SenTest: Evaluating Robustness of Sentence Encoders

Contrastive learning has proven to be an effective method for pre-training models using weakly labeled data in the vision domain. Sentence transformers are the NLP counterparts to this architecture, and have been growing in popularity due to their rich and effective sentence representations. Having effective sentence representations is paramount in multiple tasks, such as information retrieval, retrieval augmented generation (RAG), and sentence comparison. Keeping in mind the deployability factor of transformers, evaluating the robustness of sentence transformers is of utmost importance. This work focuses on evaluating the robustness of the sentence encoders. We employ several adversarial attacks to evaluate its robustness. This system uses character-level attacks in the form of random character substitution, word-level attacks in the form of synonym replacement, and sentence-level attacks in the form of intra-sentence word order shuffling. The results of the experiments strongly undermine the robustness of sentence encoders. The models produce significantly different predictions as well as embeddings on perturbed datasets. The accuracy of the models can fall up to 15 percent on perturbed datasets as compared to unperturbed datasets. Furthermore, the experiments demonstrate that these embeddings does capture the semantic and syntactic structure (sentence order) of sentences. However, existing supervised classification strategies fail to leverage this information, and merely function as n-gram detectors.

Task Arithmetic with LoRA for Continual Learning

Continual learning refers to the problem where the training data is available in sequential chunks, termed "tasks". The majority of progress in continual learning has been stunted by the problem of catastrophic forgetting, which is caused by sequential training of the model on streams of data. Moreover, it becomes computationally expensive to sequentially train large models multiple times. To mitigate both of these problems at once, we propose a novel method to continually train transformer-based vision models using low-rank adaptation and task arithmetic. Our method completely bypasses the problem of catastrophic forgetting, as well as reducing the computational requirement for training models on each task. When aided with a small memory of 10 samples per class, our method achieves performance close to full-set finetuning. We present rigorous ablations to support the prowess of our method.

My Boli: Code-mixed Marathi-English Corpora, Pretrained Language Models and Evaluation Benchmarks

The research on code-mixed data is limited due to the unavailability of dedicated code-mixed datasets and pre-trained language models. In this work, we focus on the low-resource Indian language Marathi which lacks any prior work in code-mixing. We present L3Cube-MeCorpus, a large code-mixed Marathi-English (Mr-En) corpus with 5 million tweets for pretraining. We also release L3Cube-MeBERT and MeRoBERTa, code-mixed BERT-based transformer models pre-trained on MeCorpus. Furthermore, for benchmarking, we present three supervised datasets MeHate, MeSent, and MeLID for downstream tasks like code-mixed Mr-En hate speech detection, sentiment analysis, and language identification respectively. These evaluation datasets individually consist of manually annotated \url{~}12,000 Marathi-English code-mixed tweets. Ablations show that the models trained on this novel corpus significantly outperform the existing state-of-the-art BERT models. This is the first work that presents artifacts for code-mixed Marathi research. All datasets and models are publicly released at https://github.com/l3cube-pune/MarathiNLP .

Two-stage Pipeline for Multilingual Dialect Detection

Dialect Identification is a crucial task for localizing various Large Language Models. This paper outlines our approach to the VarDial 2023 shared task. Here we have to identify three or two dialects from three languages each which results in a 9-way classification for Track-1 and 6-way classification for Track-2 respectively. Our proposed approach consists of a two-stage system and outperforms other participants' systems and previous works in this domain. We achieve a score of 58.54% for Track-1 and 85.61% for Track-2. Our codebase is available publicly (https://github.com/ankit-vaidya19/EACL_VarDial2023).

A Twitter BERT Approach for Offensive Language Detection in Marathi

Automated offensive language detection is essential in combating the spread of hate speech, particularly in social media. This paper describes our work on Offensive Language Identification in low resource Indic language Marathi. The problem is formulated as a text classification task to identify a tweet as offensive or non-offensive. We evaluate different mono-lingual and multi-lingual BERT models on this classification task, focusing on BERT models pre-trained with social media datasets. We compare the performance of MuRIL, MahaTweetBERT, MahaTweetBERT-Hateful, and MahaBERT on the HASOC 2022 test set. We also explore external data augmentation from other existing Marathi hate speech corpus HASOC 2021 and L3Cube-MahaHate. The MahaTweetBERT, a BERT model, pre-trained on Marathi tweets when fine-tuned on the combined dataset (HASOC 2021 + HASOC 2022 + MahaHate), outperforms all models with an F1 score of 98.43 on the HASOC 2022 test set. With this, we also provide a new state-of-the-art result on HASOC 2022 / MOLD v2 test set.

