Research On Negation Scope Detection And Its Application

Negation is an important subtask of information extraction that can benefit a widerange of Natural Language Processing(NLP)tasks,e.g.,sentiment analysis,medical data mining,relation extraction,and question answering.The automatic negation detection task received a special attention these recent years in the research area,as a negation word may inverse the meaning of an affirmative statement into its opposite meaning,and thus,results in a wrong prediction or interpretation of the statement.At a first sight,dealing with negation may seem easy,but in practice,this task is much more problematic,as the negation word may interact with many other linguistic phenomena and used for many different purposes.In this thesis,we aim to contribute to the ongoing research on negation extraction in the NLP area through the development of different machine learning and neural networkbased systems,which extract the negation cue(i.e.,the negation word)and identify its corresponding scope(i.e.,the tokens in the sentence that are affected by the negation cue)in a sentence.We mainly focus on the elaboration of effective systems that rely as less as possible on hand-engineered features,to reduce the engineering time and the dependence of these systems on humans,and in the same time design powerful negation cues and scopes detectors.Then,we intend to apply one of our negation scope detection models into a sentiment analysis system to show the usefulness of a scope-based negation detection method on this kind of systems comparing to traditional negation detection techniques.In the first work of this thesis,we propose an approach to identify both negation cues and their corresponding scopes in general text domain.The proposed method considers the task as a sequence labeling problem divided into two sub-tasks: a cue detection and a scope detection.For each task in hand,we use a Conditional Random Field(CRF)model,trained over a set of non-complex features,both extracted from the semantic or the context,to identify the negation cue or the tokens that are affected by this negation cue.The features used have the ability to minimize the feature engineering efforts and to capture relevant information about the relationship between the cue and the tokens within its scope.This method outperformed all the results to date in the same domain.In the second contribution,we propose a new technique,based on recurrent neural networks and word embedding,to detect the scope of negation in review texts.This work compares the performance of different recurrent neural networks(i.e.,LSTM,BiLSTM,and GRU)and a machine learning technique(i.e.,CRF)in identifying the scope of negation.The different neural networks models use a sentence and the word vector representation of the different tokens as their inputs.The novelty of this work lies in the fact that,to the best of our knowledge,this is the first application of neural networks for negation scope detection task.At the same time,the proposed method improves the scope prediction results on the SFU review corpus,and proves that recurrent neural networks are more powerful than machine learning-based methods,and this without relying on any external feature.Additionally,in the third approach,we combine the previous recurrent neural network model with a convolutional neural network to capture different genre of information that can benefit the scope detection task and improve further the robustness of our neural network model,always without relying on external features.The recurrent neural network(i.e.,Bi-LSTM)captures automatically features related to the context from the sentence,and the convolutional neural network captures syntactic features from the shortest syntactic(constituency and dependency)path between the candidate token and the cue.The proposed method proved to be robust for negation scope detection task and beat all state-of-the-art approaches in the Biomedical domain.Finally,the last contribution of this thesis is an application of our negation cue and scope detection method into a sentiment classifier.In this work,we re-implement an existing neural network sentiment classification model,for which we replace its negation detection method by our scope-based approach.A performance comparison between the sentiment classification method using the traditional negation detection and the sentiment classification using our scope-based approach shows that the identification of negation cues and their corresponding scopes in a sentence is more effective than traditional negation detection methods.In overall,the achieved results confirm the effectiveness of all the proposed techniques used to extract the negation cues and identify their corresponding scopes in different text domains from the NLP area,and their usefulness for the sentiment classification task.