WO2022003762A1 - Question answering device, question answering method, and question answering program - Google Patents

Abstract

When outputting an answer to a question by machine reading, the present invention outputs the answer in a plurality of output formats. This question answering device comprises: a calculation unit which receives, as inputs, a question and a related document used to answer the question, and calculates information indicating the relevance between the question and the related document; a plurality of output units which receive, as an input thereof, the information indicating the relevance as calculated by the calculation unit, and output answers to the question in different output formats; and a selection unit which selects a predetermined number of answers from among the answers output by the plurality of output units in the respective output formats thereof.

Description

Question answering device, question answering method and question answering program

This disclosure relates to a question answering device, a question answering method, and a question answering program.

Machine reading comprehension is known as a question answering technology that automatically answers questions entered by the user in natural language. Machine reading comprehension is a technology that inputs a question by a user and a related document (referred to as a "passage") written in natural language, and outputs a response to the input question based on the information extracted from the passage. Is.

There are various output formats when outputting the answer to the question by the machine reading comprehension, and as an example,
-A format that outputs the answer sentence generated by sentence generation based on the information extracted from the passage,
-A format that outputs labels such as YES / NO based on the information extracted from the passage.
-A format that outputs questions (questions for narrowing down the answers) generated based on the information extracted from the passage.
And so on.

Kyosuke Nishida, Itsumi Saito, Kosuke Nishida, Kazutoshi Shinoda, Atsushi Otsuka, Hisako Asano, and Junji Tomita, "Multi-style generative reading comprehension", In Proceedings of The 57th 2273-2284, 2019. Kosuke Nishida, Kyosuke Nishida, Masaaki Nagata, Itsumi Saito, Atushi Otuka, Hisako Asano and Junji Tomita, "Answering while Summarizing: Multi-task Learning for Multi-hop QA with Evidence Extraction" Computational Linguistics (ACL2019), pp.2273-2284.2019. Atsushi Otsuka, Kyosuke Nishida, Itumi Saito, Hisako Asano, Junji Tomita, Tetsuji Sato. Proposal of question answering method by machine reading comprehension focusing on clarification of question intention. Journal of the Japanese Society for Artificial Intelligence, vol.34, no.5 A-J14, pp.1-12, 2019.

Here, in order to realize a machine reading comprehension that can be output in any of the above output formats, it is conceivable to combine machine reading comprehension models corresponding to each output format.

However, if the number of machine reading models corresponding to the number of output formats is expanded and executed on the memory, a large amount of resources in the device will be consumed, which is realized in the device with limited resources. It's unlikely.

It is an object of the present disclosure to provide a question answering device, a question answering method, and a question answering program that can output in a plurality of output formats when the answer to a question is output by machine reading comprehension.

According to one aspect of the present disclosure, the question answering device is
A calculation unit that calculates information indicating the relevance of the question and the related document by inputting the question and the related document used when answering the question.
A plurality of output units that output responses to the question in different output formats by using the information indicating the relevance calculated by the calculation unit as their respective inputs.
The plurality of output units have a selection unit that selects a predetermined number of responses from each response output in each output format.

According to the present disclosure, it is possible to provide a question answering device, a question answering method, and a question answering program that can output a response to a question by machine reading comprehension in a plurality of output formats.

FIG. 1 is a diagram showing an example of a hardware configuration of a question answering device. FIG. 2 is a diagram showing a functional configuration of a question answering unit of a comparative example. FIG. 3 is a diagram showing an example of the functional configuration of the question answering unit of the question answering device according to the first embodiment. FIG. 4 is a first diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. FIG. 5 is a second diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. FIG. 6 is a third diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. FIG. 7 is a flowchart showing a flow of question answering processing by the question answering device according to the first embodiment. FIG. 8 is a diagram showing an example of the functional configuration of the question answering device according to the second embodiment. FIG. 9 is a first diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. FIG. 10 is a second diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. FIG. 11 is a third diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. FIG. 12 is a flowchart showing a flow of question answering processing by the question answering device according to the second embodiment.

Hereinafter, each embodiment will be described with reference to the attached drawings. In the present specification and the drawings, the components having substantially the same functional configuration are designated by the same reference numerals, and duplicate description thereof will be omitted.

[First Embodiment]
<Hardware configuration of question answering device>
First, the hardware configuration of the question answering device according to the first embodiment will be described. FIG. 1 is a diagram showing an example of a hardware configuration of a question answering device.

