Contributions à l'évaluation et à l'amélioration de la fidélité pour la génération de textes

Ensuring the factual faithfulness in text generation is a critical challenge, especially in knowledge-grounded tasks where the generated text must accurately reflect the content in structured or unstructured input data. While modern models have achieved good fluency and coherence, they frequently generate hallucinated outputs, such as ignoring details, introducing contradictions, or misinterpreting relations. To address these issues, this thesis presents new methods for evaluating and improving the faith- fulness of text generation. The first contribution of the thesis is FactSpotter, a novel metric to assess the faithfulness in graph-to-text generation. Unlike approaches based on n-grams or embeddings, FactSpotter evaluates whether each triple in the knowledge graph is correctly verbalised in the text. It leverages a self-supervised model to distinguish between faithful and unfaithful texts. Beyond triple-level evaluation for graph-to-text generation, FactSpotter can also be integrated into beam search as a constraint, encouraging faithful text generation without compromising fluency. Second, the thesis explores cross-text factual consistency verification, assessing whether multiple texts convey the same content. A structured discourse level format is introduced to address the limitation of triple representation. The proposed format captures not only richer details, such as direct/indirect objects, adverbials, and complements, but also discourse- level relations, including temporality, comparison, and contingency. A new dataset, DiscInfer, is annotated to train entailment-based models to detect both atomic fact- and discourse-level inconsistencies. Empirical results show that the proposed methods in the thesis enhance both faithful text generation and factual consistency evaluation. FactSpotter-guided de-coding mitigates hallucinations in text generation, while structured discourse representation enhances factual consistency verification. These contributions help to build more trustworthy text generation systems offering insights into improving fact-grounded generation across various applications.