A Implicit Discourse Relation Recognition Model Based On Convolutional Neural Network

The discourse relationship is a logical relationship between the two units of text.The implicit discourse relationship recognition is of particular concern to researchers because that it is difficulty in identifying and its high proportion in natural language.The semantic information,the information focus,the grammatical information and the features of the sentence all play an important role in the implicit discourse recognition.The model use feature engineering can only get part information of sentence,it is difficult to obtain information focus and so on.Therefore,this paperconstructs a new implicit discourse recognition model by using convolution neural network which can study the abstract features from input data.Our model can learn some important information for the implicit discourse relationship recognition,such as semantic information,information focus and abstract features,therefore our model has good performance.Experiment result shows that the model increases 4%in PDTB.The main research work in this paper includes the following parts:(1)Study the models based on traditional feature engineering and word vector representation.The model uses the classical features as the baseline of this paper.We also explore the effects of different word vectors and word vector dimensions on model.(2)Study the effects of several sentence models.Respectively using bag-of-words,serialization model and tree model to construct implicit discourse relationship recognition models.We found that semantic information,grammar information and information focus play an important role in implicit discourse recognition from experiment results.The model only uses the sentence vector with special information is not good,it requires additional information such as features,word vectors to improve model effects.(3)In order to overcome the shortcomings of several sentence models in(2),we propose a new implicit discourse relationship recognition model based on convolution neural network.The model combines the advantages of bidirectional LSTM,convolution neural network and K-Max.Pooling,which can learn semantic information,information focus and key abstract features from input data.The model is about 4%higher than the previous results.