Machine Learning for Detecting Pronominal Anaphora Ambiguity in NL Requirements

Automated or semi-automated analysis of requirements specification documents, expressed in Natural Language (NL), has always been desirable. An important precursor to this goal is the identification and correction of potentially ambiguous requirements statements. Pronominal Anaphora ambiguity is one such type of pragmatic or referential ambiguity in NL requirements, which needs attention. However, identification of such ambiguous requirements statements is a challenging task since the count of such statements is relatively lower. We present a solution to this challenge by considering the task as that of a classification problem to classify ambiguous requirements statements having pronominal anaphora ambiguity from a corpus of potentially ambiguous requirements statements with pronominal anaphora ambiguity. We show how a classifier can be trained in semi-supervised manner to detect such instances of pronominal anaphoric ambiguous requirements statements. Our study indicates a recall of 95% with Bayesian network classification algorithm.