面向知识库问答中复述问句评分的词向量构建方法

Learning Word Embeddings for Paraphrase Scoring in Knowledge Base Based Question Answering

传统的词向量构建方法基于句子内部单词间的共现概率,采用与具体任务无关的无监督训练方法实现.文中提出基于复述关系约束的词向量构建方法,用于改进知识库问答中基于词向量和词袋模型的复述问句评分.首先从复述问句库中按一定规则收集得到满足复述关系的问句对和不满足复述关系的问句对,以问句对之间的相似度不等式表示句子级的语义约束信息,再将该不等式作为约束项加入词向量训练的目标函数中.实验表明,相比传统词向量构建方法,文中方法可以提高问句间复述关系评价的准确度及知识库问答系统中问题回答的准确度.

The conventional word embeddings are learned from the co-occurrence probabilities between the words within a same sentence. The learning algorithm is task-independent and unsupervised. A method for constructing word embeddings is proposed by utilizing the constraints of paraphrasing to improve the performance of paraphrase scoring with word embeddings and bag-of-words model in knowledge base （ KB） based question answering （ QA） . In the proposed method, the pairs of paraphrase questions and non-paraphrase questions are collected respectively from a database of question paraphrases according to some designed rules. Then, the inequalities describing the similarities between the pairs of questions are adopted to represent the semantic constraint at the sentence level. These inequalities are integrated into the objective function for training word embeddings. Experimental results show that the proposed method improves the accuracies of paraphrase scoring and KB-based question answering compared with conven-tional word embedding methods.