Sentiment analysis is always puzzled by the con­text-dependent sentiment words that one word brings positive, neutral or negative meanings in different contexts. Hatzivassiloglou and Mckeown (1997) predicated the polarity of ad­jectives by using the pairs of adjectives linked by consecutive or negation conjunctions. Turney and Littman (2003) determined the polarity of sentiment words by estimating the point-wise mutual information between sentiment words and a set of seed words with strong polarity. An-dreevskaia and Bergler (2006) used a Sentiment Tag Extraction Program to extract sentiment-bearing adjectives from WordNet. Esuli and Se-basian (2006) studied the context-dependent sen­timent words in WordNet but ignored the instances in real context. Wu et al. (2008) applied collocation plus a SVM classifier in Chinese sen­timent adjectives disambiguation. Xu et al. (2008) proposed a semi-supervised learning algorithm to learn new sentiment word and their context-dependent characteristics.

Semeval-2 Task 18 is designed to provide a common framework and dataset for evaluating the disambiguation techniques for Chinese sen­timent adjectives. The HITSZ CITYU group submitted three runs corresponding to one base­line system and one improved systems (two runs).

The baseline system (HITSZ_CITYU_3) is based on collocations between sentiment words and their targets as well as their context words. For the ambiguous adjectives, 412 positive and 191 negative collocations are built from a 100-million-word corpus as the seed collocation set. Using the context words of seed collocations as features, a one-class SVM classifier is trained in the baseline system. Using HowNet-based word similarity as clue, the seed collocations are ex­panded to improve the coverage of collocation-based technique. Furthermore, a semi-supervised learning algorithm is developed to learn new col­locations between sentiment words and their tar­gets from raw corpus. Finally, the inner sentence features, such as collocations and context words, and the inter sentence features, i.e. neighboring sentence sentiments, are incorporated to deter­mine the polarity of ambiguous adjectives. The improved systems (HITSZ_CITYU_1/2) achieved 0.957 and 0.953 macro accuracy (ranked 1st and 2nd) and 0.936 and 0.933 micro accuracy (ranked 2nd and 3rd), respectively. This result shows that collocation, context-words and neighboring sentence sentiment are effective in sentiment adjectives disambiguation.

The rest of this paper is organized as follows. Section 2 presents the collocation extraction sub­system based on lexical statistics. Section 3

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 448^-51, Uppsala, Sweden, 15-16 July 2010. ©2010 Association for Computational Linguistics

presents the baseline system and Section 4 presents the improved systems. The experiment results are given in Section 5 and finally, Section 6 concludes.

A lexical statistics-based collocation extraction subsystem is developed to identify both the bi­gram and n-gram collocations of sentiment ad­jectives. This subsystem is based on our previous research on Chinese collocation extraction. It recognizes the co-occurring words of a headword as collocations which have co-occurrence fre­quency significance among all co-occurring words and co-occurrence position significance among all co-occurring positions.

For a sentiment adjective, noted as whead, any word within the [-5,+5] context window is a co-word, denoted as wco-i for 1< i < k, where k is the total number of different co-words of whead.

BI-Strength(whead,wco_i) between a head word whead and a co-word w co-i (i=1, to k) is designed to measure the co-occurrence frequency signifi­cance as follows:

BI - Strength(wheaJ, wco-i ) = (1)

where, fmaJw>head) , fmm(whead) and f (wheaJ ) are the highest, lowest and average co-occurrence fre­quencies among all the co-words of whead,, re­spectively; fmaJw>co-,), fmiJWcod and f(wco_. ) are respectively the highest, lowest and average co­occurrence frequencies of the co-words for wco-i. The value of BI-Strength(wheadwco-]) ranges from -1 to 1, and a larger value means a stronger asso­ciation. Suppose f(whead,wco-i, m) is the frequency that wco-i co-occurs with whead at position m(-5<=m<=5). The BI-Spread(whead,wco-i) is de­signed to characterizes the significance that wco-i around whead at neighbouring places as follows:

where, f (whead ,Wco-i) , fmax(Whead,,Wco-i), and fmin (whea,,dwco-i) are the average, highest, and lowest co-occurrence frequencies among all 10 posi­tions, respectively. The value of BI-Spread(whead, wco-i) ranges from 0 to 1. A larger value means that whead and wco-i tend to co-occur in one or two positions.

