融合歧义感知的检索式问答方法

Sense-Aware Retrieval-Based Question Answering via Word Ambiguity Induction

针对多义词在不同上下文中语义表达不一致的问题,提出了一个融合歧义感知的问答模型,即模型在问题-候选答案的语义匹配过程中,与外部知识源相结合,动态识别并检测出每个多义词在不同场景下的语义,并将检测到的语义信息进行特征编码后融合到语义匹配任务中,使模型能够更为准确地理解每个词的精准含义,从而做出更为精准的匹配判断.在歧义感知模型的设计上,采用基于Transformer的深度语义编码器,使其能够更加全方位地抓取到待分析歧义词以及知识源的深度语义特征,从而做出更加准确的语义消歧.在标准检索式问答数据集上(Wiki QA和TrecQA)的实验结果表明,所提出的歧义感知的问答方法能够有效融合到多个基线模型中,并捕捉到多义词在不同语境中的精准语义,使其在包含公开数据集上的问答性能MAP评估高于对应基线模型约1%,且该语义特征使得基于BERT的文本相似性匹配模型的性能优于当前先进的其它模型.

To solve the problem of inconsistent semantic expression of polysemous words in different contexts,we propose a sense-aware question-answer model.During the semantic matching process of questions and candidate answers,the model integrates with external knowledge sources to dynamically identify and detect the semantics of each polysemous word in different scenarios.The detected semantic information is encoded as features and then integrated into the semantic matching task,enabling the model to capture the exact meaning of each word and achieve better matching performance.In the design of the ambiguity perception model,we adopt a deep semantic encoder based on the Transformer,which enables it to capture more comprehensive depth semantic features of the analyzed ambiguous words and knowledge sources,making more accurate semantic disambiguation.Experimental results on standard retrieval-based Q&A datasets(WikiQA and TrecQA)demonstrate that the proposed sense-aware Q&A method can effectively be integrated into multiple baseline models,capturing the precise semantics of polysemous words in different contexts.This approach achieves a MAP evaluation performance improvement of approximately 1%compared to corresponding baselines on public datasets.Moreover,this semantic feature enables a BERT-based text matching approach to outperform other state-of-the-art models.