An Exploration into the Use of Contextual Document Clustering for Cluster Sentiment Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 140–145, Hissar, Bulgaria, 12-14 September 2011. An Exploration into the Use of Contextual Document Clustering for Cluster Sentiment Analysis Niall Rooney, Hui Wang University of Ulster {nf.rooney,h.wang} @ulster.ac.uk Fiona Browne Queen’s University,Belfast f.browne@qub.ac.uk Fergal Monaghan, Jann Müller, Alan Sergeant, Zhiwei Lin, Philip Taylor SAP Research Belfast {fergal.monaghan, jann.mueller, alan.sergeant, zhiwei.lin, philip.taylor}@sap.com Vladimir Dobrynin St Petersburg State University v.dobrynin@bk.ru Abstract In this paper we consider whether the thematic document clustering approach of Contextual Document Clustering is able to capture the overall sentiment of a cluster of documents. We provide a novel mechanism to determine the sentiment of a cluster based on the latter approach and assess the approach on three data sets formed from the NY Times annotated corpus. We demonstrate that CDC does provide a strong tendency to capture the sentiment of a cluster. 1 Introduction Sentiment analysis or opinion mining is a recent area of text classification research which tries to determine the opinion that a section of text expresses. Esuli & Sebastiani (2005) describes three subtasks:  determining whether a given piece of text has a factual nature, neutral nature or whether it expresses an opinion on its material (the Subjective-Objective (SO) polarity of the text)  determining whether a given piece of text expresses a positive or negative opinion on its subject matter orientation (the Positive-Negative (PN) polarity of the text)  determining the strength of the subject matter orientation Turney & Littman (2003) make no distinction between the latter two sub-tasks and propose a measure of semantic orientation which indicate both the direction and intensity of a text. To capture this, they focus on the semantic orientation of a word which they capture by measuring the strength of association with a set of seed words (with either absolute positive or negative polarity). They propose two measures for strength of association based on point-wise mutual information and latent semantic analysis estimated from given corpora. Other mechanisms for semantic orientation have focused on the linguistic constraints on the orientation of adjectives (e.g. the word “and” usually conjoins adjectives of the same orientation) (Hatzivassiloglou & McKeown, 1997) or that synonymous words have similar orientation 140 (Esuli & Sebastiani, 2005). The aforementioned mechanisms try to give an absolute value for the orientation of a word regardless of its context of use. Wilson et al.(2009) present a two stage classification approach to determine the contextual polarity of subjective clues (words which have been part of annotated subjective expressions) in a corpus. They based this on features primarily as a consequent of local dependency relationships (parent-child) in sentences (although they do use other features mainly at a sentence level). More recent directions in a sentiment analysis for text classification have focussed on the use of unsupervised modelling approaches for text classification. Much of this work has focussed on extending topic modelling approaches such as Probabilistic Latent Semantic Indexing (Mei et al., 2007) or Latent Dirchlet Allocation (Lin & He, 2009) to incorporate the …

Content truncated to 500 words