Assessing compositionality with phrase-level adjective-noun entailment

Compositionality is a fundamental feature of human language, allowing us to combine a finite set of primitives, such as morphemes and words, into virtually infinite combinations of meaning, such as sentences, phrases and discourses. In natural language processing (NLP), compositionality has been extensively studied through natural language inference (NLI) tasks, in which a model is asked to determine whether a hypothesis follows from a given premise. However, since NLI benchmarks generally use fully formed sentences, they are often crowded with biases and spurious correlations, which make compositionality and inference redundant to solving the proposed task. Phrase-level inference, and adjective-noun phrases especially, offer a versatile alternative to NLI. While phrases offer better control over internal distributions of a dataset, adjectives allow access to a large variety of compositional phenomena. However, this form of inference has been largely understudied, and work has been constrained with respect to the number of adjective classes, as well as the structure of the inferences. This thesis proposes to address the issues of NLI benchmarks by making full use of the potential of phrase-level adjective-noun entailment, and jointly study compositionality and inference in a fine-grained setting. To this end, it introduces PLANE, a resource to build large datasets to train and evaluate computational models. This resource will then be used to address the conflicting evidence around the compositional and inferential abilities of current large language models (LLMs), using in-context, transfer, and supervised learning experiments. Lastly, the work introduces multi-graph embeddings models (MuG), a new set of embedding models designed to address the limitations of current distributional models with respect to syntactic representation and the intrinsic evaluation of inference, and proposes to use PLANE to assess MuGs’ ability to perform phrase-level compositional entailment in a fully unsupervised manner.