unit 30, and a control unit 40.

  The receiving unit 10 receives a message from the user via the network N. When there is an utterance by the user, the receiving unit 10 receives text data corresponding to the message converted by the voice recognition data (see FIG. 2). The receiving unit 10 receives data from a server or the like installed outside via the network N.

  The storage unit 30 includes a matching model storage unit 31 and a translation model storage unit 32. The storage unit 30 is realized by, for example, a semiconductor memory element such as a RAM or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 30 may include a database that stores information necessary for selecting a response message candidate for the message.

  The matching model storage unit 31 stores a matching model. The matching model may be newly acquired and updated via the network N. The matching model storage unit 31 stores a distributed representation of response message candidates for a distributed representation of an input message and response message candidates as a set of lists.

  The translation model storage unit 32 stores a translation model. The translation model may be newly acquired and updated via the network N.

  The control unit 40 includes an analysis unit 41, a first processing unit 42, a second processing unit 43, and a generation unit 44.

  The analysis unit 41 analyzes the message received by the reception unit 10. For example, when the receiving unit 10 receives text data, the analyzing unit 41 analyzes the text data using morphological analysis or the like, and extracts a word group included in the text data.

  When the receiving unit 10 receives text data such as “Today's weather?” As a message, the analyzing unit 41 receives “today”, “no”, “weather”, “ha”, for example. "Is extracted and the content of the text data is specified.

  The first processing unit 42 generates a distributed representation of the message. The first processing unit 42 can handle messages having similar contents as the same distributed expression by using the messages as distributed expressions. The first processing unit 42 treats messages having similar contents such as “What is the weather today?” And “How about the weather today?” As the same distributed expression.

  Then, using the matching model, the first processing unit 42 performs a selection process of selecting a response message candidate from the response message candidate group based on the distributed expression of the response message candidate having a high similarity to the distributed expression of the message. Specifically, the first processing unit 42 selects a corresponding message having a distributed expression having a high cosine similarity as a response message candidate with respect to the distributed expression of the message. For example, the first processing unit 42 selects a predetermined number of corresponding messages as response message candidates from the corresponding messages in the distributed representation having a high cosine similarity with respect to the distributed representation of the message.

  For example, if the message is “What's the weather today?”, “It ’s sunny”, “It ’s cloudy”, or “I do n’t know.” Corresponding message such as "" is selected as a response message candidate. Here, the case where the weather is “sunny” will be described as an example.

  The second processing unit 43 calculates the probability value of the order in which words included in the response message candidate appear for the message using the translation model, and performs an evaluation process for evaluating the word following the message.

  Specifically, the second processing unit 43 calculates the probability value of the order in which words included in the response message candidates appear for the message using the translation model, and calculates a score that combines the probability values. To do. That is, the second processing unit 43 calculates a score as an evaluation value indicating the likelihood of the message with respect to the response message candidate selected by the first processing unit 42. Specifically, the second processing unit 43 calculates the appearance probability of the response message candidate by taking the product of the probability values.

  For example, when the message is “What's the weather today?”, “Sunny weather”, “Cloudy”, and “I don't know” are selected as response message candidates.

  In this case, the second processing unit 43 uses the translation model to calculate the probability value of the word following “What is the weather today?” In the order of appearance of the words included in each response message candidate.

  The second processing unit 43 calculates a probability value that “sunny” appears next to “what is the weather today?” Using the translation model, and further after “is the weather today?” The probability value of “desho” appearing in is calculated using a translation model. Similarly, the second processing unit 43 calculates a probability value of the order in which words appear using the translation model for the response message candidate “will be sunny”. Then, the second processing unit 43 calculates the product of the respective probability values as a score for the response message candidate “will be sunny”.

For example, the probability value of the word “sunny” following “What is the weather today?” Is “0.2”, and the probability value of the word “de” after “What is the weather today?” Is “0.4”. The probability value of the word “U” following “Today ’s weather? In this case, the score, which is the product of the probability values, is “2.88 × 10 ⁻² ” as shown in FIG. FIG. 4 is a diagram illustrating an example of scores for response message candidates. In FIG. 4, the end of the sentence is shown as “EOS (End Of Sentence)”.

  In response to the message “What's the weather today?”, If the response message candidate “maybe sunny” is likely, the score will increase.

  Further, the second processing unit 43 similarly calculates probability values for the response message candidates “cloudy” and “I don't know”, and calculates a score for each response message candidate.

