Online gaming involves large amount of people forming a large community of players who interact in real time. Toxic behavior in online chat is common and can harm players by deterring them. Thus, automated moderation is a necessity but difficult because game chat mixes domain-specific slang, deliberate obfuscation, informal "gamer" language , and tiny support for categories such as threats and extremism. This paper describes the TAGA (Token-Attribution Guided Attention) system submitted to the EEUCA 2026 Shared Task on Understanding Toxic Behavior in Gaming Communities. We propose TAGA, an architecture that employs a leave-one-out attribution method using the Detoxify toxicity scorer to compute per-token attribution scores across multiple toxicity dimensions, which are then projected into the learned attention biases that steer the model toward toxicity-indicative tokens. By preparing a five phase ablation study, we demonstrate that each component: domain-specific preprocessing, focal loss with label smoothing, attribution-guided attention pooling, and dual-model Detoxify features with strategic oversampling contributes to a cumulative gain in macro-F1 score points over the DeBERTa-v3-base baseline reported. The final system achieves a test macro-F1 score of 0.618 and, importantly, produces non-zero predictions for extreme data imbalance present in the dataset used in the shared task.

Models, such as BERT, have made a significant breakthrough in the Natural Language Processing (NLP) domain solving 11+ tasks. This is achieved by training on a large scale of unlabelled text resources and leveraging Transformers architecture making it the “Jack of all NLP trades”. However, one of the popular and challenging tasks in Sequence Classification is Short Text Classification (STC). Short Texts face the problem of being short, equivocal, and non-standard. In this paper, we address two major problems: 1. Improving STC tasks performance in Japanese language which consists of many varieties and dialects. 2. Building a light-weight Japanese BERT model with cross-domain functionality and comparable accuracy with State of the Art (SOTA) BERT models. To solve this, we propose a novel cross-domain scalable model called JLBert, which is pre-trained on a rich, diverse and less explored Japanese e-commerce corpus. We present results from extensive experiments to show that JLBert is outperforming SOTA Multilingual and Japanese specialized BERT models on three Short Text datasets by approx 1.5% across various domain.