Project Title:

Multilingual Sentiemnt Lexicons Using Sense Based Mapping to SentiWordNet


Prof. Dr. Ebru Akçapinar Sezer - Hacettpe University Ankara Turkey
Prof. Dr. Laura Kallmeyer - Heinrich-Heine-Universität Düsseldorf


Behzad Naderalvojoud, Alaettin Ucan and Behrang Qasemi Zadeh


This study proposes a cross-lingual distribution semantic model to create a sense-based sentiment lexicon for 5 target Languages by using bilingual movie subtitles. The key idea is to create a mapping between the target language terms and subjective synsets of SentiWordNet and to provide a sentimental relatedness between target language terms and English word senses. Our hypothesis is that the similar English and target language terms are used for expressing a particular sentiment in different movie subtitle sections. In fact, the co-occurrence of terms in the subtitle sections of a language is more likely similar to the co-occurrence of terms in the same sections of the other language. By applying a sense matching algorithm on the proposed semantic model, we find all possible SentiWordNet synsets associated with target terms. As some languages like Turkish are figurative, the use of words and expressions in their non-literal meaning is very prevalent in public. On the other hands, this kind of languages includes many homonyms, metaphors and figurative expressions that enlarge the domain of word senses. According to this fact, it is expected that many target keywords appear in many different synset. Therefore, the strength of the positive and negative target terms are computed by considering more complete word-sense inventory that causes more accurate polarity values. To evaluate the proposed lexicons for each language, we employ a context-sensitive lexicon based approach that uses recurrent neural network to learn the modifications of word polarities in composing the sentiment value of sentences. In fact, the objective is to use lexicon polarities to compute the sentence sentiment by considering the sentiment strength, intensification and negation.

For Multilingual Sentiment Lexicon click here

For current evaluation click here