Optimizing the Performance of Text Classification Models by Improving the Isotropy of the Embeddings using a Joint Loss Function

Abstract: Recent studies show that the spatial distribution of the sentence representations generated from pre-trained language models is highly anisotropic, meaning that the representations are not uniformly distributed among the directions of the embedding space. Thus, the expressiveness of the embedding space is limited, as the embeddings are less distinguishable and less diverse. This results in a degradation in the performance of the models on the downstream task. Most methods that define the state-of-the-art in this area proceed by improving the isotropy of the sentence embeddings by refining the corresponding contextual word representations, then deriving the sentence embeddings from these refined representations. In this thesis, we propose to improve the quality and distribution of the sentence embeddings extracted from the [CLS] token of the pre-trained language models by improving the isotropy of the embeddings. We add one feed-forward layer, referred to as the Isotropy Layer, between the model and the downstream task layers. We train this layer using a novel joint loss function that optimizes an isotropy quality measure and the downstream task loss. This joint loss pushes the embeddings outputted by the Isotropy Layer to be more isotropic, and it also retains the semantics needed to perform the downstream task. The proposed approach results in transformed embeddings with better isotropy, that generalize better on the downstream task. Furthermore, the approach requires training one feed-forward layer, instead of retraining the whole network. We quantify and evaluate the isotropy through multiple metrics, mainly the Explained Variance and the IsoScore. Experimental results on 3 GLUE datasets with classification as the downstream task show that our proposed method is on par with the state-of-the-art, as it achieves performance gains of around 2-3% on the downstream tasks compared to the baseline. We also present a small case study on one language abuse detection dataset, then interpret some of the findings in light of the results.