A computational treatment of negation and speculation in natural language

点击翻译按钮获取中文摘要

Sentiment analysis (also known as opinion mining) makes use of methods in computational linguistics and text analysis to determine whether a text (headline, news article, movie review, e-mail message, blog or Twitter post) is mostly objective or subjective and, if the latter, whether positive or negative towards its subject matter. Much progress has been made in the last few years in how to represent and analyze sentiment and opinion, but the interaction of negation, speculation and opinion remains an elusive task. For instance, in the sentence 'Some may say that it's not a good movie', the apparent positive opinion in 'good' is negated by 'not' in the same phrase, but tempered again by the speculation conveyed by 'some may say'. This sentence, in fact, seems to imply that the author actually has a favourable opinion of the film. The implication is not typically detected in current sentiment analysis systems, because speculation is rarely addressed.***The long-term objectives of my research program are to develop computational models of natural language, evaluating and implementing such models so that they can be applied to different types of texts. The short-term objectives are to find methods to extract negation and speculation, and to discover their role in the expression of sentiment and opinion.***My research will address the problem of negation and speculation by using machine learning methods to automatically find these phenomena in natural language. Machine learning methods involve classifiers which learn which features and cues are reliable indicators of a particular phenomenon, in this case negation and speculation. The result will be a software module which takes a text as input, and accurately identifies where negation and speculation lie. This module will be integrated into the sentiment analysis system developed within my lab, the Semantic Orientation Calculator (SO-CAL). ***The anticipated outcomes of this research are a set of methods for identifying negation and speculation, a module that implements those methods, and an improved SO-CAL, which will be able to determine with high accuracy the sentiment expressed in language. The software will be made available to the scientific community. Potential applications include extracting opinion about products, politicians and current events, which will be of interest to companies in social media analysis, decision-makers, journalists, moderators, and the general public, in particular those who use social media regularly. In addition to the contribution to the field of sentiment analysis, detection of negation and speculation is important in many computational linguistics tasks, including text summarization and information retrieval.