Chinese Knowledge Base Question Answering Based On Multi-task Learning And Unsupervised Learning

The knowledge base question answering(KBQA)task refers to a special question answering task which makes use of knowledge bases as the information source for answering the questions.With the help of KBQA models,users can access complex knowledge bases by describing their information needs in the form of natural language questions.Therefore,KBQA is of great application value.At present,there are abundant researches on KBQA.However,most of these researches are about KBQA in the English language and with open-domain English knowledge bases,which cannot be trivially adapted to KBQA in Chinese.Therefore,the KBQA models of the Chinese language still need further research.Besides,the existing KBQA models are mainly based on supervised learning,which means they need to rely on a certain amount of manually labeled question-answer pairs as training data,which limits the application of KBQA in low-cost small-domain scenarios.In response to the above problems,we have conducted the following three parts of research:(1)Improving the state-of-the-art model of Chinese KBQA.The state-of-the-art model of Chinese KBQA,which is named BB-KBQA,decomposes the KBQA task into three sub-tasks: question main entity recognition,entity linking,and relation prediction.We point out that the independent training of the three sub-tasks does not take advantage of the fact that these three sub-tasks are deeply related.And BB-KBQA models question main entity recognition as a sequence labeling task,which lacks explicit constraint of the uniqueness of the main entity in one question.We propose to use multi-task learning with shared parameters and pointer-network to solve those two problems.(2)Proposing an unsupervised relation prediction model based on deep metric learning.The BB-KBQA model mentioned above requires human-labeled question-answer pairs to train the question main entity recognition sub-model,which is difficult to obtain.We propose an unsupervised relation prediction model based on deep metric learning to solve the relation prediction sub-task without question-answer pairs.The model uses the text of the main entity and the candidate predicates as meta-information to build typical samples,and leverages a pre-trained language model to construct semantic representations for both the questions and the typical samples.We also propose a special semantic similarity metric function to evaluate the similarity between the questions and the typical samples.Experiments show that our model achieves good performance,which could be used in the baseline KBQA model to alleviate the dependency of human-labeled data.(3)Proposing an end to end unsupervised KBQA model based on language representation.This model uses all the triplets in the knowledge base to construct typical samples,and predict the answer of a question by finding the most similar typical sample to the question in the language representation space.It leverages approximate nearest neighborhood(ANN)technology to achieve the effect of computing semantic similarity for the question and every typical samples in the knowledge base with low computing cost,which means the model is scalable.This model does not depend on any form of human-labeled training data,making it cost less and have a wide range of application scenarios.