For example, the probability value of the word “cloudy” following “What is the weather today?” Is “0.05”, and the probability value of the word “is” next to “Today's weather? Cloudy” is “0.4”. If the probability value ending with “Is today's weather? Cloudy” is “0.7”, the score is “1.40 × 10 ⁻² ” as shown in FIG.

Also, for example, the probability value of the word “Understanding” following “What is the weather today?” Is “0.01”, and the probability value of the word “No” following “Understanding today's weather” is “0. 2 ”, the probability value of the word“ n ”following“ I don't know the weather today ”is“ 0.5 ”, and the probability value that ends with“ I do n’t know the weather today? ”Is“ 0.6 ” In this case, the score is “6.00 × 10 ⁻⁴ ” as shown in FIG.

  The generation unit 44 selects a response message candidate having the highest score calculated by the second processing unit 43 as a response message, and generates a response message.

  For example, when the scores of the above three response message candidates calculated by the second processing unit 43 are as shown in FIG. 4, the generation unit 44 selects the response message candidate with the highest score “sunny”. Select as a response message and generate a response message.

  The transmission unit 20 transmits the response message created by the generation unit 44 to the terminal device 2 and the voice synthesis server 4 (see FIG. 2) via the network N.

[4. Response message generation processing according to first embodiment]
Next, the response message generation process according to the first embodiment will be described with reference to FIG. FIG. 5 is a flowchart illustrating an example of a response message generation process according to the first embodiment.

  When a message is received by the receiving unit 10 (step S10), the first processing unit 42 performs a selection process using the matching model, and selects a response message candidate for the message (step S11).

  The second processing unit 43 performs an evaluation process using the translation model on the selected response message candidate, and calculates a score indicating the likelihood of each response message candidate (step S12).

  The generation unit 44 selects a response message candidate having the highest score as a response message, and generates a response message (step S13).

  The transmission unit 20 transmits the generated response message to the user terminal device 2 (step S14).

(Second Embodiment)
[5. Configuration of Information Processing Device 1 According to Second Embodiment]
Next, the information processing apparatus 1 according to the second embodiment will be described with reference to FIG. FIG. 6 is a diagram illustrating a configuration example of the information processing apparatus 1 according to the second embodiment.

  Here, the description will focus on the points different from the first embodiment, and the description of the same configuration as the first embodiment will be omitted.

  The storage unit 30 of the information processing apparatus 1 according to the second embodiment includes a language model storage unit 33. In addition, the control unit 40 of the information processing apparatus 1 according to the second embodiment includes a third processing unit 45.

  The language model storage unit 33 stores a language model. The language model may be newly acquired and updated via the network N.

  The language model is a model that uses sentences obtained from the web or the like as learning data and statistically summarizes the appearance rates of response message candidates used in general responses. The language model is learned and generated using an RNN (Recurrent Neural Network) technique such as LSTM (Long Short-Term Memory).

  In the language model, a score weighted according to the appearance rate of response message candidates is attached. For example, a large score is assigned to a high-frequency response message candidate whose appearance rate is higher than a predetermined rate, such as “Like” or “Sure!”.

  The 3rd process part 45 performs the adjustment process which makes low the probability value by which a high frequency response message candidate is produced | generated as a response message using a language model. Specifically, the third processing unit 45 calculates a score for each response message candidate selected by the first processing unit 42 using the language model.

  The generation unit 44 subtracts the score calculated using the language model by the third processing unit 45 from the score of the response message candidate calculated by the second processing unit 43, and the response having the largest score among the subtracted scores A message candidate is selected as a response message.

  By subtracting the score calculated by the third processing unit 45 from the score of the response message candidate calculated by the second processing unit 43, a response message candidate that is frequently used as a response message is used as a response message. Selection can be suppressed.

  The third processing unit 45 calculates, for each response message candidate, a coefficient that decreases as the score given by the language model increases, and sets the score of the response message candidate calculated by the second processing unit 43. The calculated coefficient may be multiplied.

[6. Response message generation processing according to second embodiment]
Next, a response message generation process according to the second embodiment will be described with reference to FIG. FIG. 7 is a flowchart illustrating an example of a response message generation process according to the second embodiment.

  When a message is received by the receiving unit 10 (step S20), the first processing unit 42 performs a selection process using the matching model, and selects a response message candidate for the message (step S21).

  The second processing unit 43 performs an evaluation process on the selected response message candidate using the translation model, and calculates a score indicating the likelihood of each response message candidate (step S22).

  The third processing unit 45 performs adjustment processing using the language model, and calculates a score for each response message candidate (step S23).

