Patent Application
20220229997
The present invention relates to a dialogue processing apparatus, a learning apparatus, a dialogue processing method, a learning method, and a program.
Research on machine reading comprehension has been actively conducted. Machine reading comprehension is a technique for taking a document and a question as inputs and generating an answer to the question from the document. As methods for generating answers by machine reading comprehension, there are two known modes, an “extractive mode” in which words are retrieved from a document to generate an answer and a “generative mode” in which an answer is generated also using words other than the words in the document. In general, it is known that the generative mode can generate more natural sentences as answers than the extractive mode.
The machine reading comprehension that has been actively studied is of a question-and-answer mode. That is, in the machine reading comprehension, even if a plurality of questions are input, these questions are treated independently and an answer is generated independently for each question. On the other hand, interactive machine reading comprehension where questions and answers alternate has also been studied in recent years.
Interactive machine reading comprehension is a technique for taking a document, a current question, a past question history, and a past answer history as inputs and generating an answer to the current question. Thus, in interactive machine reading comprehension, it is possible to generate an answer based on the content of a dialogue so far (that is, a question history and an answer history so far) while questions and answers alternate. For example, Non Patent Literature 1 describes a technique for outputting a start position and an end position of a range corresponding to an answer in a document as interactive machine reading comprehension for generating an answer to a current question in an extractive mode.
Non Patent Literature 1: Hsin-Yuan Huang, Eunsol Choi, and Wen-tau Yih. Flowqa: Grasping Flow in History for Conversational Machine Comprehension. arXiv preprint arXiv: 1810.06683, 2018.
However, for example, in the interactive machine reading comprehension described in Non Patent Document 1, an answer cannot be generated in a generative mode because the answer is generated in an extractive mode. Thus, the interactive machine reading comprehension of the related art can be applied to only one of the extractive mode and the generative mode.
Embodiments of the present invention have been made in view of the above points and an object of the present invention is to implement interactive machine reading comprehension that can be applied to both an extractive mode and a generative mode.
To achieve the above object, a dialogue processing apparatus according to an embodiment of the present invention includes a generation unit configured to take a question Qi that is a word sequence representing a current question in a dialogue, a document P used to generate an answer Ai to the question Qi, a question history {Qi-1, . . . , Qi-k} that is a set of word sequences representing k past questions, and an answer history {Ai-1, . . . , Ai-k} that is a set of word sequences representing answers to the k questions as inputs and to generate the answer Ai by machine reading comprehension in an extractive mode or a generative mode using pre-trained model parameters.
It is possible to implement interactive machine reading comprehension that can be applied to both the extractive mode and the generative mode.
An embodiment of the present invention (hereinafter also referred to as the “present embodiment”) will be described below. In the present embodiment, a dialogue processing apparatus 10 that implements interactive machine reading comprehension applicable to both the extractive mode and the generative mode will be described. That is, the dialogue processing apparatus 10 that implements interactive machine reading comprehension by taking a document, a current question, a past question history, and a past answer history as inputs and generating an answer to the current question in an extractive mode or a generative mode will be described. A document may also be referred to as a “passage,” “search target text,” or the like.
Overall Configuration of Dialogue Processing Apparatus 10
Interactive machine reading comprehension uses a model implemented by neural networks to generate an answer to the current question. Thus, there are a “learning time” when model parameters are learned and an “inference time” when an answer to the current question is generated using the learned model parameters. The model parameters are stored, for example, in a storage device such as an auxiliary storage device.
Overall Configuration of Dialogue Processing Apparatus 10 during Inference
An overall configuration of the dialogue processing apparatus 10 for inference will be described with reference to
As illustrated in
Each of the current question Qi and the past questions Qi-1, . . . , Qi-k is text including one or more words (that is, a character sequence). Similarly, each of the past answers Ai-1, . . . , Ai-k is text including one or more words. Similarly, the document P is text including one or more words. Hereinafter, the number of words included in the document P (that is, the length of the document P) will be denoted by T. However, T is not limited to the number of words included in the document P and may be, for example, the “sum of the number of words included in the document P and the number of special characters included therein.” Examples of the special characters include characters representing the beginning of a sentence, characters representing the end of a sentence, characters used for padding, and characters used to combine words and sentences.
