Answer Source Searching Method Based on Density Priority Strategy.
abstract
Toward a particular natural question, the goal of answer source searching is to retrieve relevant short texts which entail the answer. Effectively and precisely acquiring answer source in a set of large-scale data enables the current reading comprehension technologies to implement open-domain answer extraction. In the research area of acquiring relevant short texts, nowadays the research on the relevance computing model which is based on Statistics-based and semantic encoding approaches has achieved significant success. However, the former is efficient but used to have a lower precision. By contrast, the latter acts as precisely, though it is time-consuming and normally costs computing resources. To address the issue, this paper proposes a density superior approach. It aims to utilize the maximum density distribution of keywords in the short texts to measure relevance between questions and answer source candidates. This paper evaluates the models using the SQuAD dataset and compares their performance. Experimental results show that the proposed approach achieves more significant performance gain than others. In addition, it shows more efficient.