Adding Redundant Features for CRFs-based Sentence Sentiment Classification

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 117–126, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Adding Redundant Features for CRFs-based Sentence Sentiment Classification Jun Zhao, Kang Liu, Gen Wang National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China {jzhao, kliu, gwang}@nlpr.ia.ac.cn Abstract In this paper, we present a novel method based on CRFs in response to the two special characteristics of “contextual dependency” and “label redundancy” in sentence sentiment classification. We try to capture the contextual constraints on sentence sentiment using CRFs. Through introducing redundant labels into the original sentimental label set and organizing all labels into a hierarchy, our method can add redundant features into training for capturing the label redundancy. The experimental results prove that our method outperforms the traditional methods like NB, SVM, MaxEnt and standard chain CRFs. In comparison with the cascaded model, our method can effectively alleviate the error propagation among different layers and obtain better performance in each layer. 1 Introduction* There are a lot of subjective texts in the web, such as product reviews, movie reviews, news, editorials and blogs, etc. Extracting these subjective texts and analyzing their orientations play significant roles in many applications such as electronic commercial, etc. One of the most important tasks in this field is sentiment * Contact: Jun ZHAO, jzhao@nlpr.ia.ac.cn classification, which can be performed in several levels: word level, sentence level, passage level, etc. This paper focuses on sentence level sentiment classification. Commonly, sentiment classification contains three layers of sub-tasks. From upper to lower, (1) Subjective/Objective classification: the subjective texts are extracted from the corpus teeming with both subjective and objective texts. (2) Polarity classification: a subjective text is classified into “positive” or “negative” according to the sentimental expressions in the text. (3) Sentimental strength rating: a subjective text is classified into several grades which reflect the polarity degree of “positive” or “negative”. It is a special multi-class classification problem, where the classes are ordered. In machine learning, this kind of problem is also regarded as an ordinal regression problem (Wei Wu et al. 2005). In this paper, we mainly focus on this problem in sentiment classification. Sentiment classification in sentence level has its special characteristics compared with traditional text classification tasks. Firstly, the sentiment of each sentence in a discourse is not independent to each other. In other words, the sentiment of each sentence is related to those of other adjacent sentences in the same discourse. The sentiment of a sentence may vary in different contexts. If we detach a sentence from the context, its sentiment may not be inferred correctly. Secondly, there is redundancy among the sentiment classes, 117 especially in sentimental strength classes. For example: “I love the scenario of “No country for old man” very much!!” “This movie sounds good.” The first sentence is labeled as “highly praised” class and the second one is labeled as “something good” class. Both the sentences express positive sentiment for the movie, but …

Content truncated to 500 words