Recent Large Language Models (LLMs) have unlocked unprecedented applications of AI. As these models continue to transform human life, there are growing socio-ethical concerns around their inherent stereotypes that can lead to bias in their applications. There is an urgent need for holistic bias evaluation of these LLMs. Few such benchmarks exist today and evaluation techniques that do exist are either non-holistic or may provide a false sense of security as LLMs become better at hiding their biases on simpler tasks. We address these issues with an extensible benchmark - LLM Stereotype Index (LSI). LSI is grounded on Social Progress Index, a holistic social benchmark. We also test the breadth and depth of bias protection provided by LLMs via a variety of tasks with varying complexities. Our findings show that both ChatGPT and GPT-4 have strong inherent prejudice with respect to nationality, gender, race, and religion. The exhibition of such issues becomes increasingly apparent as we increase task complexity. Furthermore, GPT-4 is better at hiding the biases, but when displayed it is more significant. Our findings highlight the harms and divide that these LLMs can bring to society if we do not take very diligent care in their use.

Large Language Models such as GPT-3 are well-suited for text prediction tasks, which can help and delight users during text composition. LLMs are known to generate ethically inappropriate predictions even for seemingly innocuous contexts. Toxicity detection followed by filtering is a common strategy for mitigating the harm from such predictions. However, as we shall argue in this paper, in the context of text prediction, it is not sufficient to detect and filter toxic content. One also needs to ensure factual correctness and group-level fairness of the predictions; failing to do so can make the system ineffective and nonsensical at best, and unfair and detrimental to the users at worst. We discuss the gaps and challenges of toxicity detection approaches - from blocklist-based approaches to sophisticated state-of-the-art neural classifiers - by evaluating them on the text prediction task for English against a manually crafted CheckList of harms targeted at different groups and different levels of severity.

Temporal reasoning represents a vital component of human communication and understanding, yet remains an underexplored area within the context of Large Language Models (LLMs). Despite LLMs demonstrating significant proficiency in a range of tasks, a comprehensive, large-scale analysis of their temporal reasoning capabilities is missing. Our paper addresses this gap, presenting the first extensive benchmarking of LLMs on temporal reasoning tasks. We critically evaluate 8 different LLMs across 6 datasets using 3 distinct prompting strategies. Additionally, we broaden the scope of our evaluation by including in our analysis 2 Code Generation LMs. Beyond broad benchmarking of models and prompts, we also conduct a fine-grained investigation of performance across different categories of temporal tasks. We further analyze the LLMs on varying temporal aspects, offering insights into their proficiency in understanding and predicting the continuity, sequence, and progression of events over time. Our findings reveal a nuanced depiction of the capabilities and limitations of the models within temporal reasoning, offering a comprehensive reference for future research in this pivotal domain.

Zero-shot cross-lingual transfer is promising, however has been shown to be sub-optimal, with inferior transfer performance across low-resource languages. In this work, we envision languages as domains for improving zero-shot transfer by jointly reducing the feature incongruity between the source and the target language and increasing the generalization capabilities of pre-trained multilingual transformers. We show that our approach, DiTTO, significantly outperforms the standard zero-shot fine-tuning method on multiple datasets across all languages using solely unlabeled instances in the target language. Empirical results show that jointly reducing feature incongruity for multiple target languages is vital for successful cross-lingual transfer. Moreover, our model enables better cross-lingual transfer than standard fine-tuning methods, even in the few-shot setting.

In this paper, we introduce a fine-tuned transformer-based model focused on problematic webpage classification to identify webpages promoting hate and violence of various forms. Due to the unavailability of labelled problematic webpage data, first we propose a novel webpage data collection strategy which leverages well-studied short-text hate speech datasets. We have introduced a custom GPT-4 few-shot prompt annotation scheme taking various webpage features to label the prohibitively expensive webpage annotation task. The resulting annotated data is used to build our problematic webpage classification model. We report the accuracy (87.6% F1-score) of our webpage classification model and conduct a detailed comparison of it against other state-of-the-art hate speech classification model on problematic webpage identification task. Finally, we have showcased the importance of various webpage features in identifying a problematic webpage.

