Transformer-based Job Ranking Systems (JRSs) are vulnerable to societal biases inherited in unbalanced datasets. These biases often manifest as unjust job rankings, particularly disadvantaging candidates of different genders. Most bias mitigation techniques leverage candidates’ gender and align gender distributions within the embeddings of JRSs to mitigate bias. While such methods effectively align distributional properties and make JRSs agnostic to gender, they frequently fall short in addressing empirical fairness metrics, such as the performance gap across genders. In this study, we shift our attention from candidate gender to mitigate bias based on gendered language in job advertisements. We propose a novel neutrality score based on automatically discovered biased words in job ads and use it to re-rank the model’s decisions. We evaluate our method by comparing it with different bias mitigation strategies and empirically demonstrate that our proposed method not only improves fairness but can also enhance the model’s performance.

This paper presents an English to Bhojpuri machine translation (MT) system developed for the WMT25 General MT Shared Task. Given the low-resource nature of Bhojpuri, we adopt a two-stage training pipeline: unsupervised pretraining followed by supervised fine-tuning. During pretraining, we use a 300,000-sentence corpus comprising 70% Bhojpuri monolingual data and 30% English data to establish language grounding. The fine-tuning stage utilizes 29,749 bilingual English to Bhojpuri sentence pairs (including training, validation, and test sets). To adapt the system to instruction-following scenarios, we apply a novel optimization strategy: Contrastive Preference Optimization (CPO). This technique enables the model to capture fine-grained translation preferences and maintain semantic fidelity in instruction-tuned settings. We evaluate our system across multiple metrics, demonstrating moderate performance in low-resource MT tasks, particularly in diverse domains such as literary, news, social, and speech.

The study critically examines various Machine Translation systems, particularly focusing on Large Language Models, using the WMT25 Legal Domain Test Suite for translating English into Hindi. It utilizes a dataset of 5,000 sentences designed to capture the complexity of legal texts, based on word frequency ranges from 5 to 54. Each frequency range contains 100 sentences, collectively forming a corpus that spans from simple legal terms to intricate legal provisions. Six metrics were used to evaluate the performance of the system: BLEU, METEOR, TER, CHRF++, BERTScore and COMET. The findings reveal diverse capabilities and limitations of LLM architectures in handling complex legal texts. Notably, Gemini-2.5-Pro, Claude-4 and ONLINE-B topped the performance charts in terms fo human evaluation, showcasing the potential of LLMs for nuanced trans- lation. Despite these advances, the study identified areas for further research, especially in improving robustness, reliability, and explainability for use in critical legal contexts. The study also supports the WMT25 subtask focused on evaluating weaknesses of large language models (LLMs). The dataset and related resources are publicly available at https://github.com/helloboyn/WMT25-TS.

This paper presents a new neural machine translation (NMT) system aimed at low-resource language pairs: English to Kokborok and English to Bodo. The framework leverages the LLaMA3-8B-Instruct model along with LoRA-based parameter-efficient fine-tuning. For translating into Kokborok, the model undergoes an initial continued pre-training phase on a dataset containing 75,000 Kokborok and 25,000 English monolingual sentences, followed by instruction-tuning. This tuning uses a reformulated version of WMT25 dataset, adapted to the Alpaca format to support instructional goals. In the Bodo translation, the model is pre-trained on a more extensive dataset of 350,000 Bodo and 125,000 English sentences, using a similar instruction-tuning approach. LoRA adapters are used to modify the large LLaMA3 model for these low-resource settings. Testing with the WMT25 test dataset reveals modest translation results, highlighting the difficulties in translating for low-resource languages. Translating English to Bodo, the model achieved a BLEU score of 4.38, a TER of 92.5, and a chrF score of 35.4. For English to Kokborok, it yielded scores of 5.59 in chrF, 105.4 in TER, and 0.17 in BLEU. These results underscore the intricacies of the task and highlight the critical need for further data collection, domain-specific adaptations, and improvements in model design to better support underrepresented languages.

We propose a training algorithm based on retrieval-augmented generation (RAG) to obtain the most similar training samples. The training samples obtained are used as a reference to perform contextual learning-based fine-tuning of large language models (LLMs). We use the proposed method to generate headlines and extract numerical values from unstructured text. Models are made aware of the presence of numbers in the unstructured text with extended markup language (XML) tags specifically designed to capture the numbers. The headlines of unstructured text are preprocessed to wrap the number and then presented to the model. A number of mathematical operations are also passed as references to cover the chain-of-thought (COT) approach. Therefore, the model can calculate the final value passed to a mathematical operation. We perform the validation of numbers as a post-processing step to verify whether the numerical value calculated by the model is correct or not. The automatic validation of numbers in the generated headline helped the model achieve the best results in human evaluation among the methods involved.