  The generation unit 44 subtracts the score calculated using the language model by the third processing unit 45 from the score calculated by the second processing unit 43 using the translation model, and the response having the largest score among the subtracted scores A message candidate is selected as a response message, and a response message is generated (step S24).

  The transmission unit 20 transmits the generated response message to the user terminal device 2 (step S25).

[7. Modified example]
In addition to the above-described embodiment, the following modifications can be applied.

  The information processing apparatus 1 according to the modification generates a response message by switching the order of processing in the first processing unit 42 and the second processing unit 43.

  The second processing unit 43 performs an evaluation process for evaluating words following the message based on the probability value of the order of appearance of words using the translation model, and generates a response message candidate based on the evaluation result.

  Specifically, the second processing unit 43 generates a probability table indicating how much each word in the vocabulary can appear as the first word of the response message candidate when the message is assumed, The first word is determined by selecting N words (N is a preset value) with the highest probability or by performing random sampling based on a probability table. Subsequently, a probability table for the next word based on the message and the first word is generated, and the second word is similarly determined based on the probability table. By repeating such processing, the second processing unit 43 automatically generates response message candidates for the message, and generates a response message candidate group including a plurality of response message candidates.

  The first processing unit 42 performs a selection process of selecting a response message candidate that is likely to be a message from the response message candidate group generated by the second processing unit 43. Specifically, the first processing unit 42 selects a response message candidate having a distributed expression having the highest cosine similarity with respect to the distributed expression of the message from the response message candidate group, and further narrows down the response message candidates.

  The generation unit 44 selects the response message candidates narrowed down by the first processing unit 42 as a response message, and generates a response message.

  For example, even if a response message candidate that is not suitable as a response message is generated by the second processing unit 43, the distributed representation of such a response message candidate is far from the distributed representation of the message in the distributed representation space. Exists.

  Therefore, the 1st process part 42 can suppress selecting a response message candidate which is not suitable as a response message by further narrowing down a response message candidate using a matching model.

  The information processing apparatus 1 according to the modified example can diversify response message candidates for messages by generating response message candidates by the second processing unit 43.

  Further, the information processing apparatus 1 according to the modified example uses, for example, another translation model to replace the sentence with a more natural sentence when “second day” is generated as a response message candidate by the second processing unit 43. A response message such as “sunny” may be generated. Another translation model is a model that generates an appropriate sentence when an inappropriate sentence is given by using a large amount of examples of an inappropriate sentence and an appropriate sentence pair as learning data. Then, the information processing apparatus 1 selects a response message from the response message candidates generated as an appropriate sentence by the first processing unit 42.

  In addition, in the translation model, for example, a response message that is a general response when learning a translation model so that response message candidates that are general responses such as “Like” and “That's right” are difficult to create. The learning may be performed so that the probability value of the order of the appearing words is small. For example, in the set of input messages and corresponding messages in the learning data, pre-processing that thins out the ones that responded “Yes,” which is a general response to “It seems to be fine tomorrow.” Then, learn the translation model Thereby, it can suppress that the 2nd process part 43 produces | generates the response message candidate used as a general response. Therefore, the information processing apparatus 1 according to the modification can improve user satisfaction.

  In addition to the above-described modification examples, the following modification examples can be applied.

  The second processing unit 43 may calculate the sum of logarithms of the respective probability values as a score.

  The second processing unit 43 may calculate a geometric average of probability values (or an arithmetic average of logarithms of probability values) as a score. An example of calculating the geometric mean as a score is shown in FIG. In this case, the score of the response message candidate “will be sunny” is “0.412”, the score of “cloudy” is “0.241”, and the score of “I don't know” is “0.157”. . Therefore, the response message candidate “will be sunny” is selected as the response message.

  In addition, the second processing unit 43 combines a score obtained by combining the products of the respective probability values and the geometric mean of the respective probability values, or a sum of logarithms of the probability values and an arithmetic average of the logarithms of the respective probability values. A score may be calculated. For example, when the logarithm of the probability value is taken, the second processing unit 43 may add a minus sign to the logarithm to obtain the cost value. In this case, if the cost value becomes small, it is evaluated that it is likely.

  Thereby, it can suppress that the score of the response message candidate of a short sentence becomes high, can perform evaluation independent of the length of a response message candidate, and can select an appropriate response message candidate as a response message.

  The second processing unit 43 may calculate the score of the response message candidate in consideration of the user context. The context is, for example, the user's utterance time or the user's current position. For example, the second processing unit 43 changes the probability value that a word appears in accordance with the context.

