Automatically Identifying Semantic Bias in Crowdsourced Natural Language Inference Datasets
share
Natural language inference (NLI) is an important task for producing useful models of human language. Unfortunately large-scale NLI dataset production relies on crowdworkers who are prone to introduce biases in the sentences they write. In particular, without quality control they produce hypotheses from which the relational label can be predicted, without the premise, better than chance. We introduce a model-driven, unsupervised technique to find "bias clusters" in a learned embedding space of the hypotheses in NLI datasets, from which interventions and additional rounds of labeling can be performed to ameliorate the semantic bias of the hypothesis distribution of a dataset.
READ FULL TEXT