Discourse understanding is the task of inferring a situation model from a text: the entities, events, and propositions it describes, together with the referential, temporal, and rhetorical relations that connect them. This structure is largely latent, only sparsely signaled by the surface text and otherwise recoverable from context. This HDR takes that observation as its unifying theme. Cast as a machine-learning problem, discourse understanding challenges three standard assumptions of classical supervised learning, each relaxed by extending context in a different sense: the structural context that makes a document’s parts mutually dependent, the semantic context of world knowledge beyond the text, and the learning context of how competence is acquired from data. The three core chapters take these in turn, contributing structured prediction and global inference; knowledge-grounded and multimodal representations; and new corpora and methods for learning from limited supervision. A final chapter reframes this trajectory around large language models, which address all three challenges through scale alone yet remain opaque and far less sample-efficient than human learners, and proposes an interdisciplinary program at the interface of NLP and cognitive science, comparing these models against human performance and studying what they learn under cognitively plausible constraints.
defended on 28/09/2026
Discourse understanding is the task of inferring a situation model from a text: the entities, events, and propositions it describes, together with the referential, temporal, and rhetorical relations that connect them. This structure is largely latent, only sparsely signaled by the surface text and otherwise recoverable from context. This HDR takes that observation as its unifying theme. Cast as a machine-learning problem, discourse understanding challenges three standard assumptions of classical supervised learning, each relaxed by extending context in a different sense: the structural context that makes a document’s parts mutually dependent, the semantic context of world knowledge beyond the text, and the learning context of how competence is acquired from data. The three core chapters take these in turn, contributing structured prediction and global inference; knowledge-grounded and multimodal representations; and new corpora and methods for learning from limited supervision. A final chapter reframes this trajectory around large language models, which address all three challenges through scale alone yet remain opaque and far less sample-efficient than human learners, and proposes an interdisciplinary program at the interface of NLP and cognitive science, comparing these models against human performance and studying what they learn under cognitively plausible constraints.
defended on 28/09/2026
français
English