We propose a novel scaling law for general-purpose decoder-only language models (LMs) trained on multilingual data, tackling the problem of balancing languages during multilingual pretraining. A primary challenge in studying multilingual scaling is the difficulty of analyzing individual language performance due to cross-lingual transfer. To tackle this, we shift the focus from individual languages to language families. We introduce and validate a hypothesis that the test cross-entropy loss for each language family is determined solely by its own sampling ratio, independent of other languages in the mixture. This insight simplifies the complexity of multilingual scaling and make the analysis scalable to an arbitrary number of languages. Building on this hypothesis, we derive a power-law relationship that links performance with dataset size, model size and sampling ratios. This relationship enables us to predict performance across various combinations of the above three quantities, and derive the optimal sampling ratios at different model scales. To demonstrate the effectiveness and accuracy of our proposed scaling law, we perform a large-scale empirical study, training more than 100 models on 23 languages spanning 5 language families. Our experiments show that the optimal sampling ratios derived from small models (85M parameters) generalize effectively to models that are several orders of magnitude larger (1.2B parameters), offering a resource-efficient approach for multilingual LM training at scale.

Code-switching (CS), a ubiquitous phenomenon due to the ease of communication it offers in multilingual communities still remains an understudied problem in language processing. The primary reasons behind this are: (1) minimal efforts in leveraging large pretrained multilingual models, and (2) the lack of annotated data. The distinguishing case of low performance of multilingual models in CS is the intra-sentence mixing of languages leading to switch points. We first benchmark two sequence labeling tasks – POS and NER on 4 different language pairs with a suite of pretrained models to identify the problems and select the best performing char-BERT model among them (addressing (1)). We then propose a self training method to repurpose the existing pretrained models using a switch-point bias by leveraging unannotated data (addressing (2)). We finally demonstrate that our approach performs well on both tasks by reducing the gap between the switch point performance while retaining the overall performance on two distinct language pairs in both the tasks. We plan to release our models and the code for all our experiments.

Code-Mixing (CM) or language mixing is a social norm in multilingual societies. CM is quite prevalent in social media conversations in multilingual regions like - India, Europe, Canada and Mexico. In this paper, we explore the problem of Language Modeling (LM) for code-mixed Hinglish text. In recent times, there have been several success stories with neural language modeling like Generative Pre-trained Transformer (GPT) (Radford et al., 2019), Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., 2018) etc.. Hence, neural language models have become the new holy grail of modern NLP, although LM for CM is an unexplored area altogether. To better understand the problem of LM for CM, we initially experimented with several statistical language modeling techniques and consequently experimented with contemporary neural language models. Analysis shows switching-points are the main challenge for the LMCM performance drop, therefore in this paper we introduce the idea of minority positive sampling to selectively induce more sample to achieve better performance. On the contrary, all neural language models demand a huge corpus to train on for better performance. Finally, we are reporting a perplexity of 139 for Hinglish (Hindi-English language pair) LMCM using statistical bi-directional techniques.