Kabir Ahuja

Kabir Ahuja, Harshita Diddee, Rishav Hada et al. · microsoft-research

Generative AI models have shown impressive performance on many Natural Language Processing tasks such as language understanding, reasoning, and language generation. An important question being asked by the AI community today is about the capabilities and limits of these models, and it is clear that evaluating generative AI is very challenging. Most studies on generative LLMs have been restricted to English and it is unclear how capable these models are at understanding and generating text in other languages. We present the first comprehensive benchmarking of generative LLMs - MEGA, which evaluates models on standard NLP benchmarks, covering 16 NLP datasets across 70 typologically diverse languages. We compare the performance of generative LLMs including Chat-GPT and GPT-4 to State of the Art (SOTA) non-autoregressive models on these tasks to determine how well generative models perform compared to the previous generation of LLMs. We present a thorough analysis of the performance of models across languages and tasks and discuss challenges in improving the performance of generative LLMs on low-resource languages. We create a framework for evaluating generative LLMs in the multilingual setting and provide directions for future progress in the field.

Kabir Ahuja, Shanu Kumar, Sandipan Dandapat et al.

Massively Multilingual Transformer based Language Models have been observed to be surprisingly effective on zero-shot transfer across languages, though the performance varies from language to language depending on the pivot language(s) used for fine-tuning. In this work, we build upon some of the existing techniques for predicting the zero-shot performance on a task, by modeling it as a multi-task learning problem. We jointly train predictive models for different tasks which helps us build more accurate predictors for tasks where we have test data in very few languages to measure the actual performance of the model. Our approach also lends us the ability to perform a much more robust feature selection and identify a common set of features that influence zero-shot performance across a variety of tasks.

Kabir Ahuja, Sunayana Sitaram, Sandipan Dandapat et al.

Massively Multilingual Language Models (MMLMs) have recently gained popularity due to their surprising effectiveness in cross-lingual transfer. While there has been much work in evaluating these models for their performance on a variety of tasks and languages, little attention has been paid on how well calibrated these models are with respect to the confidence in their predictions. We first investigate the calibration of MMLMs in the zero-shot setting and observe a clear case of miscalibration in low-resource languages or those which are typologically diverse from English. Next, we empirically show that calibration methods like temperature scaling and label smoothing do reasonably well towards improving calibration in the zero-shot scenario. We also find that few-shot examples in the language can further help reduce the calibration errors, often substantially. Overall, our work contributes towards building more reliable multilingual models by highlighting the issue of their miscalibration, understanding what language and model specific factors influence it, and pointing out the strategies to improve the same.

Kabir Ahuja, Sandipan Dandapat, Sunayana Sitaram et al.

Although recent Massively Multilingual Language Models (MMLMs) like mBERT and XLMR support around 100 languages, most existing multilingual NLP benchmarks provide evaluation data in only a handful of these languages with little linguistic diversity. We argue that this makes the existing practices in multilingual evaluation unreliable and does not provide a full picture of the performance of MMLMs across the linguistic landscape. We propose that the recent work done in Performance Prediction for NLP tasks can serve as a potential solution in fixing benchmarking in Multilingual NLP by utilizing features related to data and language typology to estimate the performance of an MMLM on different languages. We compare performance prediction with translating test data with a case study on four different multilingual datasets, and observe that these methods can provide reliable estimates of the performance that are often on-par with the translation based approaches, without the need for any additional translation as well as evaluation costs.

Madhur Panwar, Kabir Ahuja, Navin Goyal

In-context learning (ICL) is one of the surprising and useful features of large language models and subject of intense research. Recently, stylized meta-learning-like ICL setups have been devised that train transformers on sequences of input-output pairs $(x, f(x))$. The function $f$ comes from a function class and generalization is checked by evaluating on sequences generated from unseen functions from the same class. One of the main discoveries in this line of research has been that for several function classes, such as linear regression, transformers successfully generalize to new functions in the class. However, the inductive biases of these models resulting in this behavior are not clearly understood. A model with unlimited training data and compute is a Bayesian predictor: it learns the pretraining distribution. In this paper we empirically examine how far this Bayesian perspective can help us understand ICL. To this end, we generalize the previous meta-ICL setup to hierarchical meta-ICL setup which involve unions of multiple task families. We instantiate this setup on a diverse range of linear and nonlinear function families and find that transformers can do ICL in this setting as well. Where Bayesian inference is tractable, we find evidence that high-capacity transformers mimic the Bayesian predictor. The Bayesian perspective provides insights into the inductive bias of ICL and how transformers perform a particular task when they are trained on multiple tasks. We also find that transformers can learn to generalize to new function classes that were not seen during pretraining. This involves deviation from the Bayesian predictor. We examine these deviations in more depth offering new insights and hypotheses.