  For example, in the translation model, what happens to the corresponding message for the pair of the utterance time and the input message is what was learned as a corresponding message on the premise of the input message at the time of learning. It is changed to the form to be learned and learned. For example, when generating a response message candidate, the second processing unit 43 similarly uses the utterance time as input information. In the learning data, when the utterance time is evening or night, there are very few cases where the corresponding message “good morning” appears, so the translation model learns it, and the probability that “good morning” will appear in such a time zone Lower the value.

  Thereby, the production | generation part 44 can select a response message according to a user's context. Therefore, the information processing apparatus 1 can improve user satisfaction with the response message.

  The second processing unit 43 may calculate the probability value that the word appears using the “attention mechanism”. By using the “attention mechanism”, the second processing unit 43 does not use the information in the message uniformly, but weights more strongly the portion that is strongly related to the word to be generated. The probability value is calculated while reinterpreting the information.

  Accordingly, the second processing unit 43 can accurately evaluate the likelihood of the response message candidate for the message. Therefore, the information processing apparatus 1 can provide a more accurate response message to the user, and improve the user's satisfaction with the response message.

  Further, when learning the matching model, the matching model may be learned by removing a message that is not suitable as a response message, for example, a sentence example including an abuse word or an obscene word.

  Thereby, it is possible to prevent a response message candidate including such a word from being selected.

  The analysis unit 41 may perform analysis in units of characters, or may perform analysis without extracting a word group.

[8. effect]
The information processing apparatus 1 includes a receiving unit 10, a first processing unit 42, a second processing unit 43, and a generation unit 44. The receiving unit 10 receives a user message. The first processing unit 42 performs a selection process of selecting a response message candidate for the message from the response message candidate group. The second processing unit 43 performs an evaluation process for evaluating the word following the message based on the probability value of the order in which the words appear. The generation unit 44 generates a response message for the message based on the selection process and the evaluation process.

  Thereby, the information processing apparatus 1 can improve the accuracy of the response message, and can improve the user's satisfaction with the response message.

  The second processing unit 43 performs an evaluation process on the response message candidate selected by the first processing unit 42, and calculates a score indicating the likelihood of the response message candidate for the message. The generation unit 44 generates a response message for the message based on the score.

  Thereby, since the information processing apparatus 1 performs the evaluation process on the response message candidate selected by the first processing unit 42, the load in the evaluation process is reduced, the required time is shortened, and the response message is transmitted to the user. It can be quickly transmitted to the terminal device 2. In addition, the information processing apparatus 1 can perform precise evaluation by the evaluation process, can improve the accuracy of a response message to the message, and can improve user satisfaction.

  Moreover, since the information processing apparatus 1 selects a response message candidate using the matching model by the first processing unit 42, it is possible to suppress selection (generation) of an inappropriate response message candidate.

  The second processing unit 43 generates a response message candidate for the message based on the evaluation process. The first processing unit 42 further narrows down response message candidates from the response message candidates generated by the second processing unit 43. Then, the generation unit 44 generates a response message based on the response message candidates narrowed down by the first processing unit 42.

  Thereby, the information processing apparatus 1 can diversify the response message candidates by automatically generating the response message candidates by the second processing unit 43.

  In addition, the information processing apparatus 1 can replace the response message candidate with a more natural sentence by the first processing unit 42. Therefore, user satisfaction can be improved.

  The first processing unit 42 performs a selection process of selecting a response message candidate based on the similarity between the distributed representation of the message and the distributed representation of the response message candidate.

  As a result, the information processing apparatus 1 can select a response message candidate for a similar message, can generate an appropriate response message for the similar message, and the user's satisfaction with the response message. The degree can be improved.

  The second processing unit 43 performs an evaluation process based on the user context and the probability value of the order in which words appear.

  Thereby, the information processing apparatus 1 can generate a response message according to the user's context, and can improve the user's satisfaction.

  The information processing apparatus 1 performs, for a high-frequency response message candidate whose appearance rate is higher than a predetermined rate, a third processing unit 45 that performs an adjustment process for reducing a probability value that a high-frequency response message candidate is generated as a response message Is provided. The generation unit 44 generates a response message for the message based on the selection process, the search process, and the adjustment process.

  Thereby, the information processing apparatus 1 can reduce the probability value that a high-frequency response message candidate is generated as a response message, and can improve user satisfaction.

[9. Hardware configuration]
The information processing apparatus 1 according to the above-described embodiment is realized by a computer 1000 configured as shown in FIG. 9, for example. FIG. 9 is a hardware configuration diagram illustrating an example of a computer that realizes the functions of the information processing apparatus 1. The computer 1000 includes a CPU 1100, RAM 1200, ROM 1300, HDD 1400, communication interface (I / F) 1500, input / output interface (I / F) 1600, and media interface (I / F) 1700.

