Distant Supervised Relation Extraction With Gate Mechanism And Sentence Relation

Relation extraction is one of the key basic tasks in natural language processing.Its purpose is to identify the semantic relations between two entities from natural language texts.At present,the methods of relation extraction are mainly divided into supervised,unsupervised and distant supervision.Among them,distant supervision is based on the assumption that "sentences containing the same entity pair are all expressing the corresponding relation of entity pair in the knowledge base".Distant supervision avoids time-consuming and laborious manual annotation of corpus and becomes one of the important research directions of relation extraction.Although distant supervision can quickly obtain large-scale corpus,it inevitably leads to wrong labelling problem.Although multi-instance learning and deep learning have solved the problems brought by the strong assumption of distant supervision to some extent,the existing research methods still have some deficiencies: 1)lack of effective means to filter intra-sentence noise.Not all words in the sentence are related to entities,and unrelated words obviously produce noise features.Although the introduction of dependency syntax tree can alleviate this problem to some extent,it destroys the structural information and integrity of sentences.2)The relevance between sentences in the package is ignored.Because the sentences in the package contain the same entity relation pairs,they are more or less related.At present,most researches have ignored the relation.In order to solve the above problems,we propose Distant Supervised Relation Extraction with Gate Mechanism and Sentence Relation.The gating mechanism is used to filter irrelevant features in sentences,and at the same time a more reasonable sentence weight is allocated by modeling sentence relevance,so as to obtain a more effective package-level feature representation and further improve the relationship extraction performance.The main innovations and contributions of this paper include the following aspects:(1)We propose a relation extraction method combining gating mechanisms.Aiming at the noise problem of features in distant supervision sentences,we define a new convolution operation as a gating mechanism to filter features in sentences,and introduce a soft tag mechanism as a supervision signal to be added into the gating mechanism to make the model pay more attention to entity-related features and avoid the negative impact of noise tags on feature extraction in distant supervision.(2)We introduce attention mechanism in the piece-wise pooling layer.Different weights are assigned to the segmented pooled vectors,which makes the model focus on the parts containing more effective features,thus filtering the features in sentences at the semantic level and further avoiding the negative impact of irrelevant features on the model.In addition,we also apply the soft label mechanism to the attention mechanism,making the model pay more attention to the entity itself than other parts.(3)Aiming at the problem that the existing research ignores the relevance between sentences in the package,a distant supervision relation extraction method based on sentence relevance is proposed.By modeling sentence relevance,reasonable weight is assigned to sentences in the package to obtain package level feature representation.In addition,in order to obtain better package-level features,this paper integrates the most effective features in sentences in the test phase and filters irrelevant features at the sentence level.We experiment on public datasets and the experimental results show that the performance of the model we proposed is obviously better than all baseline models,and the AUC value reaches 0.46.In addition,by setting up comparative experiments,we verify the effectiveness of gating mechanism in filtering noise features in sentences and the importance and effectiveness of merging sentence relevance.