As shown in FIG. 1, the question and answer device 100 includes a processor 101, a memory 102, an auxiliary storage device 103, an I / F (Interface) device 104, a communication device 105, and a drive device 106. The hardware of the question answering device 100 is connected to each other via the bus 107.

The processor 101 has various arithmetic devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 101 reads various programs (not shown) installed in the auxiliary storage device 103 onto the memory 102 and executes them.

The memory 102 has a main storage device such as a ROM (ReadOnlyMemory) and a RAM (RandomAccessMemory). The processor 101 and the memory 102 form a so-called computer, and the processor 101 executes various programs read on the memory 102, so that the computer realizes various functions.

For example, in the present embodiment, the computer formed by the processor 101 and the memory 102 realizes the question and answer unit 110 by executing the question and answer program read by the processor 101 on the memory 102. As will be described later, the question answering unit 110 is configured to suppress the consumption of computer resources in the question answering device 100.

The auxiliary storage device 103 stores various programs and various data used when the various programs are executed by the processor 101. For example, in the present embodiment, the auxiliary storage device 103 has a learning data set storage unit 120 and a passage storage unit 130, and various data (learning data set and passage (related document described in natural language)) can be stored. Store.

The I / F device 104 is a connection device that connects the input device 140 and the output device 141 to the question answering device 100. The I / F device 104 receives a question for the question answering device 100 via the input device 140. Further, the I / F device 104 outputs a response generated by the question answering device 100 to the input question via the output device 141. The input device 140 referred to here includes a device for converting an input question into voice data, a device for converting into text data, and the like. Similarly, the output device 141 referred to here includes a device that outputs a response by voice data, a device that outputs a response by text data, and the like.

The communication device 105 is a communication device for communicating with another device via a network.

The drive device 106 is a device for setting the recording medium 142. The recording medium 142 referred to here includes a medium such as a CD-ROM, a flexible disk, a magneto-optical disk, or the like, which records information optically, electrically, or magnetically. Further, the recording medium 142 may include a semiconductor memory or the like for electrically recording information such as a ROM or a flash memory.

The various programs installed in the auxiliary storage device 103 are installed, for example, by setting the distributed recording medium 142 in the drive device 106 and reading the various programs recorded in the recording medium 142 by the drive device 106. Will be done. Alternatively, various programs installed in the auxiliary storage device 103 may be installed by being downloaded from the network via the communication device 105.

Similarly, as for the various data stored in each storage unit of the auxiliary storage device 103, for example, the distributed recording medium 142 is set in the drive device 106, and the various data recorded in the recording medium 142 is the drive device 106. It is stored by being read by. Alternatively, various data stored in each storage unit of the auxiliary storage device 103 may be stored by being downloaded from the network via the communication device 105.

<Functional configuration of question answering section>
Next, the details of the functional configuration of the question answering unit 110 will be described. In the explanation, in order to clarify the characteristics of the functional configuration of the question answering unit 110, as a comparative example, first, the function of the question answering unit constructed by combining the number of machine reading models according to the number of output formats. The configuration will be described.

(1) Functional configuration of the question answering unit of the comparative example FIG. 2 is a diagram showing a functional configuration of the question answering unit of the comparative example. When outputting the answer to the question by machine reading comprehension, the question answering unit 200 of the comparative example is configured to be able to output in a plurality of output formats.
・ Input unit 210,
A plurality of machine reading models (in the example of FIG. 2, the first machine reading model 221 to the third machine reading model 223),
・ Selection unit 230,
Have.

The input unit 210 inputs the input question and a passage (related document written in natural language) to each of a plurality of machine reading comprehension models.

The first machine reading model 221 outputs the answer sentence generated by sentence generation as the first response based on the information extracted from the passage.

The second machine reading model 222 outputs a label such as YES / NO generated based on the information extracted from the passage as the second response.

The third machine reading model 223 outputs a question generated based on the information extracted from the passage (a question generated to narrow down the answer (referred to as a revised question)) as a third response.

The selection unit 230 selects and outputs a predetermined number of responses from the responses output from each of the first machine reading model 221 to the third machine reading model 223.

As shown in the question answering unit 200 of the comparative example, the answer to the question can be output in a plurality of output formats by combining the number of machine reading models according to the number of output formats. On the other hand, in the case of the question answering unit 200 of the comparative example, there are the following problems.
-Since the configuration is such that a number of machine reading models corresponding to the number of output formats are expanded and executed on the memory, a large amount of computer resources in the question answering device 100 are consumed.
-When selecting a predetermined number of responses from the responses output from each of the first machine reading model 221 to the third machine reading model 223, it is not possible to select an appropriate response. This is because it is not possible to determine the superiority or inferiority of the response simply by comparing the first response to the third response, and it is necessary to calculate some selection index.

