We present a constrained submission to the WMT 2026 Low-Resource Arabic-Asian Language Translation shared task for Arabic–Hindi and Arabic–Bangla translation. Our system fine-tunes the NLLB-200 distilled 600M model jointly on all four translation directions using the official task data. At inference, instead of relying on a single best checkpoint, we retain multiple training checkpoints and generate candidate translations from each checkpoint. We then apply checkpoint-pool Minimum Bayes Risk (MBR) decoding with chrF as the utility function to select the final translation. This approach exploits diversity across training checkpoints without additional training runs. On the official evaluation, our primary system ranked 4th of 8 for Hindi→Arabic, 3rd of 6 for Bangla→Arabic, 3rd of 10 for Arabic→Hindi, and 2nd of 7 for Arabic→Bangla.

We present constrained WMT 2026 submissions for low-resource Northeastern Indian languages using NLLB-200 Distilled 1.3B with multilingual LoRA adapters trained only on the official organizer-provided parallel data. The adapted systems improve over the submitted baselines for both Mizo directions and Bengali-script Manipuri→English, but show substantial variation across languages and directions. We further manually analyze approximately 100 Indic→English outputs and validate the systems on WMT 2025 data, highlighting generation artifacts and the challenges of parameter-efficient adaptation across languages, scripts, and translation directions.

We present our system for the SHROOM-Visions 2026 shared task on fine-grained hallucination detection in vision-language model (VLM) outputs. Our approach uses three token-level internal signals from Qwen2.5-VL-3B-Instruct: hidden representations, the change in token confidence when the image is removed, and predictive uncertainty. A lightweight classifier combines these signals to predict hallucinated spans and their confidence, while LoRA enables parameter-efficient adaptation of the VLM backbone. Our submission achieves IoU scores of 0.36–0.42 and Spearman correlations of 0.41–0.46 on English, French, and Italian, ranking 10th in both French and Italian. We additionally analyze continued-training experiments that improved our local validation scores but consistently degraded official test performance. Our analysis attributes this discrepancy primarily to threshold tuning on a small validation split, which was only partially mitigated by subsequent adjustments. Increasing the backbone size from 3B to 7B parameters did not resolve the issue. These findings highlight the sensitivity of hallucination detection systems to validation size and decision-threshold selection.

Sentiment analysis plays a pivotal role in understanding public opinion, particularly in the political domain where the portrayal of entities in news articles influences public perception. In this paper, we investigate the effectiveness of Large Language Models (LLMs) in predicting entity-specific sentiment from political news articles. Leveraging zero-shot and few-shot strategies, we explore the capability of LLMs to discern sentiment towards political entities in news content. Employing a chain-of-thought (COT) approach augmented with rationale in few-shot in-context learning, we assess whether this method enhances sentiment prediction accuracy. Our evaluation on sentiment-labeled datasets demonstrates that LLMs, outperform fine-tuned BERT models in capturing entity-specific sentiment. We find that learning in-context significantly improves model performance, while the self-consistency mechanism enhances consistency in sentiment prediction. Despite the promising results, we observe inconsistencies in the effectiveness of the COT prompting method. Overall, our findings underscore the potential of LLMs in entity-centric sentiment analysis within the political news domain and highlight the importance of suitable prompting strategies and model architectures.

In today’s media landscape, where news outlets play a pivotal role in shaping public opinion, it is imperative to address the issue of sentiment manipulation within news text. News writers often inject their own biases and emotional language, which can distort the objectivity of reporting. This paper introduces a novel approach to tackle this problem by reducing the polarity of latent sentiments in news content. Drawing inspiration from adversarial attack-based sentence perturbation techniques and a promptbased method using ChatGPT, we employ transformation constraints to modify sentences while preserving their core semantics. Using three perturbation methods—replacement, insertion, and deletion—coupled with a contextaware masked language model, we aim to maximize the desired sentiment score for targeted news aspects through a beam search algorithm. Our experiments and human evaluations demonstrate the effectiveness of these two models in achieving reduced sentiment polarity with minimal modifications while maintaining textual similarity, fluency, and grammatical correctness. Comparative analysis confirms the competitive performance of the adversarial attack-based perturbation methods and promptbased methods, offering a promising solution to foster more objective news reporting and combat emotional language bias in the media.

The task of event extraction (EE) aims to find the events and event-related argument information from the text and represent them in a structured format. Most previous works try to solve the problem by separately identifying multiple substructures and aggregating them to get the complete event structure. The problem with the methods is that it fails to identify all the interdependencies among the event participants (event-triggers, arguments, and roles). In this paper, we represent each event record in a unique tuple format that contains trigger phrase, trigger type, argument phrase, and corresponding role information. Our proposed pointer network-based encoder-decoder model generates an event tuple in each time step by exploiting the interactions among event participants and presenting a truly end-to-end solution to the EE task. We evaluate our model on the ACE2005 dataset, and experimental results demonstrate the effectiveness of our model by achieving competitive performance compared to the state-of-the-art methods.

This paper presents the IITKGP contribution at the Technical DOmain Identification (TechDOfication) shared task at ICON 2020. In the preprocessing stage, we applied part-of-speech (PoS) taggers and dependency parsers to tag the data. We trained a graph convolution neural network (GCNN) based system that uses the tokens along with their PoS and dependency relations as features to identify the domain of a given document. We participated in the subtasks for coarse-grained domain classification in the English (Subtask 1a), Bengali (Subtask 1b) and Hindi language (Subtask 1d), and, the subtask for fine-grained domain classification task within Computer Science domain in English language (Subtask 2a).