Measuring Context-Word Biases in Lexical Semantic Datasets

State-of-the-art contextualized models eg. BERT use tasks such as WiC and WSD to evaluate their word-in-context representations. This inherently assumes that performance in these tasks reflect how well a model represents the coupled word and context semantics. We question this assumption by presenting the first quantitative analysis on the context-word interaction required and being tested in major contextual lexical semantic tasks, taking into account that tasks can be inherently biased and models can learn spurious correlations from datasets. To achieve this, we run probing baselines on masked input, based on which we then propose measures to calculate the degree of context or word biases in a dataset, and plot existing datasets on a continuum. The analysis were performed on both models and humans to decouple biases inherent to the tasks and biases learned from the datasets. We found that, (1) to models, most existing datasets fall into the extreme ends of the continuum: the retrieval-based tasks and especially the ones in the medical domain (eg. COMETA) exhibit strong target word bias while WiC-style tasks and WSD show strong context bias; (2) AM2iCo and Sense Retrieval show less extreme model biases and challenge a model more to represent both the context and target words. (3) A similar trend of biases exists in humans but humans are much less biased compared with models as humans found semantic judgments more difficult with the masked input, indicating models are learning spurious correlations. This study demonstrates that with heavy context or target word biases, models are usually not being tested for word-in-context representations as such in these tasks and results are therefore open to misinterpretation. We recommend our framework as a sanity check for context and target word biases in future task design and model interpretation in lexical semantics.