On the other hand, the question answering unit 110 of the question answering device 100 according to the first embodiment has a configuration for solving these problems. This will be described in detail below.

(2) Functional configuration of the question answering unit FIG. 3 is a diagram showing an example of the functional configuration of the question answering unit of the question answering device according to the first embodiment. When the question answering unit 110 outputs the answer to the question by machine reading comprehension, the question answering unit 110 can output in a plurality of output formats, suppresses the consumption of computer resources, and is appropriate from among the responses in the plurality of output formats to the question. As a configuration to allow you to select a response
・ Input unit 310,
-Understanding layer 320, which functions as an input layer,
A first output layer 321 and a second output layer 322, a third output layer 323, which function as an output layer.
-Output judgment layer 324, which functions as an output layer,
・ Selection unit 330,
Have.

Of these, the input unit 310 inputs the input question and a passage (related document written in natural language) to the understanding layer 320.

Further, the understanding layer 320 is an example of a calculation unit, in which a question and a passage are input, information indicating the relationship between the question and the passage is calculated on a deep learning vector, and a state vector or a state tensor is output. The understanding layer 320 can adopt any structure as long as it can input the question and the passage and calculate the information indicating the relationship between the question and the passage.

For example, the understanding layer 320 can employ BiDAF using RNN () (see Non-Patent Document 1), Transformer-based BERT (see Non-Patent Document 2), and the like. However, the understanding layer 320 is an output that matches the format of the state vector or the state tensor that is the input of the output layer (first output layer 321 to the third output layer 323, output judgment layer 324) after the understanding layer 320. It is necessary to have a structure to do.

The first output layer 321 to the third output layer 323 are an example of the output unit, and the information (state vector or state tensor) indicating the relationship between the question and the passage output from the understanding layer 320 is used as an input. Output the response (output result of machine reading comprehension).

In the case of FIG. 3, the first output layer 321 outputs the response sentence generated by the sentence generation as the first response. Further, the second output layer 322 outputs a label such as YES / NO as a second response. Further, the third output layer 323 outputs a revised question for narrowing down the answer to the input question as a third response.

The first output layer 321 to the third output layer 323 may have any deep learning structure in outputting the first response to the third response, respectively. Further, the output format output by the first output layer 321 to the third output layer 323 is not limited to the answer sentence, the label, and the revised question, and the response of another output format may be output.

Further, in the example of FIG. 3, the configuration is such that the first to third output layers are installed, but the number of output layers to be installed is arbitrary, and even if a plurality of output layers having the same structure are installed. good. For example, two sentence generation decoder structures are installed, one output layer is an output layer learned to generate an answer sentence by sentence generation, and the other output layer generates a revised question by sentence generation. It may be an output layer learned in this way.

The output determination layer 324 is an example of the index calculation unit, and calculates the probability distribution of each response output from the first output layer 321 to the third output layer 323.

Specifically, the output determination layer 324 has a softmax layer having a number of dimensions corresponding to the number N of output formats (N = 3 in the example of FIG. 3). The N-dimensional softmax layer receives a state vector or a state tensor from the understanding layer 320, and calculates the probability distribution of each response.

Note that this configuration is effective when the number N of output formats is fixed. On the other hand, in such a configuration, when it becomes necessary to add an output format, the number of dimensions of the softmax layer cannot be accommodated, so that it is necessary to perform re-learning processing for the entire question answering unit 110.

In the selection unit 330, among the responses output from each of the first output layer 321 to the third output layer 323, the probability distribution calculated by the output determination layer 324 is the upper M (M is a predetermined number, In the example of FIG. 3, the response of M = 1) is selected and output to the user via the output device 141.

In this way, in the question answering unit 110,
・ A shared understanding layer is set up for multiple output formats (the input layer is shared), and
-The input layer (understanding layer) and the output layer (first output layer 321 to third output layer 323, output judgment layer 324) are separated, and each layer is modularized.
-The selection unit 330 selects the response to be finally output using the probability distribution of each response calculated from the output judgment layer 324 installed in the output layer as a selection index.
It was configured. As a result, according to the question answering unit 110, when the answer to the question is output by machine reading comprehension, it can be output in a plurality of output formats, and at the same time, it can be output.
・ By standardizing the input layer, the consumption of computer resources is suppressed and
・ Select an appropriate response by calculating the selection index,
Will be possible.