In the present embodiment, it is assumed that a single document P is input, although a document set including a plurality of documents may be input in the interactive machine reading comprehension. In the present embodiment, it is also assumed that the document P has a length of the number of words of about one paragraph (for example, about several hundred words).
As illustrated in
The question encoding unit 110 takes the document P and the current question Qi as inputs and calculates features {ui1, . . . , uiT} of the document P conditioned on the current question Qi using the learned model parameters (that is, encodes the current question Qi). Here, ui1, . . . , uiT are d-dimensional vectors.
The context encoding unit 120 takes the document P, the question history {Qi-1, . . . Qi-k} and the answer history {Ai-1, . . . , Ai-k} as inputs and calculates features {ui-11, . . . , ui-k1}, . . . , {ui-1T, . . . , ui-kT} of the document P conditioned on the question history {Qi-1, . . . , Qi-k} and features {vi-11, . . . , vi-k1}, . . . , {vi-1T, . . . , vi-kT} of the document P conditioned on the answer history {Ai-1, . . . , Ai-k} (that is, encodes the question history {Qi-1, . . . , Qi-k} and the answer history {Ai-1, . . . , Ai-k}) using the learned model parameters. For t=1, . . . , T, ui-1t, . . . , ui-kt, vi-1t, . . . vi-kt are d-dimensional vectors.
Here, for example, in the interactive machine reading comprehension described in Non Patent Decument 1, each answer in the answer history is expressed by a pair of start and end positions of a range corresponding to the answer in the document, while each question in the question history is expressed in text. On the other hand, in the present embodiment, each answer in the answer history {Ai-1, . . . , Ai-k} and each question in the question history {Qi-1, . . . , Qi-k} are both expressed in text as described above. Therefore, in the present embodiment, the features {ui-1t, . . . , ui-kt} (where t=1, . . . , T) of the document P regarding the question history and the features {vi-1t, . . . , vi-kt} (where t=1, . . . , T) of the document P regarding the answer history are treated equivalently.
The context combining unit 130 takes the features {ui1, . . . , uiT} calculated by the question encoding unit 110 and the features {ui-1t. . . , ui-kt} and {vi-1t, . . . , vi-kt} (where t=1, . . . , T) calculated by the context encoding unit 120 as inputs and calculates features ot (where t=1, . . . , T) using the learned model parameters. These features ot are d×(1+2k)-dimensional vectors and are each a feature reflecting the current question Qi, the question history {Qi-1, . . . , Qi-k}, and the answer history {Ai-1, . . . , Ai-k}.
The answer generation unit 140 takes the features ot (where t=1, . . . , T) calculated by the context combining unit 130 as an input and generates an answer Ai using the learned model parameters. At this time, the answer generation unit 140 generates the answer Ai in an extractive mode or a generative mode as described later.
Overall Configuration of Dialogue Processing Apparatus 10 during Learning
An overall configuration of the dialogue processing apparatus 10 during learning will be described with reference to
As illustrated in
The update unit 200 takes a correct answer to the current question Qi, which is represented by the following, and the answer Ai generated by the dialogue processing unit 100 as inputs, and updates the model parameters using a supervised learning method.
Thereby, the model parameters are learned. In the following, the correct answer to the current question Qi will also be represented by “{circumflex over ( )}Ai.”
Dialogue Process
In the following, a process for taking a current question Qi, a document P, a question history {Qi-1, . . . , Qi-k}, and an answer history {Ai-1, . . . , Ai-k} as inputs and generating an answer Ai in an extractive mode or a generative mode using learned model parameters (a dialogue process) will be described with reference to
Step S101: The question encoding unit 110 takes the document P and the current question Qi as inputs and calculates features uit (where t=1, . . . , T) of the document P conditioned on the current question Qi using the learned model parameters.
