We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as semantic or stylistic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use.To demonstrate our model’s interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by ~8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes get assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data — enabling generalization even when supervision blends semantic or stylistic factors.

We propose a new approach for the authorship attribution task that leverages the various linguistic representations learned at different layers of pre-trained transformer-based models. We evaluate our approach on two popular authorship attribution models and three evaluation datasets, in in-domain and out-of-domain scenarios. We found that utilizing various transformer layers improves the robustness of authorship attribution models when tested on out-of-domain data, resulting in a much stronger performance. Our analysis gives further insights into how our model’s different layers get specialized in representing certain linguistic aspects that we believe benefit the model when tested out of the domain.

Distributed word embeddings have become ubiquitous in natural language processing as they have been shown to improve performance in many semantic and syntactic tasks. Popular models for learning cross-lingual word embeddings do not consider the morphology of words. We propose an approach to learn bilingual embeddings using parallel data and subword information that is expressed in various forms, i.e. character n-grams, morphemes obtained by unsupervised morphological segmentation and byte pair encoding. We report results for three low resource morphologically rich languages (Swahili, Tagalog, and Somali) and a high resource language (German) in a simulated a low-resource scenario. Our results show that our method that leverages subword information outperforms the model without subword information, both in intrinsic and extrinsic evaluations of the learned embeddings. Specifically, analogy reasoning results show that using subwords helps capture syntactic characteristics. Semantically, word similarity results and intrinsically, word translation scores demonstrate superior performance over existing methods. Finally, qualitative analysis also shows better-quality cross-lingual embeddings particularly for morphological variants in both languages.