Verb Class Induction with Partial Supervision

Abstract

Dirichlet-multinomial (D-M) mixtures like latent Dirichlet allocation (LDA) are widely used for both topic modeling and clustering. Prior work on constructing Levin-style semantic verb clusters achieves state-of-the-art results using D-M mixtures for verb sense induction and clustering. We add a bias toward known clusters by explicitly labeling a small number of observations with their correct VerbNet class. We demonstrate that this partial supervision guides the resulting clusters effectively, improving the recovery of both labeled and unlabeled classes by 16%, for a joint 12% absolute improvement in F1 score compared to clustering without supervision. The resulting clusters are also more semantically coherent. Although the technical change is minor, it produces a large effect, with important practical consequences for supervised topic modeling in general.

Verb Classification Across Languages

“…For instance, Peterson et al (2016) leveraged the existing annotated VerbNet data to guide the clustering process, where an expanded Dirichlet process mixture model additionally predicted a VerbNet class for each sense of each verb. Recently, Peterson et al (2020) achieved further improvements by partially supervising the joint sense induction and clustering model of Peterson & Palmer (2018) with a limited number of directly observed sentences annotated with VerbNet class labels. However, a qualitative analysis of the created clusters revealed a number of limitations in the behavior of the proposed model, which were not adequately reflected in the results of the evaluation on the SemLink corpus (Palmer 2009), suggesting that additional analyses and evaluation benchmarks are needed to reliably assess classification quality.…”