Few-shot transfer often shows substantial gain over zero-shot transfer (CITATION), which is a practically useful trade-off between fully supervised and unsupervised learning approaches for multilingual pretained model-based systems. This paper explores various strategies for selecting data for annotation that can result in a better few-shot transfer. The proposed approaches rely on multiple measures such as data entropy using n-gram language model, predictive entropy, and gradient embedding. We propose a loss embedding method for sequence labeling tasks, which induces diversity and uncertainty sampling similar to gradient embedding. The proposed data selection strategies are evaluated and compared for POS tagging, NER, and NLI tasks for up to 20 languages. Our experiments show that the gradient and loss embedding-based strategies consistently outperform random data selection baselines, with gains varying with the initial performance of the zero-shot transfer. Furthermore, the proposed method shows similar trends in improvement even when the model is fine-tuned using a lower proportion of the original task-specific labeled training data for zero-shot transfer.

Multilingual evaluation benchmarks usually contain limited high-resource languages and do not test models for specific linguistic capabilities. CheckList is a template-based evaluation approach that tests models for specific capabilities. The CheckList template creation process requires native speakers, posing a challenge in scaling to hundreds of languages. In this work, we explore multiple approaches to generate Multilingual CheckLists. We device an algorithm –Template Extraction Algorithm (TEA) for automatically extracting target language CheckList templates from machine translated instances of a source language templates. We compare the TEA CheckLists with CheckLists created with different levels of human intervention. We further introduce metrics along the dimensions of cost, diversity, utility, and correctness to compare the CheckLists. We thoroughly analyze different approaches to creating CheckLists in Hindi. Furthermore, we experiment with 9 more different languages. We find that TEA followed by human verification is ideal for scaling Checklist-based evaluation to multiple languages while TEA gives a good estimates of model performance. We release the code of TEA and the CheckLists created at aka.ms/multilingualchecklist

Topic-sensitive query set expansion is an important area of research that aims to improve search results for information retrieval. It is particularly crucial for queries related to sensitive and emerging topics. In this work, we describe a method for query set expansion about emerging topics using vector space interpolation. We use a transformer model called OPTIMUS, which is suitable for vector space manipulation due to its variational autoencoder nature. One of our proposed methods – Dirichlet interpolation shows promising results for query expansion. Our methods effectively generate new queries about the sensitive topic by incorporating set-level diversity, which is not captured by traditional sentence-level augmentation methods such as paraphrasing or back-translation.

Leveraging shared learning through Massively Multilingual Models, state-of-the-art Machine translation (MT) models are often able to adapt to the paucity of data for low-resource languages. However, this performance comes at the cost of significantly bloated models which aren’t practically deployable. Knowledge Distillation is one popular technique to develop competitive lightweight models: In this work, we first evaluate its use in compressing MT models, focusing specifically on languages with extremely limited training data. Through our analysis across 8 languages, we find that the variance in the performance of the distilled models due to their dependence on priors including the amount of synthetic data used for distillation, the student architecture, training hyper-parameters and confidence of the teacher models, makes distillation a brittle compression mechanism. To mitigate this, we further explore the use of post-training quantization for the compression of these models. Here, we find that while Distillation provides gains across some low-resource languages, Quantization provides more consistent performance trends for the entire range of languages, especially the lowest-resource languages in our target set.

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.

Borrowing ideas from 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 conclusion 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.

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.

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 in improving calibration in the zero-shot scenario. We also find that few-shot examples in the language can further help reduce 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.

Despite state-of-the-art performance, NLP systems can be fragile in real-world situations. This is often due to insufficient understanding of the capabilities and limitations of models and the heavy reliance on standard evaluation benchmarks. Research into non-standard evaluation to mitigate this brittleness is gaining increasing attention. Notably, the behavioral testing principle ‘Checklist’, which decouples testing from implementation revealed significant failures in state-of-the-art models for multiple tasks. In this paper, we present a case study of using Checklist in a practical scenario. We conduct experiments for evaluating an offensive content detection system and use a data augmentation technique for improving the model using insights from Checklist. We lay out the challenges and open questions based on our observations of using Checklist for human-in-loop evaluation and improvement of NLP systems. Disclaimer: The paper contains examples of content with offensive language. The examples do not represent the views of the authors or their employers towards any person(s), group(s), practice(s), or entity/entities.

