Research On Event Relation Classification Based Upon Clue Mining And Feature Analysis

Event relation detection is a natural language processing technology which aims to determine the relevance and logical relation between pairwise events.The purpose of event relation detection task is to identify and decide the relation type through analyzing the feature of semantic relevance between events,which treats event as the basic semantic unit.The event relation detection contains two main tasks,event relation identification(identifying whether the event pair is related or not)and event relation type decision(deciding what relation between relevance events),respectively.We perform study and discussion on event relation detection task,and tansform it into classification problem,which can be summarized as the following three aspects:1)Tri-Training based method for event relation classification: In order to solve the problem of lacking training corpus in event relation on task,we propose a Tri-Training based method.Firstly,we train three different classifiers according to the labeled data.Then,in the collaborative training process,the new annotated event pairs used to expand each classifier are provided by the other two classifiers,that is,if the two classifiers are consistent on the predictions of the same unlabeled event pair,this event pair is considered to have a high classification confidence and will be placed into the labeled data set of the third classifier after labeling.Finally,three well-trained classifiers are used to determine the relation on test data by voting.2)Event relation classification based on Shallow Convolutional Neural Network: The collaborative training method(Tri-Training),to some extent,solves the problem caused by inadequate training data,and improves the classification performance.However,the method relies too much on classification algorithms and external corpus,and cannot explore the intrinsic interaction between events.In this section,we presents a method for event relation classification based on shallow convolution neural network.Specifically,we extract event-level and cross-event convolutional features to represent the semantic information of event and event-pair,and on which,a shallow neural network is performed to predict the relation.Overall,this method not only avoids the over-fitting problem,but also outperforms the comparison systems.3)Using visual scene relation bank to help event relation classification: At the last section of this paper,a visual scene relation bank is constructed to detect event relation.We first extract the event pairs satisfying the predefined relation template on the external document library.Moreover,we crawl Wikipedia pictures to compose visual bank required in visual secene.And then,we build a visual scene relation knowledge bank by mining the clues between event text and pictures,and which assists reasoning and prediction of event relation.The experimental results show that the proposed method can effectively improve the performance of event relations classification.