Aniket Vashishtha, Kabir Ahuja, Sunayana Sitaram

While understanding and removing gender biases in language models has been a long-standing problem in Natural Language Processing, prior research work has primarily been limited to English. In this work, we investigate some of the challenges with evaluating and mitigating biases in multilingual settings which stem from a lack of existing benchmarks and resources for bias evaluation beyond English especially for non-western context. In this paper, we first create a benchmark for evaluating gender biases in pre-trained masked language models by extending DisCo to different Indian languages using human annotations. We extend various debiasing methods to work beyond English and evaluate their effectiveness for SOTA massively multilingual models on our proposed metric. Overall, our work highlights the challenges that arise while studying social biases in multilingual settings and provides resources as well as mitigation techniques to take a step toward scaling to more languages.

Kabir Ahuja, Monojit Choudhury, Sandipan Dandapat

Borrowing ideas from {\em Production functions} in micro-economics, in this paper we introduce a framework to systematically evaluate the performance and cost trade-offs between machine-translated and manually-created labelled data for task-specific fine-tuning of massively multilingual language models. We illustrate the effectiveness of our framework through a case-study on the TyDIQA-GoldP dataset. One of the interesting conclusions of the study is that if the cost of machine translation is greater than zero, the optimal performance at least cost is always achieved with at least some or only manually-created data. To our knowledge, this is the first attempt towards extending the concept of production functions to study data collection strategies for training multilingual models, and can serve as a valuable tool for other similar cost vs data trade-offs in NLP.

Ishani Mondal, Kabir Ahuja, Mohit Jain et al.

The COVID-19 pandemic has brought out both the best and worst of language technology (LT). On one hand, conversational agents for information dissemination and basic diagnosis have seen widespread use, and arguably, had an important role in combating the pandemic. On the other hand, it has also become clear that such technologies are readily available for a handful of languages, and the vast majority of the global south is completely bereft of these benefits. What is the state of LT, especially conversational agents, for healthcare across the world's languages? And, what would it take to ensure global readiness of LT before the next pandemic? In this paper, we try to answer these questions through survey of existing literature and resources, as well as through a rapid chatbot building exercise for 15 Asian and African languages with varying amount of resource-availability. The study confirms the pitiful state of LT even for languages with large speaker bases, such as Sinhala and Hausa, and identifies the gaps that could help us prioritize research and investment strategies in LT for healthcare.

Fahim Faisal, Orevaoghene Ahia, Aarohi Srivastava et al.

Language technologies should be judged on their usefulness in real-world use cases. An often overlooked aspect in natural language processing (NLP) research and evaluation is language variation in the form of non-standard dialects or language varieties (hereafter, varieties). Most NLP benchmarks are limited to standard language varieties. To fill this gap, we propose DIALECTBENCH, the first-ever large-scale benchmark for NLP on varieties, which aggregates an extensive set of task-varied variety datasets (10 text-level tasks covering 281 varieties). This allows for a comprehensive evaluation of NLP system performance on different language varieties. We provide substantial evidence of performance disparities between standard and non-standard language varieties, and we also identify language clusters with large performance divergence across tasks. We believe DIALECTBENCH provides a comprehensive view of the current state of NLP for language varieties and one step towards advancing it further. Code/data: https://github.com/ffaisal93/DialectBench

Aniket Vashishtha, Abhinav Kumar, Atharva Pandey et al.

For text-based AI systems to interact in the real world, causal reasoning is an essential skill. Since active interventions are costly, we study to what extent a system can learn causal reasoning from symbolic demonstrations of causal axioms. Specifically, we present an axiomatic training method where the system learns from multiple demonstrations of a causal axiom (or rule), rather than incorporating the axiom as an inductive bias or inferring it from data values. A key question is whether the system would learn to generalize from the axiom demonstrations to more complex scenarios. Our results, based on applying axiomatic training to learn the transitivity axiom and d-separation rule, indicate that such generalization is possible. To avoid data contamination issues, we start with a 67 million parameter transformer model and train it from scratch. On both tasks, we find that a model trained on linear causal chains (along with some noisy variations) can generalize well to complex graphs, including longer causal chains, causal chains with reversed order, and graphs with branching.To handle diverse text inputs, the same method is extended to finetune language models. Finetuning Llama-3-8B-Instruct model on our axiomatic data leads to significant gains on causal benchmarks such as Corr2Cause and CLEAR, in some cases providing state-of-the-art performance surpassing GPT-4.

