.png){kind=icon}
.png){kind=icon}
Return
Knowledge-based question answering by tree-to-sequence learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2019.09.003.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, many knowledge bases have been constructed or populated. These knowledge bases link real-world entities by their relationships on a large scale, serving as good resources to answer factoid questions. To answer a natural language question using a knowledge base, the main task is mapping it to a structured query of the same meaning, whose results from the knowledge base will be used as the question's answers. This mapping task is non-trivial since different questions can express a same meaning and many queries can arise from a knowledge base. To fulfill the task, an important thing is to model a query's structure as it conveys a part of the meaning and affects word orders in the question. However, state-of-the-art methods based on deep learning have neglected query structures and focused only on capturing semantic correlations between a question and a simple relation chain. In this paper, we instead take a query as a tree, and encode the orders of entities and relations into its representations to better distinguish candidate queries of a given question. Overall, we first construct candidate queries of a given question from its linked entities, types and number operations. Then, we adopt an encoder-decoder framework which encodes a candidate query and decodes it into the given question, where the decoding probability is used to select the best query from candidates. During encoding, we propose a tree-based LSTM to properly model contexts of an entity or relation in a query, and one such tree-based LSTM runs from all leafs to the root while another runs reversely. During decoding, we propose to mix a generating mode and a referring mode to capture different levels of correlations between queries and questions. Experiments on two popular datasets demonstrate the effectiveness of both the tree-based encoder and the mixed-mode decoder. (C) 2019 Published by Elsevier B.V.
Keywords:
Question answering
Knowledge base
Neural network
Tree to sequence
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W