Large Language Models for Multi-label Propaganda Detection

The spread of propaganda through the internet has increased drastically over the past years. Lately, propaganda detection has started gaining importance because of the negative impact it has on society. In this work, we describe our approach for the WANLP 2022 shared task which handles the task of propaganda detection in a multi-label setting. The task demands the model to label the given text as having one or more types of propaganda techniques. There are a total of 21 propaganda techniques to be detected. We show that an ensemble of five models performs the best on the task, scoring a micro-F1 score of 59.73%. We also conduct comprehensive ablations and propose various future directions for this work.

Temporal Word Meaning Disambiguation using TimeLMs

Meaning of words constantly changes given the events in modern civilization. Large Language Models use word embeddings, which are often static and thus cannot cope with this semantic change. Thus,it is important to resolve ambiguity in word meanings. This paper is an effort in this direction, where we explore methods for word sense disambiguation for the EvoNLP shared task. We conduct rigorous ablations for two solutions to this problem. We see that an approach using time-aware language models helps this task. Furthermore, we explore possible future directions to this problem.

Spread Love Not Hate: Undermining the Importance of Hateful Pre-training for Hate Speech Detection

Pre-training large neural language models, such as BERT, has led to impressive gains on many natural language processing (NLP) tasks. Although this method has proven to be effective for many domains, it might not always provide desirable benefits. In this paper we study the effects of hateful pre-training on low resource hate speech classification tasks. While previous studies on English language have emphasized its importance, we aim to to augment their observations with some non-obvious insights. We evaluate different variations of tweet based BERT models pre-trained on hateful, non-hateful and mixed subsets of 40M tweet dataset. This evaluation is carried for Indian languages Hindi and Marathi. This paper is an empirical evidence that hateful pre-training is not the best pre-training option for hate speech detection. We show that pre-training on non-hateful text from target domain provides similar or better results. Further, we introduce HindTweetBERT and MahaTweetBERT, the first publicly available BERT models pre-trained on Hindi and Marathi tweets respectively. We show that they provide state-of-the-art performance on hate speech classification tasks. We also release a gold hate speech evaluation benchmark HateEval-Hi and HateEval-Mr consisting of manually labeled 2000 tweets each.

Continual VQA for Disaster Response Systems

Visual Question Answering (VQA) is a multi-modal task that involves answering questions from an input image, semantically understanding the contents of the image and answering it in natural language. Using VQA for disaster management is an important line of research due to the scope of problems that are answered by the VQA system. However, the main challenge is the delay caused by the generation of labels in the assessment of the affected areas. To tackle this, we deployed pre-trained CLIP model, which is trained on visual-image pairs. however, we empirically see that the model has poor zero-shot performance. Thus, we instead use pre-trained embeddings of text and image from this model for our supervised training and surpass previous state-of-the-art results on the FloodNet dataset. We expand this to a continual setting, which is a more real-life scenario. We tackle the problem of catastrophic forgetting using various experience replay methods. Our training runs are available at: https://wandb.ai/compyle/continual_vqa_final

Efficient Gender Debiasing of Pre-trained Indic Language Models

The gender bias present in the data on which language models are pre-trained gets reflected in the systems that use these models. The model's intrinsic gender bias shows an outdated and unequal view of women in our culture and encourages discrimination. Therefore, in order to establish more equitable systems and increase fairness, it is crucial to identify and mitigate the bias existing in these models. While there is a significant amount of work in this area in English, there is a dearth of research being done in other gendered and low resources languages, particularly the Indian languages. English is a non-gendered language, where it has genderless nouns. The methodologies for bias detection in English cannot be directly deployed in other gendered languages, where the syntax and semantics vary. In our paper, we measure gender bias associated with occupations in Hindi language models. Our major contributions in this paper are the construction of a novel corpus to evaluate occupational gender bias in Hindi, quantify this existing bias in these systems using a well-defined metric, and mitigate it by efficiently fine-tuning our model. Our results reflect that the bias is reduced post-introduction of our proposed mitigation techniques. Our codebase is available publicly.