Aggregation of Word Embedding and Q-learning for Arabic Anaphora Resolution

Abstract

In many linguistic situations, the repetitions of objects and entities are reduced to the pronoun. The correct interpretation of pronouns plays an important role in the construction of meaning. Thus, the resolution of the pronominal anaphors remains a very important task for most natural language processing applications. This paper presents a novel approach to resolve pronominal anaphora in Arabic texts. At first, we identify non-referential pronouns by using an iterative self-training SVM method. After, we resolve the antecedents by combining a Q-learning method with a Word2Vec based method. The Q-learning method seeks to optimize, for each anaphoric pronoun, a sequence of criteria choice to evaluate the antecedents and look for the best. It uses syntactic criteria as preference factors to favor candidate antecedents over others. The Word2Vec method uses the word embedding model AraVec 3.0. It provides the semantic similarity measures between antecedent word vectors and pronoun context vectors. To combine Q-learning and Word2Vec results, we use a ranking aggregation method. The resolution system is evaluated on literary, journalistic and technical manual texts. Its precision rate reaches until 80.82%.