Kabir Ahuja, Yuxuan Li, Andrew Kyle Lampinen

Human communication is fundamentally creative, and often makes use of subtext -- implied meaning that goes beyond the literal content of the text. Here, we systematically study whether language models can use subtext in communicative settings, and introduce four new evaluation suites to assess these capabilities. Our evaluation settings range from writing & interpreting allegories to playing multi-agent and multi-modal games inspired by the rules of board games like Dixit. We find that frontier models generally exhibit a strong bias towards overly literal, explicit communication, and thereby fail to account for nuanced constraints -- even the best performing models generate literal clues 60% of times in one of our environments -- Visual Allusions. However, we find that some models can sometimes make use of common ground with another party to help them communicate with subtext, achieving 30%-50% reduction in overly literal clues; but they struggle at inferring presence of a common ground when not explicitly stated. For allegory understanding, we find paratextual and persona conditions to significantly shift the interpretation of subtext. Overall, our work provides quantifiable measures for an inherently complex and subjective phenomenon like subtext and reveals many weaknesses and idiosyncrasies of current LLMs. We hope this research to inspire future work towards socially grounded creative communication and reasoning.

Orevaoghene Ahia, Martijn Bartelds, Kabir Ahuja et al.

Developing large audio language models (LMs) capable of understanding diverse spoken interactions is essential for accommodating the multimodal nature of human communication and can increase the accessibility of language technologies across different user populations. Recent work on audio LMs has primarily evaluated their performance on short audio segments, typically under 30 seconds, with limited exploration of long-form conversational speech segments that more closely reflect natural user interactions with these models. We introduce Brutally Long Audio Bench (BLAB), a challenging long-form audio benchmark that evaluates audio LMs on localization, duration estimation, emotion, and counting tasks using audio segments averaging 51 minutes in length. BLAB consists of 833+ hours of diverse, full-length audio clips, each paired with human-annotated, text-based natural language questions and answers. Our audio data were collected from permissively licensed sources and underwent a human-assisted filtering process to ensure task compliance. We evaluate six open-source and proprietary audio LMs on BLAB and find that all of them, including advanced models such as Gemini 2.0 Pro and GPT-4o, struggle with the tasks in BLAB. Our comprehensive analysis reveals key insights into the trade-offs between task difficulty and audio duration. In general, we find that audio LMs struggle with long-form speech, with performance declining as duration increases. They perform poorly on localization, temporal reasoning, counting, and struggle to understand non-phonemic information, relying more on prompts than audio content. BLAB serves as a challenging evaluation framework to develop audio LMs with robust long-form audio understanding capabilities.

Ashutosh Kumar, Kabir Ahuja, Raghuram Vadapalli et al.

Given a sentence (e.g., "I like mangoes") and a constraint (e.g., sentiment flip), the goal of controlled text generation is to produce a sentence that adapts the input sentence to meet the requirements of the constraint (e.g., "I hate mangoes"). Going beyond such simple constraints, recent works have started exploring the incorporation of complex syntactic-guidance as constraints in the task of controlled paraphrase generation. In these methods, syntactic-guidance is sourced from a separate exemplar sentence. However, these prior works have only utilized limited syntactic information available in the parse tree of the exemplar sentence. We address this limitation in the paper and propose Syntax Guided Controlled Paraphraser (SGCP), an end-to-end framework for syntactic paraphrase generation. We find that SGCP can generate syntax conforming sentences while not compromising on relevance. We perform extensive automated and human evaluations over multiple real-world English language datasets to demonstrate the efficacy of SGCP over state-of-the-art baselines. To drive future research, we have made SGCP's source code available

Kabir Ahuja, Melanie Sclar, Yulia Tsvetkov · uw

Stories are a fundamental aspect of human experience. Engaging deeply with stories and spotting plot holes -- inconsistencies in a storyline that break the internal logic or rules of a story's world -- requires nuanced reasoning skills, including tracking entities and events and their interplay, abstract thinking, pragmatic narrative understanding, commonsense and social reasoning, and theory of mind. As Large Language Models (LLMs) increasingly generate, interpret, and modify text, rigorously assessing their narrative consistency and deeper language understanding becomes critical. However, existing benchmarks focus mainly on surface-level comprehension. In this work, we propose plot hole detection in stories as a proxy to evaluate language understanding and reasoning in LLMs. We introduce FlawedFictionsMaker, a novel algorithm to controllably and carefully synthesize plot holes in human-written stories. Using this algorithm, we construct a benchmark to evaluate LLMs' plot hole detection abilities in stories -- FlawedFictions -- , which is robust to contamination, with human filtering ensuring high quality. We find that state-of-the-art LLMs struggle in accurately solving FlawedFictions regardless of the reasoning effort allowed, with performance significantly degrading as story length increases. Finally, we show that LLM-based story summarization and story generation are prone to introducing plot holes, with more than 50% and 100% increases in plot hole detection rates with respect to human-written originals.