Large pre-trained language models contain societal biases and carry along these biases to downstream tasks. Current in-processing bias mitigation approaches (like adversarial training) impose debiasing by updating a model’s parameters, effectively transferring the model to a new, irreversible debiased state. In this work, we propose a novel approach to develop stand-alone debiasing functionalities separate from the model, which can be integrated into the model on-demand, while keeping the core model untouched. Drawing from the concept of AdapterFusion in multi-task learning, we introduce DAM (Debiasing with Adapter Modules) – a debiasing approach to first encapsulate arbitrary bias mitigation functionalities into separate adapters, and then add them to the model on-demand in order to deliver fairness qualities. We conduct a large set of experiments on three classification tasks with gender, race, and age as protected attributes. Our results show that DAM improves or maintains the effectiveness of bias mitigation, avoids catastrophic forgetting in a multi-attribute scenario, and maintains on-par task performance, while granting parameter-efficiency and easy switching between the original and debiased models.

Diverse headline generation is an NLP task where given a news article, the goal is to generate multiple headlines that are true to the content of the article but are different among themselves. This task aims to exhibit and exploit semantically similar one-to-many relationships between a source news article and multiple target headlines. Toward this, we propose a novel model called DIVHSK. It has two components:KEYSELECT for selecting the important keywords, and SEQGEN, for finally generating the multiple diverse headlines. In KEYSELECT, we cluster the self-attention heads of the last layer of the pre-trained encoder and select the most-attentive theme and general keywords from the source article. Then, cluster-specific keyword sets guide the SEQGEN, a pre-trained encoder-decoder model, to generate diverse yet semantically similar headlines. The proposed model consistently outperformed existing literature and our strong baselines and emerged as a state-of-the-art model. We have also created a high-quality multi-reference headline dataset from news articles.

Societal biases are reflected in large pre-trained language models and their fine-tuned versions on downstream tasks. Common in-processing bias mitigation approaches, such as adversarial training and mutual information removal, introduce additional optimization criteria, and update the model to reach a new debiased state. However, in practice, end-users and practitioners might prefer to switch back to the original model, or apply debiasing only on a specific subset of protected attributes. To enable this, we propose a novel modular bias mitigation approach, consisting of stand-alone highly sparse debiasing subnetworks, where each debiasing module can be integrated into the core model on-demand at inference time. Our approach draws from the concept of diff pruning, and proposes a novel training regime adaptable to various representation disentanglement optimizations. We conduct experiments on three classification tasks with gender, race, and age as protected attributes. The results show that our modular approach, while maintaining task performance, improves (or at least remains on-par with) the effectiveness of bias mitigation in comparison with baseline finetuning. Particularly on a two-attribute dataset, our approach with separately learned debiasing subnetworks shows effective utilization of either or both the subnetworks for selective bias mitigation.

In low-resource languages, the amount of training data is limited. Hence, the model has to perform well in unseen sentences and syntax on which the model has not trained. We propose a method that addresses the problem through an encoder and an ensemble of language models. A language-specific language model performed poorly when compared to a multilingual language model. So, the multilingual language model checkpoint is fine-tuned to a specific language. A novel approach of one hot encoder is introduced between the model outputs and the CRF to combine the results in an ensemble format. Our team, Infrrd.ai, competed in the MultiCoNER competition. The results are encouraging where the team is positioned within the top 10 positions. There is less than a 4% percent difference from the third position in most of the tracks that we participated in. The proposed method shows that the ensemble of models with a multilingual language model as the base with the help of an encoder performs better than a single language-specific model.

In this paper, we describe our approaches for task six of Social Media Mining for Health Applications (SMM4H) shared task in 2021. The task is to classify twitter tweets containing COVID-19 symptoms in three classes (self-reports, non-personal reports & literature/news mentions). We implemented BERT and XLNet for this text classification task. Best result was achieved by XLNet approach, which is F1 score 0.94, precision 0.9448 and recall 0.94448. This is slightly better than the average score, i.e. F1 score 0.93, precision 0.93235 and recall 0.93235.