Traditional subword tokenisation models like BERT often degrade when processing morphologically rich languages, informal text, or noisy user inputs common in chatbot environments. This technical presentation introduces an alternative: a robust token-free, character-level sequence framework engineered explicitly for low-resource NLP pipelines. Drawing directly from successful architectural implementations, we will pull back the curtain on how a CANINE-based encoder handles character variations natively without an explicit vocabulary. We will dive deep into the math and implementation behind integrating custom Indic phoneme mapping layers and building attention-based fusion mechanisms ($\alpha = \text{softmax}(QK^T / \sqrt{d})$) to merge orthographic and pronunciation-aware feature representations. Attendees will walk away with a comprehensive structural blueprint for building error-resilient language models, supported by empirical performance metrics and post-hoc model explainability frameworks using SHAP analysis.

Session Format & Duration: 45-Minute Technical Session / Deep Dive.

Target Audience: Machine Learning Engineers, NLP Researchers, and Data Engineers working on neural network architectures and low-resource language problems.

Technical Requirements: Projector, HDMI connection, and standard audio setup.

Presentation Outline & Timing (Designed to fill a 45-minute slot perfectly):

00:00–00:10 | The Tokenisation Problem: Discussing OOV (Out-of-Vocabulary) limits, subword breakdown in Indic scripts, and why character-level modelling is required.

00:10–00:25 | Core Architecture Deep Dive: Stepping through the CANINE encoder block, phoneme feature extraction matrices, and the mathematical mechanics of the cross-attention fusion layer.

00:25–00:35 | Benchmarks & Metrics: Comparing performance metrics (Character Error Rate vs. Word Error Rate) across CANINE, ByT5, and IndicBERT baselines, highlighting a +6.1% accuracy gain over base models.

00:35–00:45 | Explainable AI & Open Q&A: Utilising SHAP analysis to visually map feature attribution (42% Character, 37% Phonetic, 21% Fusion) to demystify black-box decoder predictions.