<Learning method of question answering section>
Next, the learning method of the question answering unit 110 will be described. In the learning phase in which the question answering unit 110 is subjected to the learning process, the question answering unit 110 of the question answering device 100 according to the present embodiment first installs the comparison / change unit 410 in place of the selection unit 330. Subsequently, the question answering unit 110 of the question answering device 100 according to the present embodiment has each layer (understanding layer 320, first output layer 321 to third output layer 323, output determination layer) installed in the question answering unit 110. The learning process is performed for 324).

At that time, in the question-and-answer unit 110 of the question-and-answer device 100, a learning data set used when performing learning processing on an individual machine-reading model (first machine-reading model 221 to third machine-reading model 223) is used. use.

Specifically, the question answering unit 110 of the question answering device 100 first sets a flag indicating which output layer the learning data set is used for the learning process.

Subsequently, the question answering unit 110 of the question answering device 100 stores the correct answer data of the output determination layer 324 in the learning data set. For example, the question-and-answer unit 110 of the question-and-answer device 100 sets the vector data in which the dimensional value corresponding to the set flag is "1" and the other dimensional values are "0", and the correct answer data of the output determination layer 324. Store in the training data set as.

Subsequently, the question answering unit 110 of the question answering device 100 calculates a learning loss between the response output from the output layer corresponding to the set flag and the corresponding correct answer data, and based on the calculated learning loss. Update the parameters of the output layer and understanding layer corresponding to the set flag. At this time, the question answering unit 110 of the question answering device 100 ignores the response output from the output layer other than the output layer corresponding to the set flag.

Further, the question answering unit 110 of the question answering device 100 calculates a learning loss between the vector data output from the output determination layer and the corresponding correct answer data, and based on the calculated learning loss, the parameter of the output determination layer. To update.

FIG. 4 is a first diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. In the case of FIG. 4, it is assumed that the question answering unit 110 sets a flag indicating that the learning data set 400 = "the learning data set used for the learning process for the first output layer 321". ..

As shown in FIG. 4, the learning data set 400 includes "input data", "correct answer data of the first output layer", and "correct answer data of the output determination layer" as information items.
-Questions and passages are stored in the "input data".
-The "correct answer data of the first output layer" stores the correct answer data of the answer sentence generated by sentence generation based on the information extracted from the corresponding passage.
-The "correct answer data of the output judgment layer" includes "first dimension" to "third dimension", the value of the first dimension is "1", and the values of the second dimension and the third dimension are "0". Vector data is stored.

In FIG. 4, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 400 to the understanding layer 320, the machine reading comprehension is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, vector data of N dimensions (N = 3 in the example of FIG. 4) is output from the output determination layer 324.

In the comparison / change unit 410, the first response (answer sentence) output from the first output layer 321 and the answer sentence stored in the "correct answer data of the first output layer" of the learning data set 400. Calculate the learning loss between. Further, the comparison / change unit 410 updates the parameters of the first output layer 321 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 410, the vector data of N dimensions (N = 3 in the example of FIG. 4) output from the output judgment layer 324 and the correct answer data of the "output judgment layer" of the training data set 400. The learning loss is calculated between the vector data of the first dimension to the third dimension stored in ". Further, the comparison / change unit 410 updates the parameters of the output determination layer 324 based on the calculated learning loss.

On the other hand, FIG. 5 is a second diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. In the case of FIG. 5, a flag indicating that the learning data set 500 = "the learning data set used for the learning process for the second output layer 322" is set by the question answering unit 110. do.

As shown in FIG. 5, the learning data set 500 includes "input data", "correct answer data of the second output layer", and "correct answer data of the output determination layer" as information items.
-Questions and passages are stored in the "input data".
-In the "correct answer data of the second output layer", the correct answer data of the label such as YES / NO generated based on the information extracted from the corresponding passage is stored.
-The "correct answer data of the output judgment layer" includes "first dimension" to "third dimension", the second dimension value is "1", and the first and third dimension values are "0". Vector data is stored.

In FIG. 5, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 500 to the understanding layer 320, the machine reading comprehension is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, vector data of N dimensions (N = 3 in the example of FIG. 5) is output from the output determination layer 324.

