New frontiers in supervised word sense disambiguation: building multilingual resources and neural models on a large scale

Word Sense Disambiguation is a long-standing task in Natural Language Processing (NLP), lying at the core of human language understanding. While it has already been studied from many different angles over the years, ranging from knowledge based systems to semi-supervised and fully supervised models, the field seems to be slowing down in respect to other NLP tasks, e.g., part-of-speech tagging and dependencies parsing. Despite the organization of several international competitions aimed at evaluating Word Sense Disambiguation systems, the evaluation of automatic systems has been problematic mainly due to the lack of a reliable evaluation framework aiming at performing a direct quantitative confrontation. To this end we develop a unified evaluation framework and analyze the performance of various Word Sense Disambiguation systems in a fair setup. The results show that supervised systems clearly outperform knowledge-based models. Among the supervised systems, a linear classifier trained on conventional local features still proves to be a hard baseline to beat. Nonetheless, recent approaches exploiting neural networks on unlabeled corpora achieve promising results, surpassing this hard baseline in most test sets. Even though supervised systems tend to perform best in terms of accuracy, they often lose ground to more flexible knowledge-based solutions, which do not require training for every disambiguation target. To bridge this gap we adopt a different perspective and rely on sequence learning to frame the disambiguation problem: we propose and study in depth a series of end-to-end neural architectures directly tailored to the task, from bidirectional Long ShortTerm Memory to encoder-decoder models. Our extensive evaluation over standard benchmarks and in multiple languages shows that sequence learning enables more versatile all-words models that consistently lead to state-of-the-art results, even against models trained with engineered features. However, supervised systems need annotated training corpora and the few available to date are of limited size: this is mainly due to the expensive and timeconsuming process of annotating a wide variety of word senses at a reasonably high scale, i.e., the so-called knowledge acquisition bottleneck. To address this issue, we also present different strategies to acquire automatically high quality sense annotated data in multiple languages, without any manual effort. We assess the quality of the sense annotations both intrinsically and extrinsically achieving competitive results on multiple tasks.