Dialogue processing apparatus, learning apparatus, dialogue processing method, learning method and program
A dialogue processing apparatus includes one or more computers each including a memory and a processor configured to receive a question Q i as a word string representing a current question in an interactive machine reading comprehension task, a question history {Q 1 , . . . , Q i−1 } as a set of word strings representing previous questions, and an answer history {A 1 , . . . , A i−1 } as a set of word strings representing previous answers to the previous questions, and use a pre-learned first model parameter, to generate an encoded context vector reflecting an attribute or an importance degree of each of the previous questions and the previous answers; and receive a document P to be used to generate an answer A i to the question Q i and the encoded context vector, and use a pre-learned second model parameter, to perform matching between the document and the previous questions and previous answers, to generate the answer.
1. A learning apparatus comprising:
one or more computers each including a memory and a processor configured to:
receive, as an input, a first input word string obtained by connecting a question Q i as a word string representing a current question in an interactive machine reading comprehension task, a question history {Q 1 , . . . , Q i−1 } as a set of word strings representing previous questions in the interactive machine reading comprehension task, and an answer history {A 1 , . . . , A i−1 } as a set of word strings representing previous answers to the previous questions in the interactive machine reading comprehension task or a second input word string obtained by connecting word strings representing individual sentences to be categorized in a class categorization task, and use a first model parameter, to generate an encoded context vector u of a context represented by the first input word string or the second input word string;
receive, as inputs, a document P to be used to generate an answer A i to the question Q i and the encoded context vector u generated from the first input word string, and use a second model parameter, to perform matching between the document P and the previous questions and previous answers included in the first input word string, to generate the answer A i to the question Q i ;
receive, as an input, the encoded context vector u generated from the second input word string, and use a third model parameter, to estimate a class to which each of the sentences belongs;
use an error between the answer A i and a correct answer to the question Q i , to perform first updating to update the first model parameter and the second model parameter; and
use an error between the class and a correct answer class to which each of the sentences to be categorized belongs, to perform second updating to update the first model parameter and the third model parameter.
2. The learning apparatus according to claim 1 , wherein
the memory and the processor are further configured to use a first learning data set including learning data items each including any of the sentences to be categorized and the correct answer class to which the sentence belongs, to estimate the class and perform the second updating, and
each of the sentences to be categorized and the correct answer class correspond to each of sentences included in an original text and an importance degree indicating whether the sentence is included in a summary, and are produced from a data set including the original text and the summary or/and correspond to each of sentences included in a dialogue including the one or more sentences and an attribute indicating a topic in the dialogue or details of the dialogue, and are produced from a data set including the dialogue and a label indicating the topic therein or the details thereof.
3. The learning apparatus according to claim 2 , wherein
the memory and the processor are further configured to:
divide each of the first learning data set and a second learning data set to be used to learn the interactive machine reading comprehension task into a plurality of batches to produce a batch set including the plurality of batches;
determine, for each of the batches included in the batch set, whether the batch is a batch resulting from the division of the first learning data set or a batch resulting from the division of the second learning data set;
when the batch is determined to be the batch resulting from the division of the first learning data set, receive a second input word string obtained by connecting word strings representing individual sentences included in the learning data items included in the batch as an input, and generate the encoded context vector u; and
when the batch is determined to be the batch resulting from the division of the second learning data set, receive a first input word string obtained by connecting the question Q i , the question history {Q 1 , . . . , Q i−1 }, and the answer history {A 1 , . . . , A i−1 } which are included in each of the learning data items included in the batch as an input, and generate the encoded context vector u.
4. A learning method executed by a computer, the method comprising:
receiving, as an input, a first input word string obtained by connecting a question Q i as a word string representing a current question in an interactive machine reading comprehension task, a question history {Q 1 , . . . , Q i−1 } as a set of word strings representing previous questions in the interactive machine reading comprehension task, and an answer history {A 1 , . . . , A i−1 } as a set of word strings representing previous answers to the previous questions in the interactive machine reading comprehension task or a second input word string obtained by connecting word strings representing individual sentences to be categorized in a class categorization task, and using a first model parameter, to generate an encoded context vector u of a context represented by the first input word string or the second input word string;
receiving, as inputs, a document P to be used to generate an answer A i to the question Q i and the encoded context vector u generated from the first input word string, and using a second model parameter, to perform matching between the document P and the previous questions and previous answers included in the first input word string, to generate the answer A i to the question Q i ;
receiving, as an input, the encoded context vector u generated from the second input word string, and using a third model parameter, to estimate a class to which each of the sentences belongs;
using an error between the answer A i and a correct answer to the question Q i , to update the first model parameter and the second model parameter; and
using an error between the class and a correct answer class to which each of the sentences to be categorized belongs, to update the first model parameter and the third model parameter.