Research On Identifying Of CFN Semantic Roles Based On Distributed Word Representation

As an implementation of shallow semantic parsing, semantic role labeling has attracted more and more concentration in natural language processing fields in recent years. Moreover, outputs of semantic role labeling has been widely applied in question answering system, information extraction, machine translation and other fields. Formally, given a target word in a sentence, semantic role identification aims to identify whether a constitution in the sentence is a semantic role. Semantic role identification is one important subtask in semantic role labeling, and the other is semantic role classification.In this paper, we mainly focused on the task of semantic role identification based on the Chinese FrameNet. We used a neural network model to automatically identify semantic roles in a sentence based on distributed word representations. We further implemented a training and a testing algorithms based on neural network which can effectively employ multiple types of features.In order to derive word representations, we used three popular representation learning techniques:C&W, RNNLM and Word2Vec. All of them were developed in recent years.In this paper, we still regarded the semantic role identification task as a sequence labeling problem. Chinese characters and words were treated as two kinds of label units in our experiments respectively. When Chinese characters be used as label units, we extracted all character and their combinations as features, and integrated base chunk representations, and some other information at Chinese character-level as our features. The best F-value of our character-based model achieved 50.10%. When Chinese words be used as label units, words, part of speech(POS), locations, target words, combinations of adjacent Chinese words, combinations of adjacent POS, base chunk representations, and some other information at word-level were all integrated in our model as features. The F-value of this word-based model achieved 72.89%, which is the best results of CFN semantic role identification currently. Our results were all based on correct Chinese word segmentation beforehand.It is worth to notice that, in order to get the distributed representations of base chunks, we constucted an base chunk recognition model based on deep neural network. We furthr introduced this model to identify base chunk information of input sentences used in semantic role recognition task, and choose three hidden layers of base-chunk recognition model as their distributed representations. We concatenated these distributed base chunk representations to hidden layers of the neural networks used in semantic role identification models. The final F-value achieved 72.89%. However, F-value of model that did not used the features of base chunk represtations just achieved 72.70%. These results illustrated that impact of base chunk representations is significant.First novelty of this paper is that we implemented a neural network model which integrates multiple types of features. Second novelty is that we employed the concatenation of distributed representations of base chunk and middle layers in neural network to identify semantic role, and it further improved the state-of-the-art performance of our model.