Compositional Matrix-Space Models for Sentiment Analysis

摘要: We present a general learning-based approach for phrase-level sentiment analysis that adopts an ordinal scale and is explicitly compositional in nature. Thus, we can model the effects required accurate assignment of sentiment. For example, combining adverb (e.g., "very") with positive polar adjective "good") produces phrase ("very good") increased polarity over alone. Inspired by recent work on distributional approaches to compositionality, each word as matrix combine words using iterated multiplication, which allows modeling both additive multiplicative semantic effects. Although multiplication-based matrix-space framework has been shown be theoretically elegant way composition (Rudolph Giesbrecht, 2010), training such models done carefully: optimization non-convex requires good initial starting point. This paper presents first algorithm learning composition. In context task, our experimental results show statistically significant improvements performance bag-of-words model.