Deep Contextual Language Models (LMs) like ELMO, BERT, and their successors dominate the landscape of Natural Language Processing due to their ability to scale across multiple tasks rapidly by pre-training a single model, followed by task-specific fine-tuning. Furthermore, multilingual versions of such models like XLM-R and mBERT have given promising results in zero-shot cross-lingual transfer, potentially enabling NLP applications in many under-served and under-resourced languages. Due to this initial success, pre-trained models are being used as ‘Universal Language Models’ as the starting point across diverse tasks, domains, and languages. This work explores the notion of ‘Universality’ by identifying seven dimensions across which a universal model should be able to scale, that is, perform equally well or reasonably well, to be useful across diverse settings. We outline the current theoretical and empirical results that support model performance across these dimensions, along with extensions that may help address some of their current limitations. Through this survey, we lay the foundation for understanding the capabilities and limitations of massive contextual language models and help discern research gaps and directions for future work to make these LMs inclusive and fair to diverse applications, users, and linguistic phenomena.

Natural Language Inference (NLI) is the task of inferring the logical relationship, typically entailment or contradiction, between a premise and hypothesis. Code-mixing is the use of more than one language in the same conversation or utterance, and is prevalent in multilingual communities all over the world. In this paper, we present the first dataset for code-mixed NLI, in which both the premises and hypotheses are in code-mixed Hindi-English. We use data from Hindi movies (Bollywood) as premises, and crowd-source hypotheses from Hindi-English bilinguals. We conduct a pilot annotation study and describe the final annotation protocol based on observations from the pilot. Currently, the data collected consists of 400 premises in the form of code-mixed conversation snippets and 2240 code-mixed hypotheses. We conduct an extensive analysis to infer the linguistic phenomena commonly observed in the dataset obtained. We evaluate the dataset using a standard mBERT-based pipeline for NLI and report results.

In this paper, we explore the methods of obtaining parse trees of code-mixed sentences and analyse the obtained trees. Existing work has shown that linguistic theories can be used to generate code-mixed sentences from a set of parallel sentences. We build upon this work, using one of these theories, the Equivalence-Constraint theory to obtain the parse trees of synthetically generated code-mixed sentences and evaluate them with a neural constituency parser. We highlight the lack of a dataset non-synthetic code-mixed constituency parse trees and how it makes our evaluation difficult. To complete our evaluation, we convert a code-mixed dependency parse tree set into “pseudo constituency trees” and find that a parser trained on synthetically generated trees is able to decently parse these as well.

Code-switching is the use of more than one language in the same conversation or utterance. Recently, multilingual contextual embedding models, trained on multiple monolingual corpora, have shown promising results on cross-lingual and multilingual tasks. We present an evaluation benchmark, GLUECoS, for code-switched languages, that spans several NLP tasks in English-Hindi and English-Spanish. Specifically, our evaluation benchmark includes Language Identification from text, POS tagging, Named Entity Recognition, Sentiment Analysis, Question Answering and a new task for code-switching, Natural Language Inference. We present results on all these tasks using cross-lingual word embedding models and multilingual models. In addition, we fine-tune multilingual models on artificially generated code-switched data. Although multilingual models perform significantly better than cross-lingual models, our results show that in most tasks, across both language pairs, multilingual models fine-tuned on code-switched data perform best, showing that multilingual models can be further optimized for code-switching tasks.

Multilingual communities exhibit code-mixing, that is, mixing of two or more socially stable languages in a single conversation, sometimes even in a single utterance. This phenomenon has been widely studied by linguists and interaction scientists in the spoken language of such communities. However, with the prevalence of social media and other informal interactive platforms, code-switching is now also ubiquitously observed in user-generated text. As multilingual communities are more the norm from a global perspective, it becomes essential that code-switched text and speech are adequately handled by language technologies and NUIs.Code-mixing is extremely prevalent in all multilingual societies. Current studies have shown that as much as 20% of user generated content from some geographies, like South Asia, parts of Europe, and Singapore, are code-mixed. Thus, it is very important to handle code-mixed content as a part of NLP systems and applications for these geographies.In the past 5 years, there has been an active interest in computational models for code-mixing with a substantive research outcome in terms of publications, datasets and systems. However, it is not easy to find a single point of access for a complete and coherent overview of the research. This tutorial is expecting to fill this gap and provide new researchers in the area with a foundation in both linguistic and computational aspects of code-mixing. We hope that this then becomes a starting point for those who wish to pursue research, design, development and deployment of code-mixed systems in multilingual societies.