In the comparison / change unit 410, between the second response (label) output from the second output layer 322 and the label stored in the "correct answer data of the second output layer" of the learning data set 500. Calculate the learning loss with. Further, the comparison / change unit 410 updates the parameters of the second output layer 322 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 410, the vector data of N dimensions (N = 3 in the example of FIG. 5) output from the output judgment layer 324 and the correct answer data of the "output judgment layer" of the training data set 500. The learning loss is calculated between the vector data of the first dimension to the third dimension stored in ". Further, the comparison / change unit 410 updates the parameters of the output determination layer 324 based on the calculated learning loss.

On the other hand, FIG. 6 is a third diagram showing an operation example in the learning phase of the question answering device according to the first embodiment. In the case of FIG. 6, it is assumed that the question answering unit 110 sets a flag indicating that the learning data set 600 = "the learning data set used for the learning process for the third output layer 323". ..

As shown in FIG. 6, the learning data set 600 includes "input data", "correct answer data of the third output layer", and "correct answer data of the output determination layer" as information items.
-Questions and passages are stored in the "input data".
-The "correct answer data of the third output layer" stores the correct answer data of the revised question generated based on the information extracted from the corresponding passage.
-The "correct answer data of the output judgment layer" includes "first dimension" to "third dimension", the value of the third dimension is "1", and the values of the first dimension and the second dimension are "0". Vector data is stored.

In FIG. 6, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 600 to the understanding layer 320, the machine reading comprehension is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, vector data of N dimensions (N = 3 in the example of FIG. 6) is output from the output determination layer 324.

In the comparison / change unit 410, the third response (revised question) output from the third output layer 323 and the revised question stored in the "correct answer data of the third output layer" of the learning data set 600. Calculate the learning loss between. Further, the comparison / change unit 410 updates the parameters of the third output layer 323 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 410, the vector data of N dimensions (N = 3 in the example of FIG. 6) output from the output judgment layer 324 and the correct answer data of the "output judgment layer" of the training data set 600. The learning loss is calculated between the vector data of the first dimension to the third dimension stored in ". Further, the comparison / change unit 410 updates the parameters of the output determination layer 324 based on the calculated learning loss.

As described above, in the question answering unit 110 of the question answering device 100, the learning data sets 400 to 600 are used to cover the understanding layer 320, the first output layer 321 to the third output layer 323, and the output determination layer 324. The learning process is performed sequentially.

<Flow of question answering process>
Next, the flow of the question answering process by the question answering device 100 will be described. FIG. 7 is a flowchart showing a flow of question answering processing by the question answering device according to the first embodiment. Of these, steps S701 to S703 represent processing in the learning phase, and steps S704 to S707 represent processing in the response phase.

In step S701, the question answering unit 110 performs learning processing on the understanding layer 320, the first output layer 321 and the output determination layer 324 using the learning data set 400.

In step S702, the question answering unit 110 performs learning processing on the understanding layer 320, the second output layer 322, and the output determination layer 324 using the learning data set 500.

In step S703, the question answering unit 110 performs learning processing on the understanding layer 320, the third output layer 323, and the output determination layer 324 using the learning data set 600.

In step S704, the input unit 310 of the question answering unit 110 accepts the input of the question and the passage, and inputs the input question and the passage to the understanding layer 320.

In step S705, the first output layer 321 to the third output layer 323 take the state vector output from the understanding layer 320 as an input, and output the first response to the third response.

In step S706, the output determination layer 324 of the question answering unit 110 outputs the selection index by calculating the probability distribution of the first response to the third response by inputting the state vector output from the understanding layer 320. do.

In step S707, the selection unit 330 of the question answering unit 110 selects the predetermined upper M responses based on the selection index output from the output determination layer 324, and outputs the selected response.

<Summary>
As is clear from the above description, the question answering device 100 according to the first embodiment is
-It has an understanding layer that calculates information indicating the relationship between the question and the passage by inputting the question and the passage.
-It has a first output layer to a third output layer that outputs first to third responses that are different output formats from each other, using information indicating relevance calculated by the understanding layer as each input.
-It has an output judgment layer that calculates the probability distribution of the first response to the third response based on the information indicating the relevance output by the understanding layer. Further, it has a selection unit for selecting a predetermined number of responses using the probability distribution of the first response to the third response calculated by the output determination layer as a selection index.

Thereby, according to the question answering device 100 according to the first embodiment, it is possible to output the answer to the question in a plurality of output formats, suppress the consumption of computer resources, and select an appropriate response. Become.

That is, according to the first embodiment, when a response to a question is output by machine reading comprehension, a highly feasible question answering device, a question answering method, and a question answering program that can be output in a plurality of output formats are provided. be able to.

