In this paper, we present the masala-chai team’s participation in the LegalLens 2024 shared task and detail our approach to predicting legal entities and performing natural language inference (NLI) in the legal domain. We experimented with various transformer-based models, including BERT, RoBERTa, Llama 3.1, and GPT-4o. Our results show that state-of-the-art models like GPT-4o underperformed in NER and NLI tasks, even when using advanced techniques such as bootstrapping and prompt optimization. The best performance in NER (accuracy: 0.806, F1 macro: 0.701) was achieved with a fine-tuned RoBERTa model, while the highest NLI results (accuracy: 0.825, F1 macro: 0.833) came from a fine-tuned Llama 3.1 8B model. Notably, RoBERTa, despite having significantly fewer parameters than Llama 3.1 8B, delivered comparable results. We discuss key findings and insights from our experiments and provide our results and code for reproducibility and further analysis at https://github.com/rosequ/masala-chai

Word-level language detection is necessary for analyzing code-switched text, where multiple languages could be mixed within a sentence. Existing models are restricted to code-switching between two specific languages and fail in real-world scenarios as text input rarely has a priori information on the languages used. We present a novel unsupervised word-level language detection technique for code-switched text for an arbitrarily large number of languages, which does not require any manually annotated training data. Our experiments with tweets in seven languages show a 74% relative error reduction in word-level labeling with respect to competitive baselines. We then use this system to conduct a large-scale quantitative analysis of code-switching patterns on Twitter, both global as well as region-specific, with 58M tweets.