Here, for example, a model called Bidirectional Encoder Representations from Transformers (BERT) described in Reference 1 below can be employed as a model of a neural network that implements the question encoding unit 110.
Reference 1:
J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv: 1810.04805, 2018. The model of the neural network that implements the question encoding unit 110 may be an encoding model other than BERT. Use of a “model with knowledge for language understanding acquired by performing pre-learning using a Transformer model” such as the BERT can achieve higher accuracy. For the Transformer model, refer to Reference 2 below.
Reference 2
A. Vaswani, N. Shazeer, N. Parmar, and J. Uszkoreit. Attention is All You Need, arXiv preprint arXiv: 1706.03762, 2017. When theBERT described in Reference 1 above is employed as a model of the neural network that implements the question encoding unit 110, the question encoding unit 110 combines the document P and the current question Qi with a separator token [SEP] and inputs the combination to the model. Thereby, features uit (where t=1, . . . , T) of the document P conditioned on the current question Qi are obtained.
The model of the neural network that implements the question encoding unit 110 is not limited to the BERT. Any model can be employed as a model of the neural network that implements the question encoding unit 110 as long as the model of the neural network is a model that takes a document P and a current question Qi as inputs and calculates features uit (where t=1, . . . , T) of the document P conditioned on the current question Qi.
Step S102: The context encoding unit 120 takes the document P, the question history Qi-1, . . . , Qi-k, and the answer history {Ai-1, . . . , Ai-k} as inputs and calculates features {ui-1t, . . . , ui-kt} (where t=1, . . . , T) of the document P conditioned on the question history {Qi-1, . . . , Qi-k} and features {vi-1t, . . . , vi-kt} (where t=1, . . . , T) of the document P conditioned on the answer history {Ai-1, . . . , Ai-k} using the learned model parameters.
Here, for example, the BERT described in Reference 1 above can be employed as a model of the neural network that implements the context encoding unit 120. When the BERT is employed as a model of the neural network that implements the context encoding unit 120, the context encoding unit 120 includes a model that encodes the question history {Qi-1, . . . , Qi-k} and a model that encodes the answer history {Ai-1, . . . , Ai-k}. Thus, in this case, the context encoding unit 120 combines each question in the question history and the document P with a separator token [SEP] and inputs the combination to the model that encodes the question history {Qi-1, . . . , Qi-k}. Similarly, the context encoding unit 120 combines each answer in the answer history and the document P with a separator token [SEP] and inputs the combination to the model that encodes the answer history {Ai-1, . . . Ai-k}.
The model of the neural network that implements the context encoding unit 120 may be an encoding model other than the BERT. Use of a “model with knowledge for language understanding acquired by performing pre-learming using a Transformer model” such as the BERT can achieve higher accuracy. For the Transformer model, refer to Reference 2 above.
More specifically, for example, when encoding a past question Qi-j (where j|{1, . . . , k}), the context encoding unit 120 combines the past question Qi-j and the document P with a separator token [SEP] and inputs the combination to the model that encodes the question history {Qi-1, . . . , Qi-k}. Thereby, features {ui-j1, . . . , ui-jT} of the document P conditioned on the past question Qi-j are obtained.
Similarly, for example, when encoding a past answer Ai-j (where j∈{1, . . . , k}), the context encoding unit 120 combines the past answer Ai-j and the document P with a separator token [SEP] and inputs the combination to the model that encodes the answer history {Ai-1, . . . , Ai-k}. Thereby, features {vi-j1, . . . , vi-jT} of the document P conditioned on the past answers Ai-j are obtained.
In this manner, features {ui-1t, . . . , ui-kt} (where t=1, . . . , T) of the document P conditioned on the past questions Qi-j (where j=1, . . . , k) and features {vi-1t, . . . , Vi-kt} (where t=1, . . . , T) of the document P conditioned on the past answers Ai-j (where j=1, . . . , k) are calculated.