[Second Embodiment]
In the first embodiment, it is assumed that the number of output formats is fixed, and when a new output format is added, the question answering unit 110 is assumed to perform relearning processing for the entire question answering unit 110. Configured.

On the other hand, in the second embodiment, the question answering unit is configured so that the question answering unit 110 as a whole does not need to be relearned even when a new output format is added. Hereinafter, the second embodiment will be described focusing on the differences from the first embodiment.

<Functional configuration of question answering section>
First, the functional configuration of the question answering unit of the question answering device according to the second embodiment will be described. FIG. 8 is a diagram showing an example of the functional configuration of the question answering unit of the question answering device according to the second embodiment.

The difference from the functional configuration shown in FIG. 3 is that in the case of the question answering unit 800 of FIG. 8, individual scores are calculated as selection indexes for each of the first output layer 321 to the third output layer 323. The point is that the first output determination layer 801 to the third output determination layer 803 are installed. Further, in the case of the question answering unit 800 of FIG. 8, the function of the selection unit 810 is different from the function of the selection unit 330 of FIG.

The first output determination layer 801 is an example of the index calculation unit, and has a logit layer that receives the state vector of the first output layer 321 and calculates a scalar value of 0 to 1.0 as the first score.

Similarly, the second output determination layer 802 is an example of the index calculation unit, and has a logit layer that receives the state vector of the second output layer 322 and calculates a scalar value of 0 to 1.0 as the second score. ..

Similarly, the third output determination layer 803 is an example of the index calculation unit, and has a logit layer that receives the state vector of the third output layer 323 and calculates a scalar value of 0 to 1.0 as the third score. ..

The selection unit 810 selects the response corresponding to the predetermined top M scores based on the first score to the third score calculated by the first output judgment layer 801 to the third output judgment layer 803. And output.

As described above, in the question answering device 100 according to the second embodiment, the first output determination layer 801 to the third output for calculating individual scores for each of the first output layer 321 to the third output layer 323. A judgment layer 803 is installed. As a result, according to the second embodiment, even if a new output format is added, it is sufficient to perform learning processing on the added new output layer, output judgment layer, and understanding layer, and the learning has already been completed. It is not necessary to perform re-learning processing on the output layer and the output judgment layer of.

<Learning method of question answering section>
Next, the learning method of the question answering unit 800 will be described. In the learning phase in which the question answering unit 110 is subjected to the learning process, the question answering unit 800 of the question answering device 100 according to the present embodiment first installs the comparison / change unit 910 in place of the selection unit 810. Subsequently, the question answering unit 800 of the question answering device 100 according to the present embodiment has each layer (understanding layer 320, first output layer 321 to third output layer 323, first) installed in the question answering unit 800. The learning process is performed on the output determination layer 801 to the third output determination layer 803).

At that time, in the question and answer device 100, as in the first embodiment, the learning data used when performing the learning process on the individual machine reading model (first machine reading model 221 to third machine reading model 223). Use the set.

Specifically, the question answering unit 800 of the question answering device 100 first sets a flag indicating which output layer the learning data set is used for the learning process.

Subsequently, the question answering unit 800 of the question answering device 100 stores the correct answer data of the first score to the third score calculated by the first output determination layer 801 to the third output determination layer 803 in the learning data set. do. For example, the question answering unit 800 of the question answering device 100 stores correct answer data having a score corresponding to the set flag of "1.0" in the learning data set.

Subsequently, the question answering unit 800 of the question answering device 100 calculates a learning loss between the response output from the output layer corresponding to the set flag and the corresponding correct answer data, and based on the calculated learning loss. Update the parameters of the output layer and understanding layer corresponding to the set flag. At this time, the question answering unit 800 of the question answering device 100 ignores the response output from the output layer other than the output layer corresponding to the set flag.

Further, the question answering unit 800 of the question answering device 100 calculates a learning loss between the score output from the output determination layer corresponding to the set flag and the corresponding correct answer data, and based on the calculated learning loss. Update the parameters of the output judgment layer corresponding to the set flag. At this time, the question answering device 100 ignores the scores output from the output determination layers other than the output determination layer corresponding to the set flag.

FIG. 9 is a first diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. In the case of FIG. 9, the flag indicating that the learning data set 900 = "the learning data set used for the learning process for the first output layer 321 and the first output judgment layer 801" is the question answering unit 800. It is assumed that it is set by.

