Transformer-based language models have taken many fields in NLP by storm. BERT and its derivatives dominate most of the existing evaluation benchmarks, including those for Word Sense Disambiguation (WSD), thanks to their ability in capturing context-sensitive semantic nuances. However, there is still little knowledge about their capabilities and potential limitations in encoding and recovering word senses. In this article, we provide an in-depth quantitative and qualitative analysis of the celebrated BERT model with respect to lexical ambiguity. One of the main conclusions of our analysis is that BERT can accurately capture high-level sense distinctions, even when a limited number of examples is available for each word sense. Our analysis also reveals that in some cases language models come close to solving coarse-grained noun disambiguation under ideal conditions in terms of availability of training data and computing resources. However, this scenario rarely occurs in real-world settings and, hence, many practical challenges remain even in the coarse-grained setting. We also perform an in-depth comparison of the two main language model-based WSD strategies, namely, fine-tuning and feature extraction, finding that the latter approach is more robust with respect to sense bias and it can better exploit limited available training data. In fact, the simple feature extraction strategy of averaging contextualized embeddings proves robust even using only three training sentences per word sense, with minimal improvements obtained by increasing the size of this training data.

The evaluation of several tasks in lexical semantics is often limited by the lack of large amounts of manual annotations, not only for training purposes, but also for testing purposes. Word Sense Disambiguation (WSD) is a case in point, as hand-labeled datasets are particularly hard and time-consuming to create. Consequently, evaluations tend to be performed on a small scale, which does not allow for in-depth analysis of the factors that determine a systems' performance. In this paper we address this issue by means of a realistic simulation of large-scale evaluation for the WSD task. We do this by providing two main contributions: First, we put forward two novel approaches to the wide-coverage generation of semantically aware pseudowords (i.e., artificial words capable of modeling real polysemous words); second, we leverage the most suitable type of pseudoword to create large pseudosense-annotated corpora, which enable a large-scale experimental framework for the comparison of state-of-the-art supervised and knowledge-based algorithms. Using this framework, we study the impact of supervision and knowledge on the two major disambiguation paradigms and perform an in-depth analysis of the factors which affect their performance.