Learning Dynamic User Representations and Exploiting Personalization in NLP

Abstract

Nowadays, with the improvement of hardware and computations capabilities, there has been an increase of language models size and language understanding capabilities. However, these Natural Language Processing (NLP) models often treat language as universally understood, disregarding the socio- and psycholinguistic insights about the impact of speaker and audience characteristics on communication. In the dynamic environment of social media, users not only create connections but also tailor their language to the context, their audience and their affiliation to different sociodemographic group. Hence, in the evolving landscape of language technologies, there is an increasing demand for personalized systems that can mirror a user's individual style. Recognizing this, our thesis posits that a deeper understanding of users' social and semantic networks is essential to enhancing language understanding. <br /> Central to this thesis is the development of methodologies for capturing user's context, and integrating this context into NLP models in order to enhance their performance. Furthermore, there is a diversity of perspectives within different user groups concerning various subjective topics or situations. By integrating user-specific information, our models seek to better interpret how an individual perceive an utterance, and improve the performance of text classification in subjective NLP tasks, where there is a variety of perspectives. Additionally, users interactions with a community, are influenced by the evolving topics of interest and the prevailing views within their groups. This evolving landscape posits the need for dynamic user representations, that can capture evolving aspects of user behavior and social interactions. By leveraging social and semantic graphs, we construct models that effectively encapsulate the changing nature of user behavior over time. Modeling these temporal patterns of users' interactions, can provide insights into how their opinions or behaviors change, in addition to predicting future behaviors. In general, these approaches aim to significantly augment text representation in NLP. <br /> Finally, we propose an evaluation framework across diverse tasks, such as sarcasm detection, misinformation spreading, perspective classification, and personalized language generation, to showcase the effectiveness and versatility of our approaches. Overall, our research makes a significant leap in integrating user-specific information into NLP models for a variety of tasks, paving the way for more nuanced and context-aware language processing.