As shown in FIG. 9, the learning data set 900 includes "input data", "correct answer data of the first output layer", and "correct answer data of the output judgment layer" as information items.
-Questions and passages are stored in the "input data".
-The "correct answer data of the first output layer" stores the correct answer data of the answer sentence generated by sentence generation based on the information extracted from the corresponding passage.
-The "correct answer data of the output determination layer" stores the correct answer data of the score (first score) output from the first output determination layer 801.

In FIG. 9, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 900 to the understanding layer 320, the machine reading comprehension is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, the first score to the third score are output from the first output determination layer 801 to the third output determination layer 803.

In the comparison / change unit 910, the first response (answer sentence) output from the first output layer 321 and the answer sentence stored in the "correct answer data of the first output layer" of the learning data set 900. Calculate the learning loss between. Further, the comparison / change unit 910 updates the parameters of the first output layer 321 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 910, the first score output from the first output judgment layer 801 and the value stored in the first score of the "correct answer data of the output judgment layer" of the learning data set 900. Calculate the learning loss between. Further, the comparison / change unit 910 updates the parameters of the first output determination layer 801 based on the calculated learning loss.

On the other hand, FIG. 10 is a second diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. In the case of FIG. 10, the flag indicating that the learning data set 1000 = "the learning data set used for the learning process for the second output layer 322 and the second output judgment layer 802" is the question response unit 800. It is assumed that it is set by.

As shown in FIG. 10, the learning data set 1000 includes "input data", "correct answer data of the second output layer", and "correct answer data of the output determination layer" as information items.
-Questions and passages are stored in the "input data".
-In the "correct answer data of the second output layer", the correct answer data of the label such as YES / NO generated based on the information extracted from the corresponding passage is stored.

In the "correct answer data of the output judgment layer", the correct answer data of the score (second score) output from the second output judgment layer 802 is stored.

In FIG. 10, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 1000 to the understanding layer 320, the machine reading comprehension is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, the first score to the third score are output from the first output determination layer 801 to the third output determination layer 803.

In the comparison / change unit 910, between the second response (label) output from the second output layer 322 and the label stored in the "correct answer data of the second output layer" of the learning data set 1000. Calculate the learning loss with. Further, the comparison / change unit 910 updates the parameters of the second output layer 322 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 910, the second score output from the second output judgment layer 802 and the value stored in the second score of the "correct answer data of the output judgment layer" of the learning data set 1000. Calculate the learning loss between. Further, the comparison / change unit 910 updates the parameters of the second output determination layer 802 based on the calculated learning loss.

On the other hand, FIG. 11 is a third diagram showing an operation example in the learning phase of the question answering device according to the second embodiment. In the case of FIG. 11, the flag indicating that the learning data set 1100 = "the learning data set used for the learning process for the third output layer 323 and the third output determination layer 803" is the question response unit 800. It is assumed that it is set by.

As shown in FIG. 11, the learning data set 1100 includes "input data", "correct answer data of the third output layer", and "correct answer data of the output determination layer" as information items.
-Questions and passages are stored in the "input data".
-The "correct answer data of the third output layer" stores the correct answer data of the revised question generated based on the information extracted from the corresponding passage.
-The "correct answer data of the output determination layer" stores the correct answer data of the score (third score) output from the third output determination layer 803.

In FIG. 11, when the input unit 310 inputs the input data (a set of a question and a passage) of the learning data set 1100 to the understanding layer 320, machine learning is performed from the first output layer 321 to the third output layer 323. As the output result of, the first response to the third response are output. Further, the first score to the third score are output from the first output determination layer 801 to the third output determination layer 803.

In the comparison / change unit 910, the third response (revised question) output from the third output layer 323 and the revised question stored in the "correct answer data of the third output layer" of the learning data set 1100. Calculate the learning loss between. Further, the comparison / change unit 910 updates the parameters of the third output layer 323 and the understanding layer 320 based on the calculated learning loss.

Similarly, in the comparison / change unit 910, the third score output from the third output judgment layer 803 and the value stored in the third score of the "correct answer data of the output judgment layer" of the learning data set 1100. Calculate the learning loss between. Further, the comparison / change unit 910 updates the parameters of the third output determination layer 803 based on the calculated learning loss.

<Flow of question answering process>
Next, the flow of the question answering process by the question answering device 100 according to the second embodiment will be described. FIG. 12 is a flowchart showing a flow of question answering processing by the question answering device according to the second embodiment. The difference from the flowchart described with reference to FIG. 7 in the first embodiment is in steps S1201 to S1203.

