知识本体与词向量结合的词义相似度强化学习计算方法

Hybrid semantic similarity calculation of knowledge ontology and word vector based on reinforcement learning

从语言学上分析了基于知识本体与基于词向量的词语语义相似度计算方法的不同特点,指出两类方法进行语义计算的优点与不足,并据此提出了基于强化学习策略的混合式语义相似度计算方法。该方法不但能够解决使用单一方法无法解决的语义计算问题,而且计算结果更接近人工判定结果,与MC30人工判定值的皮尔逊相关系数达到0.917。最后指出了提升该类算法性能的途径,一方面,可将更多的语言学信息作为观测变量,使得学习算法面临的环境更接近真实语言环境;另一方面,可使用更高层次的语言学任务结果作为奖励,增加算法解决问题的能力。

This paper analyzes the different characteristics of semantic similarity calculation methods based on ontology and word vector,points out the advantages and disadvantages of the two methods,and proposes a hybrid semantic similarity calculation method based on reinforcement learning strategy.This method can not only solve the problem of semantic computing which can not be solved by a single method,but also the calculation result is closer to the result of manual judgment,and the Pearson correlation coefficient with MC30 manual judged value is 0.917.Finally,the way to improve the performance with this kind of algorithm is pointed out,on the one hand,more linguistic information can be used as observation variables to make the learning environment closer to the real language environment;on the other hand,higher level linguistic task results can be used as rewards to increase the ability of the algorithm to solve problems.