Judging method for enhancing answer quality ranking by depicting cause-effect dependent relationships and time sequence influence mechanisms

Abstract

The invention discloses a judging method for enhancing answer quality ranking by depicting cause-effect dependent relationships and time sequence influence mechanisms. The method comprises the steps of 1) using each question and its answers ranked in a timed sequence as a training data set; 2) performing unsupervised learning on text in the training data set via Paragraph2Vec model to obtain a text expression model, and constructing question and answer recessive expressions respectively; 3) introducing question-answer cause-effect dependent relationships and answer-answer time sequence influence mechanisms to a traditional long-short term memory model; 4) using an answer ranking model acquired by learning to rank candidate answers to questions based on the question and answer recessive expressions. Compared with common answer quality judging methods, the method of the invention further explores time-sequence-based interaction between answers and discovers forming law of high-quality answers. The performance in ranking answer qualities acquired herein is better than that of traditional judging methods based on text-semantic relevance.

Description

technical field [0001] The invention relates to question-and-answer text retrieval, in particular to a method for enhancing answer quality ranking by using a characterization of causal dependencies and a timing influence mechanism. Background technique [0002] Question and answer retrieval is an important technical field with practical significance, and sorting question and answer texts according to their semantic relevance is an important technology in this field. In the retrieval process, this technology sorts the relevance of the question and each candidate answer, and displays the sorting result to the user, which has great value in the application of answer quality sorting. [0003] The traditional answer quality ranking method generally first learns a semantic space for the question and answer data, and then maps the question and the answer in the semantic space to form the corresponding feature vector. Afterwards, a manually-specified relevance measurement function ...

AI Technical Summary

Problems solved by technology

This method uses the feature vector in the semantic space as the representation of the question-answering data, and only considers the semantic relevance between the answers to the questions (that is, the "causal dependency" in this paper), while ignoring the relationship between candidate answers under the same question. Therefore, it is difficult to dig out the complex influence mechanism that exists between the answers

Images