ANFIS-based model for improved paraphrase rating prediction

Paraphrase rating is an important problem with very interesting applications in plagiarism detection, language translation, text summarization, question answering, web search and information retrieval. In this paper, we present an adaptive neuro-fuzzy inference system (ANFIS) based model for automatic rating of semantic equivalence of pairs of sentences. Using a corpus of human-judged sentence pairs, lexical similarity metrics are first computed. Then, a model is constructed for predicting the mean of the rates assigned by a number of human beings. The correlation with the actual ratings and the prediction errors are studied for individual metrics as well as the model output using a nonlinear logistic regression function. The experimental results showed that much higher correlations and low error rates can be achieved with the proposed method compared to those obtained with individual metrics.