Kabir Ahuja, Vidhisha Balachandran, Madhur Panwar et al. · cmu, microsoft-research

Transformers trained on natural language data have been shown to learn its hierarchical structure and generalize to sentences with unseen syntactic structures without explicitly encoding any structural bias. In this work, we investigate sources of inductive bias in transformer models and their training that could cause such generalization behavior to emerge. We extensively experiment with transformer models trained on multiple synthetic datasets and with different training objectives and show that while other objectives e.g. sequence-to-sequence modeling, prefix language modeling, often failed to lead to hierarchical generalization, models trained with the language modeling objective consistently learned to generalize hierarchically. We then conduct pruning experiments to study how transformers trained with the language modeling objective encode hierarchical structure. When pruned, we find joint existence of subnetworks within the model with different generalization behaviors (subnetworks corresponding to hierarchical structure and linear order). Finally, we take a Bayesian perspective to further uncover transformers' preference for hierarchical generalization: We establish a correlation between whether transformers generalize hierarchically on a dataset and whether the simplest explanation of that dataset is provided by a hierarchical grammar compared to regular grammars exhibiting linear generalization.

Prachi Jain, Ashutosh Sathe, Varun Gumma et al. · microsoft-research

Pretrained Language Models (PLMs) are widely used in NLP for various tasks. Recent studies have identified various biases that such models exhibit and have proposed methods to correct these biases. However, most of the works address a limited set of bias dimensions independently such as gender, race, or religion. Moreover, the methods typically involve finetuning the full model to maintain the performance on the downstream task. In this work, we aim to modularly debias a pretrained language model across multiple dimensions. Previous works extensively explored debiasing PLMs using limited US-centric counterfactual data augmentation (CDA). We use structured knowledge and a large generative model to build a diverse CDA across multiple bias dimensions in a semi-automated way. We highlight how existing debiasing methods do not consider interactions between multiple societal biases and propose a debiasing model that exploits the synergy amongst various societal biases and enables multi-bias debiasing simultaneously. An extensive evaluation on multiple tasks and languages demonstrates the efficacy of our approach.

Somnath Kumar, Vaibhav Balloli, Mercy Ranjit et al.

Large language models (LLMs) have revolutionized various domains but still struggle with non-Latin scripts and low-resource languages. This paper addresses the critical challenge of improving multilingual performance without extensive fine-tuning. We introduce a novel dynamic learning approach that optimizes prompt strategy, embedding model, and LLM per query at runtime. By adapting configurations dynamically, our method achieves significant improvements over static, best and random baselines. It operates efficiently in both offline and online settings, generalizing seamlessly across new languages and datasets. Leveraging Retrieval-Augmented Generation (RAG) with state-of-the-art multilingual embeddings, we achieve superior task performance across diverse linguistic contexts. Through systematic investigation and evaluation across 18 diverse languages using popular question-answering (QA) datasets we show our approach results in 10-15% improvements in multilingual performance over pre-trained models and 4x gains compared to fine-tuned, language-specific models.

Satwik Bhattamishra, Kabir Ahuja, Navin Goyal

While recurrent models have been effective in NLP tasks, their performance on context-free languages (CFLs) has been found to be quite weak. Given that CFLs are believed to capture important phenomena such as hierarchical structure in natural languages, this discrepancy in performance calls for an explanation. We study the performance of recurrent models on Dyck-n languages, a particularly important and well-studied class of CFLs. We find that while recurrent models generalize nearly perfectly if the lengths of the training and test strings are from the same range, they perform poorly if the test strings are longer. At the same time, we observe that recurrent models are expressive enough to recognize Dyck words of arbitrary lengths in finite precision if their depths are bounded. Hence, we evaluate our models on samples generated from Dyck languages with bounded depth and find that they are indeed able to generalize to much higher lengths. Since natural language datasets have nested dependencies of bounded depth, this may help explain why they perform well in modeling hierarchical dependencies in natural language data despite prior works indicating poor generalization performance on Dyck languages. We perform probing studies to support our results and provide comparisons with Transformers.

Satwik Bhattamishra, Kabir Ahuja, Navin Goyal

Transformers have supplanted recurrent models in a large number of NLP tasks. However, the differences in their abilities to model different syntactic properties remain largely unknown. Past works suggest that LSTMs generalize very well on regular languages and have close connections with counter languages. In this work, we systematically study the ability of Transformers to model such languages as well as the role of its individual components in doing so. We first provide a construction of Transformers for a subclass of counter languages, including well-studied languages such as n-ary Boolean Expressions, Dyck-1, and its generalizations. In experiments, we find that Transformers do well on this subclass, and their learned mechanism strongly correlates with our construction. Perhaps surprisingly, in contrast to LSTMs, Transformers do well only on a subset of regular languages with degrading performance as we make languages more complex according to a well-known measure of complexity. Our analysis also provides insights on the role of self-attention mechanism in modeling certain behaviors and the influence of positional encoding schemes on the learning and generalization abilities of the model.