ASAPP: Automatic Semantic Alignment for Phrases applied to Portuguese

Abstract

We present two distinct approaches to the ASSIN shared evaluation task where, given a collection with pairs of sentences, in Portuguese, poses the following challenges: (a)~computing the semantic similarity between the sentences of each pair; and (b)~testing whether one sentence paraphrases or entails the other. The first approach, dubbed Reciclagem, is exclusively based on heuristics computed on Portuguese semantic networks. The second, dubbed ASAPP, is based on supervised machine learning. The results of Reciclagem enable an indirect comparison of Portuguese semantic networks. They were then used as features of the ASAPP approach, together with lexical and syntactic features. After comparing our results with those in the gold collection, it is clear that ASAPP consistently outperforms Reciclagem. This happens both for European Portuguese and Brazilian Portuguese, where the entailment performance reaches an accuracy of 80.28% +- 0.019, and the semantic similarity scores are 66.5% +- 0.021 correlated with those given by humans.