In step S1201, the question answering unit 110 performs learning processing on the understanding layer 320, the first output layer 321 and the first output determination layer 801 using the learning data set 900.

In step S1202, the question answering unit 110 performs learning processing on the understanding layer 320, the second output layer 322, and the second output determination layer 802 using the learning data set 1000.

In step S1203, the question answering unit 110 performs learning processing on the understanding layer 320, the third output layer 323, and the third output determination layer 803 using the learning data set 1100.

<Summary>
As is clear from the above description, the question answering device 100 according to the second embodiment is
-It has an understanding layer that calculates information indicating the relationship between the question and the passage by inputting the question and the passage.
-It has a first output layer to a third output layer that outputs first to third responses that are different output formats from each other, using information indicating relevance calculated by the understanding layer as each input.
A first, which receives the state vectors of the first output layer to the third output layer and calculates individual scores (first score to third score) for the first output layer to the third output layer. It has an output determination layer to a third output determination layer. Further, it has a selection unit for selecting a predetermined number of responses using the first score to the third score calculated by the first output determination layer to the third output determination layer as a selection index.

As a result, according to the question answering device 100 according to the second embodiment, the answer to the question can be output in a plurality of output formats as in the first embodiment, the consumption of computer resources is suppressed, and an appropriate response is made. Can be selected. In addition, according to the question answering device 100 according to the second embodiment, even when a new output format is added, it is not necessary to perform relearning processing for the entire question answering unit.

That is, according to the second embodiment, when a response to a question is output by machine reading comprehension, a more feasible question answering device, a question answering method, and a question answering program that can be output in a plurality of output formats are provided. can do.

[Other embodiments]
In the first embodiment and the second embodiment, the case where different output judgment layers are installed has been described, but whether the output judgment layer in the first embodiment is installed or the output determination in the second embodiment is performed. The decision to install a layer is optional. For example, it may be determined in consideration of the task and purpose to be set, the system configuration of the question answering device, and the like.

Further, in the first embodiment and the second embodiment, it has been described that the learning phase and the response phase are executed in the same question answering device 100. However, the learning phase and the response phase may be configured to be executed by separate devices. In this case, the device that executes the response phase does not need to have the learning data set storage unit 120, and the comparison / change units 410 and 910 are not installed.

It should be noted that the present invention is not limited to the configurations shown here, such as combinations with other elements in the configurations and the like described in the above embodiments. These points can be changed without departing from the spirit of the present invention, and can be appropriately determined according to the application form thereof.

100: Question response device 110: Question response unit 120: Learning data set storage unit 130: Passage storage unit 310: Input unit 320: Understanding layer 321: First output layer 322: Second output layer 323: Third Output layer 324: Output judgment layer 330: Selection unit 400 to 600: Learning data set 801: First output judgment layer 802: Second output judgment layer 803: Third output judgment layer 810: Selection unit 900 to 1100 : Training dataset

Claims (5)

A calculation unit that calculates information indicating the relevance of the question and the related document by inputting the question and the related document used when answering the question.
A plurality of output units that output responses to the question in different output formats by using the information indicating the relevance calculated by the calculation unit as their respective inputs.
A question answering device having a selection unit that selects a predetermined number of responses from each response output by the plurality of output units in each output format. It further has an index calculation unit that receives the information indicating the relevance calculated by the calculation unit and calculates the probability distribution of each response output by the plurality of output units in each output format.
The question answering device according to claim 1, wherein the selection unit selects a predetermined number of responses based on the probability distribution of each response. Further having a plurality of index calculation units for receiving information calculated when the plurality of output units output each response and calculating a score for each of the plurality of output units.
The question answering device according to claim 1, wherein the selection unit selects a predetermined number of responses based on a score for each of the plurality of output units. A calculation step of calculating information indicating the relevance of the question and the related document by inputting the question and the related document used for answering the question.
A plurality of output steps that output a response to the question in different output formats by using the information indicating the relevance calculated in the calculation step as each input, and
A question answering method including a selection step of selecting a predetermined number of responses from each response output in each output format in the plurality of output steps. A calculation step of calculating information indicating the relevance of the question and the related document by inputting the question and the related document used for answering the question.
A plurality of output steps that output a response to the question in different output formats by using the information indicating the relevance calculated in the calculation step as each input, and
A question answering program for causing a computer to execute a selection process of selecting a predetermined number of responses from each response output in each output format in the plurality of output processes.

Images