Graph Representation Learning And Its Application To Opinion Extraction In Multi-target Scenario

Target-oriented Opinion Words Extraction is a hot topic in text data processing in recent years.It is helpful to screen out useful information from massive jumbled text data and provide information support for military intelligent decision making.The task is to extract the opinion corresponding to the given target in a sentence.For example,for the given target “box” in the sentence: “they hid the control room key in an old box.”,the corresponding opinion to be extracted is “old”.There are three deficiencies in the current study.First,the information content of the target embedded in the potential opinion representations(that is,all the possible word representations of the corresponding opinion)is simple.This is because current models generally only consider the one-way information interaction from static target informa-tion(i.e.,information that does not change with training,such as the semantic information and location information of the target itself)to potential opinion.Second,the target infor-mation fails to adjust its weight distribution adaptively in the representations of potential opinion.This is because the current model still relies on distance rules to allocate target information.Third,the current study did not optimize the model for the scenario with multiple targets in the sentence,resulting in a significant decline in model performance in the multi-target scenario.It is considered that graph representation learning based on graph neural network has outstanding advantages in modeling complex structural rela-tions and strengthening node information interaction.This paper tries to use it to solve the above shortcomings,and puts forward the following three methods.Aiming at the first and second problems,this paper tries to construct the graph rep-resentation learning network structure of two-way interaction between target and opinion information,and uses the attention mechanism on the graph to realize the dynamic adap-tive embedding of target information.For the third problem,this paper proposes two methods of progressive type.First,we use graph theory knowledge to optimize the information diffusion path of a given target,so as to enhance the distinction between given and non-given target word information and improve the pertinence of extraction in multi-target scenario.Second,we use the effective auxiliary information to further improve the performance of the model in multi-target scenario.Experimental results show that the performance of the proposed method is significantly better than that of the comparison model.