Understanding and Reasoning with Negation
Md Mosharaf Hossain · 2022
In this dissertation, I start with an analysis of negation in eleven benchmark corpora covering six Natural Language Understanding (NLU) tasks. With a thorough investigation, I first show that (a) these benchmarks contain fewer negations compared to general-purpose English and (b) the few negations they contain are often unimportant. Further, my empirical studies demonstrate that state-of-the-art transformers trained using these corpora obtain substantially worse results with the instances that contain negation, especially if the negations are important. Second, I investigate whether translating negation is also an issue for modern machine translation (MT) systems. My studies find that indeed the presence of negation can significantly impact translation quality, in some cases resulting in reductions of over 60%. In light of these findings, I investigate strategies to better understand the semantics of negation. I start with identifying the focus of negation. I develop a neural model that takes into account the scope of negation, context from neighboring sentences, or both. My best proposed system obtains an accuracy improvement of 7.4% over prior work. Further, I analyze the main error categories of the systems through a detailed error analysis. Next, I explore more practical ways to understand the semantics of negation. I consider revealing the meaning of negation by revealing their affirmative interpretations. First, I propose a question-answer driven approach to create AFIN, a collection of 3,001 sentences with verbal negations and their affirmative interpretations. Then, I present an automated procedure to collect pairs of sentences with negation and their affirmative interpretations, resulting in over 150,000 pairs. Experimental results demonstrate that leveraging these pairs helps (a) a T5 system generate affirmative interpretations from negations in AFIN and (b) state-of-the-art transformers solve natural language understanding tasks, including natural language inference and sentiment analysis. Furthermore, I develop a plug-and-play affirmative interpretation generator that is potentially useful in improving a number of natural language understanding tasks where negation poses a challenge.