The model of the neural network that implements the context encoding unit 120 is not limited to the BERT and any model can be employed as long as it is a model that takes a document P, a question history {Qi-1, . . . , Qi-k}, and an answer history {Ai-1, . . . , Ai-k} as inputs and calculates features {ui-1t, . . . , ui-kt} (where t=1, . . . , T) of the document P conditioned on the question history {Qi-1, . . . Qi-k} and features {vi-1t, . . . , vi-kt} (where t=1, . . . , T) of the document P conditioned on the answer history {Ai-1, . . . , Ai-k}.
Step S103: The context combining unit 130 takes the features {ui1, . . . , uiT} calculated by the question encoding unit 110 and the features {ui1, . . . , uiT} and {vi-1t, . . . , vi-kt} (where t=1, . . . , T) calculated by the context encoding unit 120 as inputs and calculates features ot (where t=1, . . . , T) using the trained model parameters. Specifically, the context combining unit 130 calculates a feature ot for each t=1, . . . , T as ot=[uit; ui-1t; . . . ; ui-kt; vi-1t; . . . ; vi-kt]. Here, the operator “;” represents concatenation across dimension. As a result, features ot (where t=1, . . . , T) are calculated. In this case, each feature ot (where t=1, . . . , T) is a d×(1+2k)-dimensional vector.
Here, any model can be employed as a model of the neural network that implements the context combining unit 130 as long as the model of the neural network is a model that can take the features {ui1, . . . , uiT} and the features {ui-1t, . . . , ui-kt} and {vi-1t, . . . , vi-kt} (where t=1, . . . , T) as inputs and calculate features ot (where t=1, . . . , T) that reflect the current question Qi, the question history {Qi-1, . . . , Qi-k}, and the answer history {Ai-1, . . . , Ai-k}. Specifically, for example, a recurrent neural network (RNN) or the like can be employed.
However, the context combining unit 130 does not necessarily have to be implemented by a neural network. The context combining unit 130 may perform processing of combining uit, {ui-1t, . . . , ui-kt}, and {vi-1t, . . . , vi-kt} using the operator “;” for each t=1, . . . , T.
Step S104: The answer generation unit 140 takes the features ot (where t=1, . . . , T) calculated by the context combining unit 130 as an input and generates an answer Ai in an extractive mode or a generative mode using the trained model parameters.
(1) When answer Ai is generated in extractive mode For example, a model called BiDirectional Attention Flow (BiDAF) described in Reference 3 below can be employed as a model of the natural network that implements the answer generation unit 140 when the answer Ai is generated in the extractive mode. In this case, the answer generation unit 140 is implemented by a modeling layer and an output layer of BiDAF. Specifically, the answer generation unit 140 inputs the features ot (where t=1, . . . , T) to the modeling layer of BiDAF described in Reference 3 and outputs the probability that each word in the document P is a start position of the answer Ai and the probability that each word is an end position thereof as an output of the output layer. Then, the answer generation unit 140 estimates a range in the document P corresponding to the answer Ai using a dynamic programming method based on these probabilities and then generates the answer Ai by retrieving words in the range from the document P.
Reference 3:
Min Joon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. Bidirectional Attention Flow for Machine Comprehension. In ICLR 2017.
The model of the neural network that implements the answer generation unit 140 when the answer Ai is generated in an extractive mode is not limited to BiDAF. Any model can be employed as long as it is a model that takes features ot (where t=1, . . . , T) as an input and generates an answer Ai in the extractive mode.
