Research On Joint Information Extraction Methods In Low Resource Situations

With the rapid development of the Internet and computers,the amount of information from the network has exploded exponentially.Information extraction technology that can help people get information accurately and quickly has attracted attention.Traditional pipelined information extraction methods had the drawbacks of error accumulation and information loss between subtasks,which could not satisfy the needs when new relationships appeared in the field.Therefore,the research on joint information extraction in low-resource scenarios had strong research significance and application value.This paper studies information extraction approaches,low-resource scenario methods,and joint extraction methods.The specific contents and contributions are as follows:Firstly,this paper investigates and analyzes the relevant theoretical knowledge of joint information extraction in low-resource scenarios,including the basic principle of semantic representation encoder,the network structure of classification model,and activation function.The basic theory of contrastive learning and data augmentation methods are also studied.Secondly,a Semantic-Consistent Learning(SCL)method for joint information extraction is proposed that does not rely on external knowledge.By studying existing joint information extraction approaches in low-resource scenarios,it is found that current work heavily relies on large-scale artificial external knowledge,rather than just the corpus itself,which requires expensive human resources.Therefore,this paper proposes Semantic Consistency Learning(SCL)for joint information extraction,which takes full advantage of the semantic consistency in different scenarios to guides the extraction of new relational triplets based on only a single sample instance of each class.Experiments show that this method has better performance in low-resource scenarios than the baseline models.Finally,in order to optimize the Semantic Consistency Learning Method,this paper presents a template sentence evaluation method based on instance-level contrastive learning encoder and class-level contrastive learning encoder.This method models the intra-class consistency and cross-class discrimination of samples to optimize the selection of template sentences in the Semantic Consistency Learning Method.Experiments show the great performance and the semantic expression ability of the encoder and prove the rationality and validity.The joint extraction method proposed in this paper has better performance than the baseline models in low resource scenarios,and has competitive application value.