The word pairs satisfying, (1) BI-Strength(whead Wco-j)>Ko and (2) BI-Spread(whead, wco-i)>U0, are extracted as bi-gram collocations, where K0 and U0 are empirical threshold.

Based on the extracted bi-gram collocations, the appearance of each co-word in each position around whead is analyzed. For each of the possible relative distances from whead, only words occupy­ing the position with a probability greater than a given threshold T are kept. Finally, the adjacent words satisfying the threshold requirement are combined as n-gram collocations.

The baseline system incorporates collocation and context words as features in a one-class SVM classifier. It consists of two steps:

STEP 1: To match a test instance containing seed collocation set. If the instance cannot be matched by any collocations, go to STEP 2.

STEP 2: Use a trained classifier to indentify the sentiment of the word.

The collocations of 14 testing sentiment adjec­tives are extracted from a 100-million-word cor­pus. Collocations with obvious and consistent sentiment are manually identified. 412 positive and 191 negative collocations are established as the seed collocation set.

We think that the polarity of a word can be de­termined by exploiting the association of its co-occurring words in sentence. We assume that, the two instances of an ambiguous sentiment adjec­tives that have similar neighboring nouns may have the same polarity. Gamon and Aue (2005) made an assumption to label sentiment terms. We extract 13,859 sentences containing collo­cations between negative adjective and targets in seed collocation set or collocations between ambiguous adjective and negative modifier (such as ÜiP too) as the training data. These sentences are assume negative. A single-class classifier is then trained to recognize negative sentences. Three types of features are used:

(1) Context features include bag of words within context in window of [-5, +5]

(2) Collocation features contain bi-grams in window [-5,+5]

(3) Collocation features contain n-grams in window [-5,+5]

In our research, SVM with linear kernel is employed and the open source SVM package -LIBSVM is selected for the implementation.

The preliminary experiment shows that the base­line system is not satisfactory, especially the coverage is low. It is observed that the seed col­location set covers 17.54% of sentences contain­ing the ambiguous adjectives while the colloca­tions between adjective and negative modifier covers only 11.28%. Therefore, we expand the sentiment adjective-target collocation set based on word similarity and a semi-supervised learn­ing algorithm orderly. We then incorporate both inner-sentence features (collocations, context words, etc.) and inter-sentence features in the improved systems for sentiment adjectives dis­ambiguation.

The ACL-SEMEVAL task 18 testing dataset contains 14 ambiguous adjectives and 2,917 in­stances. HITSZ_CITYU group submitted three runs. Run-1 and Run-2 are two runs correspond­ing to the improved system and Run-3 is the baseline system. The achieved performances are listed in Table 1.

It is observed that the improved systems achieve promising results which is obviously higher than the baseline. They are ranked 1st and 2nd in Macro Accuracy evaluation and 2nd and 3rd in Micro Accuracy evaluation among 16 submit­ted runs, respectively.

In this paper, we proposed similarity-based and semi-supervised based methods to expand the adjective-target seed collocation set. Meanwhile, we incorporate both inner-sentence (collocations and context words) and inter-sentence features in a two-class SVM classifier for the disambigua­tion of sentiment adjectives. The achieved prom­ising results show the effectiveness of colloca­tion features, context words features and senti­ment of neighboring sentences. Furthermore, we found that the neighboring sentence sentiments are important features for the disambiguation of sentiment ambiguous adjectives, which is ob­viously different from the traditional word sense disambiguation that emphasize the inner-sentence features.