Negation handling is an important sub-task in Sentiment Analysis. Negation plays a significant role in written text. Negation terms in sentence often changes the polarity of entire sentence from positive to negative or vice versa, resulting in the opposite meaning of the sentence than what is observed by the machine learning based linguistic model. As automatic opinion mining has become very important in this digital era, proper handling of negation term is the need of the hour. In any natural language negations can be formulated both explicitly or implicitly while their use is very much domain-specific. Existing negation handling techniques follow rule-based approach and mainly used in medical domain. Due to the complex syntactic structure of negation, it is hard to build general purpose machine learning based negation handling model on user review or conversational text data. In this paper, we investigate negation components i.e., cue and scope in a sentence which determine the polarity shift in sentence. We propose LSTM based deep neural network model for negation handling task where the model automatically learns the negation features from labeled input training dataset. We used ConanDoyle story corpus for model training and testing, which is pre-annotated with negation information. The proposed model first identify negation cues in each sentence and then using bidirectional LSTM extracts the relationship between cue and other words to identify scope of the cue in sentences. We derived word level features for model training to determine correct polarity of the sentence. Result shows that the LSTM based nonlinear language models perform comparatively better than the traditional state of the art SVM, HMM or CRF based models. BiLSTM achieved best result, F1 measures 93.34%, outperform traditional rule based model in negation handling task.
Full Text Finder