Cost-Performance Optimization for Processing Low-Resource Language Tasks Using Commercial LLMs

Large Language Models (LLMs) exhibit impressive zero/few-shot inference and generation quality for high-resource languages(HRLs). A few of them have been trained in low-resource languages (LRLs) and give decent performance. Owing to the prohibitive costs of training LLMs, they are usually used as a network service, with the client charged by the count of input and output tokens. The number of tokens strongly depends on the script and language, as well as the LLM's sub-word vocabulary. We show that LRLs are at a pricing disadvantage, because the well-known LLMs produce more tokens for LRLs than HRLs. This is because most currently popular LLMs are optimized for HRL vocabularies. Our objective is to level the playing field: reduce the cost of processing LRLs in contemporary LLMs while ensuring that predictive and generative qualities are not compromised. As means to reduce the number of tokens processed by the LLM, we consider code-mixing, translation, and transliteration of LRLs to HRLs. We perform an extensive study using the IndicXTREME dataset, covering 15 Indian languages, while using GPT-4 (one of the costliest LLM services released so far) as a commercial LLM. We observe and analyze interesting patterns involving token count, cost,and quality across a multitude of languages and tasks. We show that choosing the best policy to interact with the LLM can reduce cost by 90% while giving better or comparable performance, compared to communicating with the LLM in the original LRL.

How (un)ethical are instruction-centric responses of LLMs? Unveiling the vulnerabilities of safety guardrails to harmful queries

In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models.

Mask-up: Investigating Biases in Face Re-identification for Masked Faces

AI based Face Recognition Systems (FRSs) are now widely distributed and deployed as MLaaS solutions all over the world, moreso since the COVID-19 pandemic for tasks ranging from validating individuals' faces while buying SIM cards to surveillance of citizens. Extensive biases have been reported against marginalized groups in these systems and have led to highly discriminatory outcomes. The post-pandemic world has normalized wearing face masks but FRSs have not kept up with the changing times. As a result, these systems are susceptible to mask based face occlusion. In this study, we audit four commercial and nine open-source FRSs for the task of face re-identification between different varieties of masked and unmasked images across five benchmark datasets (total 14,722 images). These simulate a realistic validation/surveillance task as deployed in all major countries around the world. Three of the commercial and five of the open-source FRSs are highly inaccurate; they further perpetuate biases against non-White individuals, with the lowest accuracy being 0%. A survey for the same task with 85 human participants also results in a low accuracy of 40%. Thus a human-in-the-loop moderation in the pipeline does not alleviate the concerns, as has been frequently hypothesized in literature. Our large-scale study shows that developers, lawmakers and users of such services need to rethink the design principles behind FRSs, especially for the task of face re-identification, taking cognizance of observed biases.

GRAFFORD: A Benchmark Dataset for Testing the Knowledge of Object Affordances of Language and Vision Models

We investigate the knowledge of object affordances in pre-trained language models (LMs) and pre-trained Vision-Language models (VLMs). Transformers-based large pre-trained language models (PTLM) learn contextual representation from massive amounts of unlabeled text and are shown to perform impressively in downstream NLU tasks. In parallel, a growing body of literature shows that PTLMs fail inconsistently and non-intuitively, showing a lack of reasoning and grounding. To take a first step toward quantifying the effect of grounding (or lack thereof), we curate a novel and comprehensive dataset of object affordances -- GrAFFORD, characterized by 15 affordance classes. Unlike affordance datasets collected in vision and language domains, we annotate in-the-wild sentences with objects and affordances. Experimental results reveal that PTLMs exhibit limited reasoning abilities when it comes to uncommon object affordances. We also observe that pre-trained VLMs do not necessarily capture object affordances effectively. Through few-shot fine-tuning, we demonstrate improvement in affordance knowledge in PTLMs and VLMs. Our research contributes a novel dataset for language grounding tasks, and presents insights into LM capabilities, advancing the understanding of object affordances. Codes and data are available at https://github.com/sayantan11995/Affordance

Zero shot VLMs for hate meme detection: Are we there yet?

Multimedia content on social media is rapidly evolving, with memes gaining prominence as a distinctive form. Unfortunately, some malicious users exploit memes to target individuals or vulnerable communities, making it imperative to identify and address such instances of hateful memes. Extensive research has been conducted to address this issue by developing hate meme detection models. However, a notable limitation of traditional machine/deep learning models is the requirement for labeled datasets for accurate classification. Recently, the research community has witnessed the emergence of several visual language models that have exhibited outstanding performance across various tasks. In this study, we aim to investigate the efficacy of these visual language models in handling intricate tasks such as hate meme detection. We use various prompt settings to focus on zero-shot classification of hateful/harmful memes. Through our analysis, we observe that large VLMs are still vulnerable for zero-shot hate meme detection.

