基于实体词语义相似度的中文实体关系抽取

Chinese entity relation extraction based on entity semantic similarity

为了探索语义相似度在中文实体关系抽取上的作用,提出由实体词在《同义词词林》中的5层编码构建成的《同义词词林》编码树和由关系实例中的实体词,各个类别中所有实体词计算相似度后求得的平均值构建成的实体词语义相似度树2种新特征,并连同已有的《同义词词林》编码、实体类型信息共4种特征探究其对抽取性能的影响。单一特征的试验中,实体类型特征效果最好,F值达到了小类84.9、大类83.2;组合特征的试验中,实体类型和《同义词词林》编码树的组合特征效果最好,大类小类的F值都比实体类型特征提高了2.5,3种组合特征性能不升反降。试验结果表明《同义词词林》编码树是对实体类型的有效补充,但过多的特征会造成信息冗余,使抽取性能下降。

In order to explore the impact of the semantic similarity on the Chinese entity relation extraction,two newfeatures were proposed,which were the ＂Tong Yi Ci Cilin ＂ code tree constructed with the entities ＇5 layer code in＂Tong Yi Ci Cilin＂and the entity semantic similarity tree constructed with the average of the semantic similarity between the entity word in relation instance and all entity words in each category of relation. The impact on the relation extraction performance of these two newfeatures together with the existing ＂Tong Yi Ci Cilin＂code feature and the entity type information feature was explored. In the cases with single features,the entity type feature got the best performance,and the F values of subtype and type were 84. 9 and 83. 2; In the cases with combination features,the combination of the entity type feature and the ＂Tong Yi Ci Cilin＂code tree feature got the best performance,the F values of both subtype and type were 2. 5 higher than the entity type feature. But the performance of three combinations features became poorer instead of better. The results showed that the＂Tong Yi Ci Cilin＂code tree was an effective supplement of the entity type information,but excessive features may result in information redundancy and poor performance.