Research On Frame Disambiguation Based On Neural Network

The research process of natural language processing is a transition from rule-based research to statistical-based research.During this period,the research field has achieved many brilliant achievements.However,no matter what method is used,the ultimate goal is to achieve free communication between people and machines,the so-called semantic understanding.Imagine that the process of human learning and understanding of language is always done in the corresponding environment.Based on this,in the 1970 s,Mr.Fillmore proposed the concept of frame semantic,which is to understand the words in the language.Meaning has the conceptual structure,different frameworks have different conceptual structures.The frame semantics makes the understanding of the semantics of sentences can be carried out under certain circumstances,thus avoiding the ambiguity problems in natural language to some extent and improving the accuracy of computer understanding semantics.But even with the framework,there are still ambiguities.For example,when the grand festival is coming,I want to <tgt>send a Chinese New Year greeting to Chinese friends;"Think" as the target word of this sentence,can provoke a framework There are "points of view","thinking","craving",and the meaning of sentences in different frameworks is completely different,so it is necessary to disambiguate the framework at this time.Frame disambiguation,that is,according to the context information of the target word,automatically an appropriate framework for the target word from the existing framework library.The study of frame disambiguation is usually treated as a classification problem.By analyzing the sentences with natural language processing tools,and manually extracting the features obtained after analysis,and then training a classification model with machine learning methods,the target words are finally classified into its most suitable framework.At present,in machine learning existing classification models(such as support vector machine,maximum entropy model,neural network,etc.),the neural network can approximate any function because it can approximate any function,but in some cases,It has not been applied in the frame disambiguation research.At the same time,the features extracted manually in the past are the features of part of speech and dependent syntactic relations,but lack of semantic dependence analysis.In view of the above problems,the main contents of this paper are as follows:(1)Construct a frame disambiguation model using neural networks.Because the neural network was not used in the frame disambiguation before,in order to compare the disambiguation results in the literature,this paper also uses the support vector machine to construct the frame disambiguation model;(2)When the frame disambiguation feature is selected,the semantic dependency analysis feature is added to the existing part of speech and dependent syntactic relationship.The semantic dependence analysis relationship can explore the relationship between the components in the sentence from another dimension.Its addition to existing features can provide more adequate features for frame disambiguation;The training corpus and test corpus in this paper are 7 words in the Chinese FrameNet(CFN)established by Shanxi University,1519 examples.In the experiment,the best result achieved by SVM is 80.67%,and the best result obtained by neural network is 84.42%.It is verified that neural network can be used in framework disambiguation problem and can obtain better results.At the same time,semantic dependence analysis relationship The addition of features can also increase the accuracy of frame disambiguation.