Low-Resource Counterspeech Generation for Indic Languages: The Case of Bengali and Hindi

With the rise of online abuse, the NLP community has begun investigating the use of neural architectures to generate counterspeech that can "counter" the vicious tone of such abusive speech and dilute/ameliorate their rippling effect over the social network. However, most of the efforts so far have been primarily focused on English. To bridge the gap for low-resource languages such as Bengali and Hindi, we create a benchmark dataset of 5,062 abusive speech/counterspeech pairs, of which 2,460 pairs are in Bengali and 2,602 pairs are in Hindi. We implement several baseline models considering various interlingual transfer mechanisms with different configurations to generate suitable counterspeech to set up an effective benchmark. We observe that the monolingual setup yields the best performance. Further, using synthetic transfer, language models can generate counterspeech to some extent; specifically, we notice that transferability is better when languages belong to the same language family.

Context Matters: Pushing the Boundaries of Open-Ended Answer Generation with Graph-Structured Knowledge Context

In the continuously advancing AI landscape, crafting context-rich and meaningful responses via Large Language Models (LLMs) is essential. Researchers are becoming more aware of the challenges that LLMs with fewer parameters encounter when trying to provide suitable answers to open-ended questions. To address these hurdles, the integration of cutting-edge strategies, augmentation of rich external domain knowledge to LLMs, offers significant improvements. This paper introduces a novel framework that combines graph-driven context retrieval in conjunction to knowledge graphs based enhancement, honing the proficiency of LLMs, especially in domain specific community question answering platforms like AskUbuntu, Unix, and ServerFault. We conduct experiments on various LLMs with different parameter sizes to evaluate their ability to ground knowledge and determine factual accuracy in answers to open-ended questions. Our methodology GraphContextGen consistently outperforms dominant text-based retrieval systems, demonstrating its robustness and adaptability to a larger number of use cases. This advancement highlights the importance of pairing context rich data retrieval with LLMs, offering a renewed approach to knowledge sourcing and generation in AI systems. We also show that, due to rich contextual data retrieval, the crucial entities, along with the generated answer, remain factually coherent with the gold answer.

Probing LLMs for hate speech detection: strengths and vulnerabilities

Recently efforts have been made by social media platforms as well as researchers to detect hateful or toxic language using large language models. However, none of these works aim to use explanation, additional context and victim community information in the detection process. We utilise different prompt variation, input information and evaluate large language models in zero shot setting (without adding any in-context examples). We select three large language models (GPT-3.5, text-davinci and Flan-T5) and three datasets - HateXplain, implicit hate and ToxicSpans. We find that on average including the target information in the pipeline improves the model performance substantially (~20-30%) over the baseline across the datasets. There is also a considerable effect of adding the rationales/explanations into the pipeline (~10-20%) over the baseline across the datasets. In addition, we further provide a typology of the error cases where these large language models fail to (i) classify and (ii) explain the reason for the decisions they take. Such vulnerable points automatically constitute 'jailbreak' prompts for these models and industry scale safeguard techniques need to be developed to make the models robust against such prompts.

BanglaAbuseMeme: A Dataset for Bengali Abusive Meme Classification

The dramatic increase in the use of social media platforms for information sharing has also fueled a steep growth in online abuse. A simple yet effective way of abusing individuals or communities is by creating memes, which often integrate an image with a short piece of text layered on top of it. Such harmful elements are in rampant use and are a threat to online safety. Hence it is necessary to develop efficient models to detect and flag abusive memes. The problem becomes more challenging in a low-resource setting (e.g., Bengali memes, i.e., images with Bengali text embedded on it) because of the absence of benchmark datasets on which AI models could be trained. In this paper we bridge this gap by building a Bengali meme dataset. To setup an effective benchmark we implement several baseline models for classifying abusive memes using this dataset. We observe that multimodal models that use both textual and visual information outperform unimodal models. Our best-performing model achieves a macro F1 score of 70.51. Finally, we perform a qualitative error analysis of the misclassified memes of the best-performing text-based, image-based and multimodal models.

Auditing Gender Analyzers on Text Data

AI models have become extremely popular and accessible to the general public. However, they are continuously under the scanner due to their demonstrable biases toward various sections of the society like people of color and non-binary people. In this study, we audit three existing gender analyzers -- uClassify, Readable and HackerFactor, for biases against non-binary individuals. These tools are designed to predict only the cisgender binary labels, which leads to discrimination against non-binary members of the society. We curate two datasets -- Reddit comments (660k) and, Tumblr posts (2.05M) and our experimental evaluation shows that the tools are highly inaccurate with the overall accuracy being ~50% on all platforms. Predictions for non-binary comments on all platforms are mostly female, thus propagating the societal bias that non-binary individuals are effeminate. To address this, we fine-tune a BERT multi-label classifier on the two datasets in multiple combinations, observe an overall performance of ~77% on the most realistically deployable setting and a surprisingly higher performance of 90% for the non-binary class. We also audit ChatGPT using zero-shot prompts on a small dataset (due to high pricing) and observe an average accuracy of 58% for Reddit and Tumblr combined (with overall better results for Reddit). Thus, we show that existing systems, including highly advanced ones like ChatGPT are biased, and need better audits and moderation and, that such societal biases can be addressed and alleviated through simple off-the-shelf models like BERT trained on more gender inclusive datasets.