Research On Automatic Negative Focus Detection In Natural Language Text

In natural language text,Negation is a universal but complicated linguistic phenomenon.And it is usually used to express people's attitude towards a certain point of view.A negative expression usually contains a negative trigger word(such as "no")which reverses the semantics of the expression itself or one aspect of it.It has received considerable attention from the NLP community over the last decade,since a negated statement often carries both an explicit negative focus and implicit positive meanings.The research on negation processing is mainly based on the following three aspects:1)cue detection,which finds negative triggers or expressions in text;2)scope resolution,which determines the grammatical scope in a sentence affected by a negative cue;and 3)focus detection,which identifies the most prominently negated part in a negative statement.This paper mainly focuses on the third aspect,and the specific research contents include the following three parts:1.Negative cue detection and negative focus detection based on bidirectional LSTM and CRF fusion model.The research of negative cue mainly includes traditional methods based on vocabulary(dictionary),statistics or sequence tagging.The existing negative focus detection methods mainly focus on rule-based methods and feature engineering-based methods which rely on domain experts for template or feature design and require a lot of manpower and time.Therefore,this paper draws lessons from the successful experience of deep learning methods in various natural language processing studies,and automatically learns parameters and deep semantic feature representation through neural networks.Experiments show that this method can automatically learn effective features and outperform the best system.2.Negative focus detection based on the contextual attention mechanism.At present,the research of the negative focus mostly uses the lexical,syntactic and semantic features in the sentence to identify negative focus,but ignores the context of the sentence.However,through the analysis of the development dataset,the emphasized part of the author may have a great difference in different contexts.Therefore,this paper,based on the methods of attention mechanism and topic model,realizes the negative focus detection system based on the word-level attention mechanism and the topic-level attention mechanism.Also,we discuss the influence of the context information on the performance.The experiment shows that our approach can effectively capture the context-dependent relationship among the adjacent sentences,and achieves the best performance on negative focus detection of SEM'12 dataset,yielding a great improvement over the state-of-the-art3.Construction of a Chinese negative focus corpus.Nowadays,the research progress of the negative focus dectection on Chinese is slow,and the main reason is the lack of a Chinese-oriented corpus.Therefore,we mark the Chinese negative focus corpus based on the CNeSp corpus,which is the first corpus of the negative focus detection study for Chinese,including the contents of the three fields of financial article,product review and scientific literature.The scale of the corpus is 16841 sentences with a total of 4039 cases in which negative focus is included.At the same time,we make an analysis of the relevant statistics and use LSTM model to construct a baseline system.Also,we provide the support of the corpus resources for the Chinese-oriented researchThis paper is devoted to the research on the automatic detection of negative focus in natural language text.Firstly,a series of methods are proposed to improve the performance of related tasks effectively,and secondly,trying to promote the progress of the research in Chinese.It is hoped that this research will have certain reference value for the related research in this field in the future,and promote the development of deep natural language understanding.