基于ARSER深度匹配模型的专利检索策略研究

Research on Patent Retrieval Strategy based on ARSER Deep Matching Model

摘要:

专利匹配是专利检索工作的重要组成部分,其核心是如何快速准确地发现与查询专利相似的其他专利文件。以往专利检索研究中,专利匹配通常基于标题、短文本,或者利用TF-IDF、Word2vec等方法基于专利说明书进行浅层语义文本匹配。为充分利用专利说明书深层语义信息,该文提出了深度语义文本匹配模型ARSER(mAtching-aggRegation based on Sentence level Embedding with beRt),模型编码层利用BERT获得专利说明书向量嵌入,通过注意力机制分别得到一项专利说明书中的句子向量在另一项专利说明书中的最佳匹配向量,计算两者匹配结果作为局部匹配,最后融合局部匹配得到最终匹配结果。实验结果表明,该文提出的ARSER模型在专利匹配任务上性能优于其他基准方法。

Abstract:

Patent matching is an important part of patent retrieval. Its core is how to quickly and accurately find other patent documents similar to a given patent. In previous patent retrieval research, patent matching is usually based on short text such as title and abstract, or shallow semantic text matching based on patent specification by using TF-IDF, Word2vec and other methods. In order to make full use of the deep semantic information of the patent specification, this paper proposes a deep semantic text matching model ARSER ( mAtching-aggRegation based on Sentence level Embedding with beRt ). The patent specification embedding obtained by BERT is used as the model coding layer. Through the attention mechanism, the sentence embedding of one patent specification finds its best matching vector in another patent specification, and the matching result is calculated as the local matching. Finally, the local matching is fused to obtain the final matching result. The experimental results show that the ARSER model proposed in this paper performs better than other baseline methods in patent matching tasks.