In this paper, we demonstrate an Interactive Machine Translation interface, that assists human translators with on-the-fly hints and suggestions. This makes the end-to-end translation process faster, more efficient and creates high-quality translations. We augment the OpenNMT backend with a mechanism to accept the user input and generate conditioned translations.

A large percentage of the world’s population speaks a language of the Indian subcontinent, what we will call here Indic languages, comprising languages from both Indo-European (e.g., Hindi, Bangla, Gujarati, etc.) and Dravidian (e.g., Tamil, Telugu, Malayalam, etc.) families, upwards of 1.5 Billion people. A universal characteristic of Indic languages is their complex morphology, which, when combined with the general lack of sufficient quantities of high quality parallel data, can make developing machine translation (MT) for these languages difficult. In this paper, we describe our efforts towards developing general domain English–Bangla MT systems which are deployable to the Web. We initially developed and deployed SMT-based systems, but over time migrated to NMT-based systems. Our initial SMT-based systems had reasonably good BLEU scores, however, using NMT systems, we have gained significant improvement over SMT baselines. This is achieved using a number of ideas to boost the data store and counter data sparsity: crowd translation of intelligently selected monolingual data (throughput enhanced by an IME (Input Method Editor) designed specifically for QWERTY keyboard entry for Devanagari scripted languages), back-translation, different regularization techniques, dataset augmentation and early stopping.

Telugu is the fifteenth most commonly spoken language in the world with an estimated reach of 75 million people in the Indian subcontinent. At the same time, it is a severely low resourced language. In this paper, we present work on English–Telugu general domain machine translation (MT) systems using small amounts of parallel data. The baseline statistical (SMT) and neural MT (NMT) systems do not yield acceptable translation quality, mostly due to limited resources. However, the use of synthetic parallel data (generated using back translation, based on an NMT engine) significantly improves translation quality and allows NMT to outperform SMT. We extend back translation and propose a new, iterative data augmentation (IDA) method. Filtering of synthetic data and IDA both further boost translation quality of our final NMT systems, as measured by BLEU scores on all test sets and based on state-of-the-art human evaluation.

Getting manually labeled data in each domain is always an expensive and a time consuming task. Cross-domain sentiment analysis has emerged as a demanding concept where a labeled source domain facilitates a sentiment classifier for an unlabeled target domain. However, polarity orientation (positive or negative) and the significance of a word to express an opinion often differ from one domain to another domain. Owing to these differences, cross-domain sentiment classification is still a challenging task. In this paper, we propose that words that do not change their polarity and significance represent the transferable (usable) information across domains for cross-domain sentiment classification. We present a novel approach based on χ2 test and cosine-similarity between context vector of words to identify polarity preserving significant words across domains. Furthermore, we show that a weighted ensemble of the classifiers enhances the cross-domain classification performance.

Training language models for Code-mixed (CM) language is known to be a difficult problem because of lack of data compounded by the increased confusability due to the presence of more than one language. We present a computational technique for creation of grammatically valid artificial CM data based on the Equivalence Constraint Theory. We show that when training examples are sampled appropriately from this synthetic data and presented in certain order (aka training curriculum) along with monolingual and real CM data, it can significantly reduce the perplexity of an RNN-based language model. We also show that randomly generated CM data does not help in decreasing the perplexity of the LMs.

We offer a fluctuation smoothing computational approach for unsupervised automatic short answer grading (ASAG) techniques in the educational ecosystem. A major drawback of the existing techniques is the significant effect that variations in model answers could have on their performances. The proposed fluctuation smoothing approach, based on classical sequential pattern mining, exploits lexical overlap in students’ answers to any typical question. We empirically demonstrate using multiple datasets that the proposed approach improves the overall performance and significantly reduces (up to 63%) variation in performance (standard deviation) of unsupervised ASAG techniques. We bring in additional benchmarks such as (a) paraphrasing of model answers and (b) using answers by k top performing students as model answers, to amplify the benefits of the proposed approach.

State-of-the-art statistical machine translation (SMT) technique requires a good quality parallel data to build a translation model. The availability of large parallel corpora has rapidly increased over the past decade. However, often these newly developed parallel data contains contain significant noise. In this paper, we describe our approach for classifying good quality parallel sentence pairs from noisy parallel data. We use 10 different features within a Support Vector Machine (SVM)-based model for our classification task. We report a reasonably good classification accuracy and its positive effect on overall MT accuracy.