  The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400 and controls each unit. The ROM 1300 stores a boot program executed by the CPU 1100 when the computer 1000 is started up, a program depending on the hardware of the computer 1000, and the like.

  The HDD 1400 stores programs executed by the CPU 1100, data used by the programs, and the like. The communication interface 1500 receives data from other devices via the network N and sends the data to the CPU 1100, and transmits data determined by the CPU 1100 to other devices via the network N.

  The CPU 1100 controls an output device such as a display and a printer and an input device such as a keyboard and a mouse via the input / output interface 1600. The CPU 1100 acquires data from the input device via the input / output interface 1600. Further, the CPU 1100 outputs the determined data to the output device via the input / output interface 1600.

  The media interface 1700 reads a program or data stored in the recording medium 1800 and provides it to the CPU 1100 via the RAM 1200. The CPU 1100 loads such a program from the recording medium 1800 onto the RAM 1200 via the media interface 1700, and executes the loaded program. The recording medium 1800 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory. Etc.

  For example, when the computer 1000 functions as the information processing apparatus 1 according to the embodiment, the CPU 1100 of the computer 1000 implements the function of the control unit 40 by executing a program loaded on the RAM 1200. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them. However, as another example, these programs may be acquired from other devices via the network N.

  As described above, some of the embodiments and modifications of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various aspects can be made based on the knowledge of those skilled in the art including the aspects described in the disclosure line of the invention. It is possible to carry out the present invention in other forms that have been modified and improved.

[10. Other]
In addition, among the processes described in the above-described embodiments and modifications, all or a part of the processes described as being automatically performed can be manually performed, or are described as being performed manually. All or part of the processing can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

  In addition, the above-described embodiments and modifications can be combined as appropriate within a range that does not contradict processing contents.

  In addition, the “section (module, unit)” described above can be read as “means” or “circuit”. For example, the receiving unit 10 can be read as receiving means or a receiving circuit.

DESCRIPTION OF SYMBOLS 1 Information processing apparatus 2 Terminal apparatus 10 Reception part 20 Transmission part 30 Storage part 40 Control part 41 Analysis part 42 1st process part 43 2nd process part 44 Generation part 45 3rd process part

Claims (8)

A receiver for receiving a user message;
A first processing unit that performs a selection process of selecting a response message candidate for the message from a response message candidate group;
A second processing unit for performing an evaluation process for evaluating a word following the message based on a probability value of the order in which the words appear;
An information processing apparatus comprising: a generation unit that generates a response message to the message based on the selection process and the evaluation process. The second processing unit includes:
Performing the evaluation process on the response message candidate selected by the first processing unit, and calculating a score indicating the likelihood of the response message candidate for the message;
The generator is
The information processing apparatus according to claim 1, wherein the response message for the message is generated based on the score. The second processing unit includes:
Generating response message candidates for the message based on the evaluation process;
The first processing unit includes:
Further narrowing down the response message candidates from the response message candidates generated by the second processing unit;
The generator is
The information processing apparatus according to claim 1, wherein the response message is generated based on the response message candidates narrowed down by the first processing unit. The first processing unit includes:
The selection process of selecting the response message candidate based on a similarity between the distributed representation of the message and the distributed representation of the response message candidate is performed. The information processing apparatus described in 1. The second processing unit includes:
The information processing apparatus according to claim 1, wherein the evaluation process is performed based on the user's context and the probability value of the order in which the words appear. A third processing unit that performs an adjustment process for reducing a probability value that the high-frequency response message candidate is generated as the response message with respect to a high-frequency response message candidate having an appearance rate higher than a predetermined rate;
The generator is
The information processing apparatus according to any one of claims 1 to 5, wherein the response message for the message is generated based on the selection process, the evaluation process, and the adjustment process. An information processing method executed by an information processing apparatus,
A receiving process for receiving a user message;
A first processing step of performing a selection process of selecting a response message candidate for the message from a response message candidate group;
A second processing step of performing an evaluation process for evaluating a word following the message based on a probability value of the order in which the words appear;
An information processing method comprising: generating a response message for the message based on the selection process and the evaluation process. A receiving procedure for receiving the user's message;
A first processing procedure for performing a selection process of selecting a response message candidate for the message from a response message candidate group;
A second processing procedure for performing an evaluation process for evaluating a word following the message based on a probability value of the order in which the words appear;
A program for causing a computer to execute a generation procedure for generating a response message to the message based on the selection process and the evaluation process.

Images