(2) When answer Ai is generated in generative mode For example, a model called a sequence-to-sequence attentional model described in Reference 4 below or a model called a pointer-generator can be employed as a model of the neural network that implements the answer generation unit 140 when the answer Ai is generated in a generative mode. In this case, the answer generation unit 140 is implemented by an encoder and a decoder of the sequence-to-sequence attentional model or an encoder and a decoder of the pointer-generator. Specifically, the answer generation unit 140 inputs the features ot (where t=1, . . . , T) to the encoder of the sequence-to-sequence attentional model or the pointer-generator described in Reference 4 and calculates probabilities that weight words of the document P (an attention distribution described in Reference 4). Then, the answer generation unit 140 inputs the probabilities to the decoder (the decoder of the sequence-to-sequence attentional model or the pointer-generator described in the decoder reference 4), and generates an answer Ai. At this time, the answer generation unit 140 generates the answer Ai by coping words from the document P or generating words from a predetermined vocabulary set based on probabilities that weight words in the document P and probabilities that weight words in the predetermined vocabulary set (a vocabulary distribution described in Reference 4).
Reference Literature 4
Abigail See, Peter J. Liu, and Christopher D. Manning. Get to the Point: Summarization with Pointer-generator Networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1073-1083, July 2017.
The model of the neural network that implements the answer generation unit 140 when the answer Ai is generated in a generative mode is not limited to the sequence-to-sequence attentional model or the pointer-generator, and any model can be employed as long as it is a model that takes features ot (where t=1, . . . , T) as an input and generates an answer Ai in a generative mode.
The dialogue processing apparatus 10 according to the present embodiment can take the current question Qi, the document P, the question history {Qi-1, . . . Qi-k}, and the answer history {Ai-1, . . . , Ai-k} as inputs and generate an answer Ai in an extractive mode or a generative mode as described above. That is, the dialogue processing apparatus 10 according to the present embodiment can input the answer history {Ai-1, . . . , Ai-k} in which each past answer Ai-j (where j=1, . . . , k) is text and thus either a model that generates the answer Ai in the extractive mode or a model that generates the answer Ai in the generative mode can be employed as a model of the neural network that implements the answer generation unit 140.
Further, in the dialogue processing apparatus 10 according to the present embodiment, the features {ui-1t, ui-kt} (where t=1, . . . , T) of the document P regarding the question history and the features {vi-1t, . . . , vi-kt} (where t=1, . . . , T) of the document P regarding the answer history can be treated equivalently because both the question history {Qi-1, . . . , Qi-k} and the answer history {Ai-1, . . . , Ai-k} are text.
Learning Process
In the following, a process of learning model parameters of neural networks that implement the functional units included in the dialogue processing unit 100 (the question encoding unit 110, the context encoding unit 120, the context combining unit 130, and the answer generation unit 140) (a learning process) will be described with reference to
Here, a dataset of learning data used for the learning process may include, for example, a dataset (QuAC) described in Reference 5 below.
Reference 5:
Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wentau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. 2018. QuAC: Question Answering in Context. In EMNLP. ArXiv: 1808.07036. Each piece of learning data includes a current question Qi, a document P, a question history {Qi-1, . . . , Qi-k}, an answer history {Ai-1, . . . , Ai-k}, and a correct answer {circumflex over ( )}Ai. When the model of the neural network that implements the answer generation unit 140 is a model that generates the answer Ai in an extractive mode, the correct answer {circumflex over ( )}Ai is expressed by a pair of start and end positions of a range corresponding to the correct answer in the document P. On the other hand, when the model of the neural network is a model that generates the answer Ai in a generative mode, the correct answer {circumflex over ( )}Ai is expressed by one or more words representing the correct answer.
Steps S201 to S204 of
Step S205: The update unit 200 takes the correct answer {circumflex over ( )}Ai and the answer Ai generated by the dialogue processing unit 100 as inputs and updates the model parameters using a known supervised learning method. When the model of the neural network that implements the answer generation unit 140 is a model that generates the answer Ai in an extractive mode, the update unit 200 calculates an error between the answer Ai and the correct answer {circumflex over ( )}Ai using the start and end positions used when generating the answer Ai.
The above step S205 is executed, for example, for each mini-batch composed of a predetermined number of pieces of learning data.
In the above manner, the dialogue processing apparatus 10 according to the present embodiment can train the model parameters of the neural networks that implement the functional units included in the dialogue processing unit 100.
