Research On Aspect-Opinion Pair Extraction Methods In Sentiment Analysis

Sentiment analysis tasks can help users quickly mine key opinions of reviewers from massive review data,so as to obtain decision-making basis.With more than two decades of research on various sentiment analysis tasks,sentiment analysis systems are currently widely used in many fields such as business,politics,and medicine.Aspect based sentiment analysis task is a fine-grained sentiment analysis research,which can help users to dig deeper into the comments of a reviewer on various aspects of an opinion target,rather than judging the sentimental tendency of the reviewer as a whole.Aspect extraction is a key task in aspect based sentiment analysis task,and it is also a prerequisite for executing the downstream tasks.In past research,scholars usually extract aspects and opinions separately or simultaneously without pairing,so additional pairing operation needs to be performed before judging the reviewer's sentimental tendency toward each aspect.Therefore,in recent years,some scholars have formally proposed the Aspect-Opinion Pair Extraction(AOPE)task.AOPE task aims to capture each aspect with its corresponding opinions in a review.Entity recognition and relation detection are two fundamental subtasks of AOPE.Entity recognition is responsible for extracting all aspect and opinion terms,and relation detection is responsible for pairing.Although recent works have considered interaction,their model structures show that the two subtasks are still relatively independent during calculation,and the number of interactions is not much.Furthermore,since AOPE task has only been formally proposed for two years,many AOPE models focus on the joint training of subtasks,but do not pay attention to the syntactic information providing clues for both subtasks,or their methods of introducing syntactic information still depend on the external algorithms.To address the above issues,we propose a model for Synchronously Tracking Entities and Relations(STER)to deal with AOPE task.Specifically,we design a network consisting of a bank of gated RNNs,where we can track all entities of a review sentence in parallel.STER utilizes three features,i.e.,context,syntax and relation,to learn the representation of each tracked entity without external knowledge,and synchronously calculates the correlated degree between all entities using these entity representations at each time step.The entity representation and the correlated degree are highly dependent during calculation.Finally,they will be used for entity recognition and relation detection,respectively.Therefore,in STER,the two subtasks of AOPE can achieve sufficient interaction,which enhances their mutual heuristic effect heavily.To verify the effectiveness and adaptiveness of our model,we conduct experiments on two annotation versions of SemEval datasets.The results demonstrate that STER not only achieves advanced performances but adapts to different annotation strategies well.To study the impact of different relation detection and entity labeling algorithms on the extraction performance,we also design a Convolution-based Triple Relational Features(CTRF)model for AOPE.CTRF utilizes a two-layer CNN to learn the relations between entities from triple features of dot product similarity,cosine similarity,and bilinear similarity,and then calculates the probability of two entities belonging to an aspect-opinion pair.Later,we perform entity labeling in CTRF and its baseline models that use different relation detection algorithms,adopting two methods of CRF and Softmax regression.The experimental results show that CTRF performs better than other baseline models in the extraction tasks with its relation detection algorithm,and CRF performs better in the entity labeling task compared to Softmax regression.