| Information extraction aims to extract structured information from massive data,and entity relationship extraction is one of the important tasks of information extraction,which usually relies on large-scale labeled samples to train models,but massive exogenous data are not easy to obtain and the labeling work is time-consuming and laborious.However,it is not easy to obtain large amount of exogenous data,and the labeling work is laborious.To address the above problems,this paper investigates two aspects of relationship extraction and triadic group extraction,and the main work is as follows:(1)To address the problem that the information of the sample itself is not fully utilized in the case of few samples,the entity self-attention module is introduced.(1)The word-level vector representation of the sentence is extracted first,and then the head and tail entity vectors are input into the self-attentive layer for weighting,so that the relational classifier can pay more attention to the semantic information of entity pairs.This method proved to be effective in compensating for the underutilization of existing information in the case of few samples.(2)The En Att Concept FERE model is proposed to address the problem that entity concepts cannot be reasonably selected in the data enhancement stage.Since the specific semantics of entities in a sentence has a direct impact on the selection of concepts,the En Att Concept FERE model compares the entity-specific semantics with each concept corresponding to the entity by adding an entity-level vector representation module,and selects the appropriate concept for entity concept enhancement.Experiments on the Few Rel dataset with three settings,5-way 1-shot,5-way 5-shot and 10-way 1-shot,achieve accuracy rates of 91.74%,93.69% and 76.78%,respectively,outperforming TD-Proto,Concept FERE and other relational extraction models.(3)The BLRel ATE model is proposed to address the problem that the relationship prototypes cannot be represented correctly and the feature capture capability of the prototype matcher is insufficient.The model is based on the task decomposition strategy of identifying relations first and then extracting head and tail entities for triad extraction.For the determination of relationship prototypes,the weighted summed representation of word vectors of sentences is used as an important component of relationship prototypes.The prototype matcher uses a CNN-Bi LSTM model for text feature extraction,a CNN network for local feature extraction,and an overlay Bi LSTM network for global semantic capture,allowing query instances with unknown labels to be matched to a more accurate support set prototype.Experiments on the Few Rel dataset achieve F1 values of 43.26% and 41.83%under the 5-way 5-shot and 10-way 10-shot task settings,and their results surpass those of joint extraction models such as Rel ATE,validating the importance of establishing effective relational prototypes and improving the feature extraction capability of the prototype matcher. |
