This research has been supported by Ministerio de Ciencia e Innovation - Spanish Government (grant no. TIN2009-13391-C04-01), and Consel-leria d'Educacion-Generalitat Valenciana (grant no. PROMETEO/2009/119 and ACOMP/2010/ 288).

Recent years have marked the beginning and ex­pansion of the Social Web, in which people free­ly express and respond to opinion on a whole variety of topics. Moreover, at the time of taking a decision, more and more people search for in­formation and opinions expressed on the Web on their matter of interest and base their final deci­sion on the information found (Pang and Lee, 2008). Nevertheless, the high quantity of data that has to be analysed imposed the development of specialized Natural Language Processing (NLP) systems that automatically extract, classi­fy and summarize the opinions available on the web on different topics. Research in this field, of opinion mining (sentiment analysis), has ad­dressed the problem of extracting and classifying opinions from different perspectives and at dif­ferent levels, depending on various factors. While determining the overall opinion on a mov­ie is sufficient for taking the decision to watch it or not, when buying a product, people are inter­ested in the individual opinions on the different product characteristics. Especially in this con­text, opinion mining systems are confronted with a difficult problem: the fact that the adjectives used to express opinion have different polarities depending on the characteristic they are men­tioned with. For example, "high price" is nega­tive, while "high resolution" is positive. There­fore, specialized methods have to be employed to correctly determine the contextual polarity of such words and thus accurately assign polarity to the opinion.

This is the aim of the SemEval 2010 Task 18 -Disambiguation of Sentiment Ambiguous Adjec­tives (Wu and Jin, 2010). In the following sec­tions, we first present state-of-the art approaches towards polarity classification of opinions, sub­sequently describing our approach in the SemEv-al task. Finally, we present the results we ob­tained in the evaluation and our plans for future work.

Subjectivity analysis is defined by (Wiebe, 1994) as the "linguistic expression of somebody's opi­nions, sentiments, emotions, evaluations, beliefs and speculations". Sentiment analysis, on the other hand, is defined as the task of extracting, from a text, the opinion expressed on an object (product, person, topic etc.) and classifying it as positive, negative or neutral. The task of senti­ment analysis, considered a step further to sub­jectivity analysis, is more complex than the lat­ter, because it involves an extra step: the classifi­cation of the retrieved opinion words according to their polarity. There are a series of techniques that were used to obtain lexicons of subjective words - e.g. the Opinion Finder lexicon (Wilson et al., 2005) and opinion words with associated polarity. (Hu and Liu, 2004) start with a set of seed adjectives ("good" and "bad") and apply synonymy and antonymy relations in WordNet. A similar approach was used in building Word­Net Affect (Strapparava and Valitutti, 2004), starting from a larger set of seed affective words, classified according to the six basic categories of emotion (joy, sadness, fear, surprise, anger and disgust) and expanding the lexicon using paths in WordNet. Another related method was used in the creation of SentiWordNet (Esuli and Sebas-tiani, 2005), using a set of seed words whose po­larity was known and expanded using gloss simi­larity. The collection of appraisal terms in (Whi-telaw et al., 2005), the terms also have polarity assigned. MicroWNOp (Cerini et al., 2007), another lexicon containing opinion words with their associated polarity, was built on the basis of a set of terms extracted from the General Inquirer lexicon and subsequently adding all the synsets in WordNet where these words appear. Other methods built sentiment lexicons using the local context of words. (Pang et al., 2002) built a lex­icon of sentiment words with associated polarity value, starting with a set of classified seed adjec­tives and using conjunctions ("and") disjunctions ("or", "but") to deduce orientation of new words in a corpus. (Turney, 2002) classifies words ac­cording to their polarity on the basis of the idea that terms with similar orientation tend to co-occur in documents. Thus, the author computes the Pointwise Mutual Information score between seed words and new words on the basis of the number of AltaVista hits returned when querying the seed word and the word to be classified with the "NEAR" operator. In our work in (Balahur and Montoyo, 2008a), we compute the polarity of new words using "polarity anchors" (words whose polarity is known beforehand) and Nor­malized Google Distance (Cilibrasi and Vitanyi, 2006) scores. Another approach that uses the po­larity of the local context for computing word polarity is (Popescu and Etzioni, 2005), who use a weighting function of the words around the context to be classified.

In the SemEval 2010 Task 18, the participants were given a set of contexts in Chinese, in which 14 dynamic sentiment ambiguous adjectives are selected. They are: ^jbig, /h|small, #|many, ;P |few, rftjhigh, {Ë|low, J?|thick, #|thin, ^jdeep, ^shallow, J!|heavy, g|light, |ÊL;^|huge, J!;^ |grave. The task was to automatically classify the polarity of these adjectives, i.e. to detect whether their sense in the context is positive or negative. The contexts were given in two forms: as plain text, in which the adjective to be classified was marked; in the second for, the text was tokenized and the tokens were tagged with part of speech (POS). There was no training set provided.

Our approach uses a set of opinion mining re­sources and an opinion mining system that is implemented to work for English. This is why, the first step we took in our approach was to translate the given contexts into English using the Google Translator1. In order to perform this task, we first split the initial file into 10 smaller files, using a specialized program - GSplit3 2. The OpAL adjective polarity disambiguation system combines supervised methods with unsu-pervised ones. In order to judge the polarity of the adjectives, it uses three types of judgments. The first one is the general polarity of the con­text, determined by our in-house opinion mining system - based on SVM machine learning on the NTCIR data and the EmotiBlog (Boldrini et al., 2009) annotations and different subjectivity, opi­nion and emotion lexica (Opinion Finder, Mi­croWordNet Opinion, General Inquirer, Word­Net Affect, emotion triggers (Balahur and Mon-toyo, 2008b). The second one is the local polari­ty, given by the highest number of results ob­tained when issuing queries containing the clos­est noun with the adjective to be disambiguated followed by the conjunction "AND" and a prede­fined set of 6 adjectives whose polarity is nonambiguous - 3 positive - "positive", "beautiful", "good" and 3 negative - "negative", "ugly", "bad". An example of such queries is "price high and good". The third component is made up of rules, depending on the presence of specific modifiers in a window of 4 words before the ad­jective. The final verdict is given based on the vote given by the majority of the three compo­nents, explained in detail in the next sections:

Table 1 and Table 2 present the results obtained by the OpAL system in the SemEval 2010 Task 18 competition. The system ranked fifth, with a Micro accuracy of 0.76037 and sixth, with a Ma­cro accuracy of 0.7037.

In this article we presented our approach towards the disambiguation of polarity ambiguous adjec­tives depending on the context in which they ap­pear. The OpAL system's run was based on three subcomponents working in English - one assess­ing the overall polarity of the context using an opinion mining system, the second assessing the local polarity using Google queries formed by expressions containing the noun present in the context before the adjective to be classified and the third one evaluating contextual polarity based on the adjective's default value and the modifiers around it. The final output is based on the vote given by the majority of the three components. The approach had a good performance, the OpAL system run ranking fifth among 16 runs. Future work includes the separate evaluation of the three components and their combination in a unique approach, using machine learning, as well as a thorough assessment of errors that are due to translation.