On the Inductive Bias of Masked Language Modeling: From Statistical to Syntactic Dependencies

We study how masking and predicting tokens in an unsupervised fashion can give rise to linguistic structures and downstream performance gains. Recent theories have suggested that pretrained language models acquire useful inductive biases through masks that implicitly act as cloze reductions for downstream tasks. While appealing, we show that the success of the random masking strategy used in practice cannot be explained by such cloze-like masks alone. We construct cloze-like masks using task-specific lexicons for three different classification datasets and show that the majority of pretrained performance gains come from generic masks that are not associated with the lexicon. To explain the empirical success of these generic masks, we demonstrate a correspondence between the Masked Language Model (MLM) objective and existing methods for learning statistical dependencies in graphical models. Using this, we derive a method for extracting these learned statistical dependencies in MLMs and show that these dependencies encode useful inductive biases in the form of syntactic structures. In an unsupervised parsing evaluation, simply forming a minimum spanning tree on the implied statistical dependence structure outperforms a classic method for unsupervised parsing (58.74 vs. 55.91 UUAS).

Too Much in Common: Shifting of Embeddings in Transformer Language Models and its Implications

The success of language models based on the Transformer architecture appears to be inconsistent with observed anisotropic properties of representations learned by such models. We resolve this by showing, contrary to previous studies, that the representations do not occupy a narrow cone, but rather drift in common directions. At any training step, all of the embeddings except for the ground-truth target embedding are updated with gradient in the same direction. Compounded over the training set, the embeddings drift and share common components, manifested in their shape in all the models we have empirically tested. Our experiments show that isotropy can be restored using a simple transformation.

Benchmarking Transformer-based Language Models for Arabic Sentiment and Sarcasm Detection

The introduction of transformer-based language models has been a revolutionary step for natural language processing (NLP) research. These models, such as BERT, GPT and ELECTRA, led to state-of-the-art performance in many NLP tasks. Most of these models were initially developed for English and other languages followed later. Recently, several Arabic-specific models started emerging. However, there are limited direct comparisons between these models. In this paper, we evaluate the performance of 24 of these models on Arabic sentiment and sarcasm detection. Our results show that the models achieving the best performance are those that are trained on only Arabic data, including dialectal Arabic, and use a larger number of parameters, such as the recently released MARBERT. However, we noticed that AraELECTRA is one of the top performing models while being much more efficient in its computational cost. Finally, the experiments on AraGPT2 variants showed low performance compared to BERT models, which indicates that it might not be suitable for classification tasks.