A Probabilistic Model to Determine the Coherence of Texts in Interactive Question Answering Systems



Evaluation plays an important role in interactive question answering systems like many computational linguistics fields. The coherence between the questions and the answers exchanged between the user and the system is one of the important criteria in evaluating these systems. In this paper, a new approach to determine the degree of coherence of generated text by the IQA systems is presented. The proposed model is a probabilistic model in which for feature extraction, the similarity between different N-grams is derived based on four defined criteria. Then using a prediction of the best density function among the 18 functions considered for each feature, a model for determining the coherence is selected. The results of implementation on two databases provided by several interactive question answering systems indicate that the proposed probabilistic model is highly adapted and its accuracy in determining the degree of coherence in the conversation text has been made. The Kolmogorov-Smirnov, Anderson, Darling and Cramer van Meys trials were used to matching or non-matching probability density function. According to the presented results, the probability density factor with the least error was the best performance in determining the coherence of each conversation.