From Semantic to Emotional Space Mitra Mohtarami 1 , Man Lan 2 , and Chew Lim Tan 1 1 Department of Computer Science, National University of Singapore; 2 Institute for Infocomm Research mitra@comp.nus.edu.sg; mlan@i2r.a-star.edu.sg; tancl@comp.nus.edu.sg Abstract This paper proposes an effective approach to model the emotional space of words to infer their Sense Sentiment Similarity (SSS). SSS reflects the distance between the words regarding their senses and underlying sentiments. We propose a probabilistic approach that is built on a hidden emotional model in which the basic human emotions are considered as hidden. This leads to predict a vector of emotions for each sense of the words, and then to infer the sense sentiment similarity. The effectiveness of the proposed approach is investigated in two Natural Language Processing tasks: Indirect yes/no Question Answer Pairs Inference and Sentiment Orientation Prediction. Introduction Sentiment analysis or opinion mining aims to enable computers to derive sentiment from human language. In this paper, we aim to address sense sentiment similarity that aims to infer the similarity between word pairs with respect to their senses and underlying sentiments. Previous works employed semantic similarity measures to estimate sentiment similarity of word pairs (Kim and Hovy 2004; Turney and Littman 2003). However, it has been shown that although the semantic similarity measures are good for relating semantically related words like "car" and "automobile" (Islam et al., 2008), but are less effective to capture sentiment similarity (Mohtarami et al., 2012). For example, using Latent Semantic Analysis (Landauer et al., 1998), the semantic similarity of "excellent" and "superior" is greater than the similarity between "excellent" and "good". However, the intensity of sentiment in "excellent" is more similar to "superior" than "good". That is, sentiment similarity of "excellent" and "superior" should be greater than "excellent" and "good". Copyright © 2013, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. This paper shows that not only semantic similarity measures are less effective, considering just the total sentiment of words (as positive or negative) is also not sufficient to accurately infer sentiment similarity between words senses. The reason is that, although the opinion words can be categorized into positive and negative sentiments with different sentiment intensity values, they carry different human emotions. For instance, consider a fixed set of emotions e = [anger, disgust, sadness, fear, guilt, interest, joy, shame, surprise] where each dimension ranges from 0 to 1. Given the above emotions, the emotion vectors and the sentiment orientation (SO) of the words "doleful", "rude" and "smashed" will be as follows (Neviarouskaya et al., 2007; Neviarouskaya et al., 2009): e(rude) =[0.2,0.4,0,0,0,0,0,0,0], SO(rude)=-0.2-0.4=-0.6 e(doleful)=[0,0,0.4,0,0,0,0,0,0], SO(doleful)=-0.4 e(smashed)=[0,0,0.8,0.6,0,0,0,0,0], SO(smashed)=-1.4 All the three words have negative sentiment and SO of "doleful" is closer to "rude" than "smashed". However, the emotional vectors indicate that "rude" only carries the emotions "anger" and "disgust", while "doleful" and "smashed" carry the same emotion "sadness". As such, considering the emotional space of words, the word "doleful" should be closer to "smashed" than "rude". This paper shows that using emotional vectors of the words is more effective than using semantic similarity measures or considering sentiment of the words (as positive or negative) to infer sense sentiment similarity. To achieve this aim, we propose a probabilistic approach by combining semantic and emotional spaces. Furthermore, we show the utility of sentiment similarity in Indirect yes/no Question Answer Pairs (IQAPs) Inference and Sentiment Orientation (SO) prediction tasks explained as follows: in Probabilistic Sense Sentiment Analysis Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence 711