International Journal of Computer Processing of Languages

This paper presents our work on determining the sentiment polarity of sentences and articles. Feature combinations is the first novelty of our approaches. In order to explore better feature sets, we do some data analysis of the MPQA corpus and a few experiments to evaluate combinations. Through these efforts, our approach achieves 77.0% accuracy in MPQA data set. Another novelty of this work is that Semi-Supervised Conditional Maximum Entropy (SCME) modeling is used to combine labeled and unlabeled data. Experimental results show that our approach can significantly improve performance. At sentence-level, our approach achieves 78.2% accuracy in MPQA data set, the relative improvement given by semi-supervised technique is 5.2% over the supervised method.

Song sentiment analysis attracts much attention in research areas such as acoustic signal processing (ASP) and natural language processing (NLP). The text-based efforts in the NLP community are found to be interesting. As a popular text-based solution, the lyric-based sentiment classification approach adopts the vector space model (VSM) to represent lyric text and assigns songs sentiment labels such as light-hearted and heavy-hearted. Four problems render the term-based VSM model ineffective. Firstly, many words within song lyrics contribute little to expressing sentiment, but they are equally considered as term features. Secondly, nouns and verbs being used to express sentiment are ambiguously used in natural language text. But the ambiguity is not addressed. Thirdly, negations and modifiers accompanying the sentiment keywords make particular contributions to sentiment expressing, but the contributions are not reflected. Lastly, song lyric is usually very short and data sparseness problem is serious. To address these problems, the sentiment vector space model (s-VSM) is proposed in this paper to represent song lyric document based on sentiment unit. With the s-VSM model, the support vector machines (SVM) classification algorithm is applied to perform song sentiment analysis. Experimental results show that lyric is helpful to song sentiment analysis and the s-VSM model outperforms the VSM model significantly in performing the task of song sentiment analysis.

This paper investigates sentiments in blog data using machine learning techniques such as support vector machines (SVM) and conditional random fields (CRF). Bloggers collaboratively contribute to the formation of blog corpora, which contain a rich set of icon-tagged text expressions. First, a collocation model on the expressions is used to construct a lexicon which provides keyword features for subsequent classification models. SVM and CRF classifiers use these keywords as features to assign a sentence sentiment bipolarity or arousal-valence space categories. Our methods determine a sentiment category by taking contexts into account. Experiments show that CRF classifiers outperform SVM classifiers and the performance benefits from increasing the training size of blog materials.

Stock market reports in on-line news are widely used by amateurs to make quick investment decisions. Financial analysts often give opinions about trends of stock markets based on past and present economic event indicators. These opinions commonly appear in text form and are abundant over the Internet. It is tedious and time consuming for users to browse through such text manually let alone to understand the embedded opinions. To overcome this shortcoming, automatic trend predication methods have been proposed. Under conventional methods, reports are represented using bag of words and trend prediction is treated as a 3-way trend classification problem, i.e. trend as 'up', 'down' or 'stable'. In this paper, we propose a new pattern-based opinion mining method for market trend predication. Experiments show that (1) pattern-based classification is more effective than its word-based counterpart for feature representation; and (2) opinion mining outperforms event-based classification for trend predication. The task of opinion mining gets more difficult when the users are exposed to opinions from more than one analyst. The question becomes whose opinions should he/she trust? This lays down our second research objective, i.e. to study different opinion incorporation strategies. Intuitively, one would trust the opinion supported by the majority. However, we show that on the contrary, the user is better off trusting the most credible analyst.