Experimental Results
Next, the results of an experiment performed using the dialogue processing apparatus 10 according to the present embodiment for comparison with the related art will be described. Datasets used were the following three extractive types of QuAC (datasets used for machine reading comprehension in an extractive mode).
The learning set was used only for training the model parameters.
In the dialogue processing apparatus 10 according to the present embodiment, the question encoding unit 110 and the context encoding unit 120 were implemented by BERT and the answer generation unit 140 was implemented by a modeling layer and an output layer of BiDAF. The following were BERT setting conditions.
The maximum sequence length is a word length allowed to be input to the BERT. The slide width is used when inputting a word sequence having a length exceeding the maximum sequence length into the BERT. The word sequence is input to the BERT while sliding the word sequence by the slide width. In this experiment, the maximum sequence length is 384 tokens and the slide width is 128 tokens. Therefore, when a word sequence exceeding 384 tokens is input to the BERT, word sequences are input to the BERT such that the word sequences before and after the slide overlap partially.
Here, Table 1 below shows a comparison result of F-meaures (an F-measure when the development set was used and an F-measure when the test set was used) between the method of the related art described in Non Patent Document 1 (FlowQA) and the method of the dialogue processing apparatus 10 according to the present embodiment (the method of the present embodiment).
As shown in the above comparison results, the method of the present embodiment has higher F-measures in both the development and test sets than the method of the related art. Thus, it can be seen that the method of the present embodiment can generate an appropriate answer with higher accuracy than the method of the related art.
Hardware Configuration of Dialogue Processing Apparatus 10
Finally, a hardware configuration of the dialogue processing apparatus 10 according to the present embodiment will be described with reference to
As illustrated in
The input device 301 is, for example, a keyboard, a mouse, or a touch panel. The display device 302 is, for example, a display. The dialogue processing apparatus 10 may not include at least one of the input device 301 and the display device 302.
The external I/F 303 is an interface with an external device. The external device includes a recording medium 303a and the like. The dialogue processing apparatus 10 can perform reading and writing from and to the recording medium 303a or the like via the external I/F 303. Examples of the recording medium 303a include a compact disc (CD), a digital versatile disc (DVD), a secure digital (SD) memory card, and a universal serial bus (USB) memory card. The recording medium 303a may record one or more programs that implement each functional unit included in the dialogue processing apparatus 10 (for example, the dialogue processing unit 100 and the update unit 200).
The RAM 304 is a volatile semiconductor memory that temporarily holds programs and data. The ROM 305 is a non-volatile semiconductor memory that can hold programs and data even when power is off.
The processor 306 is an arithmetic unit such as a central processing unit (CPU) or a graphics processing unit (GPU) and reads a program or data from the ROM 305 or the auxiliary storage device 308 and loads the program or data to the RAM 304 to execute processing. Each functional unit included in the dialogue processing apparatus 10 is implemented, for example, by a process of causing the processor 306 to execute one or more programs stored in the auxiliary storage device 308. The dialogue processing apparatus 10 may include both the CPU and the GPU as the processor 306 or may include only one of the CPU and the GPU.
The communication I/F 307 is an interface for connecting the dialogue processing apparatus 10 to a communication network. One or more programs that implement each functional unit included in the dialogue processing apparatus 10 may be acquired (downloaded) from a predetermined server device or the like via the communication I/F 307.
The auxiliary storage device 308 is, for example, a hard disk drive (HDD) or a solid state drive (SSD) and is a non-volatile storage device that stores programs and data. The programs and data stored in the auxiliary storage device 308 include, for example, an OS and one or more programs that implement each functional unit included in the dialogue processing apparatus 10.
The dialogue processing apparatus 10 according to the present embodiment can realize the dialogue process and the learning process described above by having the hardware configuration illustrated in
The present invention is not limited to the above embodiment disclosed in detail and various changes and modifications can be made without departing from the description of the scope of claims.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2019/021142 | 5/28/2019 | WO | 00 |
