Large Language Models (LLMs) exhibit strong multilingual performance despite being predominantly trained on English-centric corpora. This raises a fundamental question: How do LLMs achieve such multilingual capabilities? Focusing on languages written in non-Roman scripts, we investigate the role of Romanization—the representation of non-Roman scripts using Roman characters—as a potential bridge in multilingual processing. Using mechanistic interpretability techniques, we analyze next-token generation and find that intermediate layers frequently represent target words in Romanized form before transitioning to native script, a phenomenon we term Latent Romanization. Further, through activation patching experiments, we demonstrate that LLMs encode semantic concepts similarly across native and Romanized scripts, suggesting a shared underlying representation. Additionally, for translation into non-Roman script languages, our findings reveal that when the target language is in Romanized form, its representations emerge earlier in the model’s layers compared to native script. These insights contribute to a deeper understanding of multilingual representation in LLMs and highlight the implicit role of Romanization in facilitating language transfer.

This study addresses the challenge of extending Large Language Models (LLMs) to non-English languages, specifically those using non-Roman scripts. We propose an approach that utilizes the romanized form of text as an interface for LLMs, hypothesizing that its frequent informal use and shared tokens with English enhance cross-lingual alignment. Our approach involve the continual pretraining of a English LLM like Llama 2 on romanized text of non-English, non-Roman script languages, followed by instruction tuning on romanized data. The results indicate that romanized text not only reduces token fertility by 2x-4x but also matches if not outperforms native script representation across various NLU, NLG and MT tasks. Moreover, the embeddings computed on romanized text exhibit closer alignment with their English translations than those from the native script. Our approach presents a promising direction for leveraging the power of English LLMs in languages traditionally underrepresented in NLP research.

Open Information Extraction (OIE) is a structure prediction (SP) task in Natural Language Processing (NLP) that aims to extract structured n-ary tuples - usually subject-relation-object triples - from free text. The word embeddings in the input text can be enhanced with linguistic features, usually Part-of-Speech (PoS) and Syntactic Dependency Parse (SynDP) labels. However, past enhancement techniques cannot leverage the power of pre-trained language models (PLMs), which themselves have been hardly used for OIE. To bridge this gap, we are the first to leverage linguistic features with a Seq2Seq PLM for OIE. We do so by introducing two methods - Weighted Addition and Linearized Concatenation. Our work gives any neural OIE architecture the key performance boost from both PLMs and linguistic features in one go. In our settings, this shows wide improvements of up to 24.9%, 27.3% and 14.9% on Precision, Recall and F1 scores respectively over the baseline. Beyond this, we address other important challenges in the field: to reduce compute overheads with the features, we are the first ones to exploit Semantic Dependency Parse (SemDP) tags; to address flaws in current datasets, we create a clean synthetic dataset; finally, we contribute the first known study of OIE behaviour in SP models.

Realizing the recent advances from Natural Language Processing (NLP) to the legal sector poses challenging problems such as extremely long sequence lengths, specialized vocabulary that is usually only understood by legal professionals, and high amounts of data imbalance. The recent surge of Large Language Models (LLM) has begun to provide new opportunities to apply NLP in the legal domain due to their ability to handle lengthy, complex sequences. Moreover, the emergence of domain-specific LLMs has displayed extremely promising results on various tasks. In this study, we aim to quantify how general LLMs perform in comparison to legal-domain models (be it an LLM or otherwise). Specifically, we compare the zero-shot performance of three general-purpose LLMs (ChatGPT-3.5, LLaMA-70b and Falcon-180b) on the LEDGAR subset of the LexGLUE benchmark for contract provision classification. Although the LLMs were not explicitly trained on legal data, we observe that they are still able to classify the theme correctly in most cases. However, we find that their mic-F1/mac-F1 performance are upto 19.2/26.8% lesser than smaller models fine-tuned on the legal domain, thus underscoring the need for more powerful legal-domain LLMs.