Multi-entity Sentiment Scoring

International Conference RANLP 2009 - Borovets, Bulgaria, pages 258–263 Multi-entity Sentiment Scoring Karo Moilanen and Stephen Pulman Oxford University Computing Laboratory { Karo.Moilanen | Stephen.Pulman }@comlab.ox.ac.uk Abstract We present a compositional framework for modelling entity-level sentiment (sub)contexts, and demonstrate how holistic multi-entity polar- ity scoring emerges as a by-product of composi- tional sentiment parsing. A data set of five anno- tators’ multi-entity judgements is presented, and a human ceiling is established for the challenging new task. The accuracy of an initial implemen- tation, which includes both supervised learning and heuristic distance-based scoring methods, is 5.6∼6.8 points below the human ceiling amongst sentences and 8.1∼8.7 points amongst phrases. Keywords Entity-level sentiment analysis, sentiment scoring, sentiment parsing, sentiment annotation, compositional semantics 1 Introduction The ability to detect author sentiment towards vari- ous entities in text is a fundamental goal in sentiment analysis, and holds great promise for many applica- tions. Entities, which can comprise anything from mentions of people or organisations to concrete or even abstract objects, condition what a text is ulti- mately about. Besides the intrinsic value of entity scoring, the success of document- and sentence-level analysis is also decided by how accurately entities in them can be modelled. Deep entity analysis unfortu- nately presents the most difficult challenges, be they linguistic or computational. One of the most recent developments in the area - compositional semantics - has shown potential for sentence- and expression-level analysis in both logic-oriented [11],[9] and machine learning-oriented [3] paradigms. Our goal in this paper is to further that avenue by extending it to entity-level sentiment analysis. Entity-level approaches have so far involved rela- tively shallow methods which usually presuppose some pre-given topic or entity of relevance to be classified or scored (§5.3). Other proposals have attempted specific semantic sentiment roles such as evident sen- timent HOLDERs, SOURCEs, TARGETs, or EXPERI- ENCERs (§5.2). What characterises these approaches is that only a few specific entities in text are analysed while all others are left unanalysed. While shallow approaches can capture some amount of explicitly ex- pressed sentiment, they ignore all layers of implicit sentiment pertaining to a multitude of other entities. We believe that access to these rich layers is required for deeper logical sentiment reasoning in the future. We take a different view on the problem and investi- gate the possibility of a holistic multi-entity analysis in that we make no categorical distinctions between individual entity mentions, topics, or sentiment roles of any kind. We instead refer to all base nouns simply as entity markers which may (or may not) serve the above metafunctions, and aim at classifying all such markers in sentences using a single, unified approach. For the sentence in Ex. 1, we envisage a classifier that classifies all of the bracketed entities as positive(+), neutral(N), or negative(-) (NB. / = ‘or’): (1) “Here’s the [thing](N)/(+): Other [studies](N)/(+) have found that [clergy](+), and not [psychologists](-)/(+) or other men- tal [health](+) [experts](+)/(-), are the most common [source](+)/(N) of [help](+) sought in [times](N)/(-) of psychological [distress](-).” …

Content truncated to 500 words