Context-enriched Natural Language Generation models

Abstract

With the advent of conversational agents or chatbots and the widespread digital dissemination of information on critically sensitive domains, such as governance and healthcare, the need to model natural language generation to target human interlocutors has become all the more important. While Natural Language Processing (NLP) models have gained substantial attention, only in recent years has research opened new paths for tackling Human-Computer Interaction (HCI) from the perspective of natural language. Among those, just a few have investigated how to develop language production models using a Human-Centered Design (HCD) perspective to build context-enriched Natural Language Generation (NLG) models from their conceptions. Drawing on the tenets of the HCD methodology, this thesis compiles three studies aimed at enriching NLG models with context in order to generate texts for successful human communication, offering a productive improvement in user comprehension and experience. The first study focuses on a particular task of NLG models known as Referring Expression Generation (REG). We address the gap in state-of-the-art REG models when dealing with entities not seen during their training, compromising the comprehension of the final text. Drawing on the results of the automatic and human evaluation and an ablation study using the WebNLG corpus, this first approach contributes to generating more meaningful referring expressions to unseen entities. The second study collaborates with the development of DaMata, a robot journalist that automatically produces daily and monthly reports about the deforestation of the Legal Amazon area. With the context enrichment of the REG component, DaMata can provide more diverse information and reach wider audiences. Moreover, DaMata's pipeline framework ensures high-fidelity data publishing and enhances user experience. Finally, in the third study, we focus on developing a human-centered NLG model within a conversational question-answering system in the healthcare domain. We created personas to model interpersonal interaction (expert and non-expert users) in that specific domain. The third approach presented competitive results with a human-centered context enrichment, giving direction to more adequate relations between humans and NLP - Human-Centered NLP. In addition, we showed that an HCD approach benefits language generation from the perspective of "machines" but mainly humans, who are the primary actors within NLP. Overall, this thesis contributes with new directions to low-resource contexts (languages other than English and sensitive domains) where the need to promote effective communication is essential.