Tht Impact Of Pre Suppositional Lanuguage Ie Language On Participants

"In linguistics and philosophy of language, semantics and pragmatics are presented as two complementary and intertwined aspects of meaning in language. The former is concerned with the literal context-free meaning of words and sentences whereas the latter focuses on the intended meaning, one that is context-dependent. While NLP research has focused in the past mostly on semantics, the goal of this thesis is to develop computational models that leverage this pragmatic knowledge in language that is crucial to performing many NLP tasks correctly. Through the body of work presented in this thesis, I aim at highlighting the importance of pragmatics in the context of natural language generation (NLG) and natural language understanding (NLU) systems. One overarching theme in the thesis is the notion of presupposition. A presupposition can be thought of as the collection of facts and assumptions that are not explicitly stated and taken for granted in a particular discourse. Presuppositions are ubiquitous in natural language and understanding them facilitates smooth communication. As such, studying them in the context of NLG and NLU systems is valuable. To that end, this thesis presents four fundamental contributions: First, I present a neural model for definiteness prediction (DP) -- the task of determining whether a noun phrase should be definite or indefinite. This work contrasts with previous work which relied on heavily-engineered linguistic features. Second, I evaluate how such a model plays a role in the improvement of extractive summaries. By using the decisions made by a neural DP model as a lightweight post-editing step, I show how we can obtain modified extractive summaries that are substantially preferred by human judges. Third, I present the task of adverbial presupposition triggering detection. This task focuses on detecting contexts where adverbs (e.g., ``again'') trigger presuppositions (``John ate again'' presupposes ``he ate before''). For this task, I present a neural model that uses a new attention mechanism in an RNN architecture and show that its use results in better prediction performance over a number of baselines without introducing additional parameters. In the final contribution, I focus on the popular NLU task of Natural Language Inference (NLI) which is concerned with predicting whether a statement entails another. Specifically, I look into learning models that belong to the popular family of large pre-trained transformer models and that are finetuned on NLI data that includes some instances involving pragmatic effects. I investigate how these models perform on NLI cases involving presupposition. Through a series of contrastive test cases, I show that when these models sometimes perform well on NLI cases involving presupposition, they are essentially exploiting superficial cues that might not generalize to other datasets. Finally, to improve the performance of said transformer models on NLI cases involving presupposition, I investigate how transfer learning can be used in the context of NLI to leverage tasks that are heavily based on notions of discourse and pragmatics. The models and results presented in this thesis lay the groundwork for further research that focuses on pragmatic effects in natural language generation and natural language understanding, and demonstrate the importance of focusing on these effects in the design of powerful NLP systems"--