llmlingua-2-xlm-roberta-large-meetingbank

Classifies individual tokens in meeting transcripts as important or unimportant using XLM-RoBERTa-large architecture fine-tuned on the MeetingBank dataset. The model performs sequence-level token classification by processing the entire transcript context through a 24-layer transformer encoder, then applying a classification head to each token position to predict importance scores. This enables selective compression of meeting content by identifying which tokens carry semantic weight for downstream LLM processing.

Leverages XLM-RoBERTa's cross-lingual transfer capabilities to understand and classify tokens across 100+ languages using a single unified model. The architecture uses shared multilingual embeddings and transformer layers trained on Common Crawl data, allowing the fine-tuned meeting classifier to generalize to non-English meeting transcripts without language-specific retraining. Token representations are contextualized through bidirectional attention, enabling the model to disambiguate polysemous words and understand language-specific importance markers.

Performs token importance classification using bidirectional transformer attention, where each token's importance score is computed by attending to all surrounding tokens in the full meeting transcript. The model uses 24 transformer layers with multi-head attention (16 heads, 1024 hidden dimensions) to build rich contextual representations, then applies a classification head to predict token importance. This bidirectional approach enables the model to understand that a token's importance depends on its discourse role (e.g., a speaker name is important if followed by a decision, but unimportant if just introducing a comment).

Processes multiple meeting transcripts in parallel using dynamic padding, where sequences are padded to the longest length in the batch rather than a fixed maximum length. The model uses HuggingFace's DataCollator pattern to group variable-length transcripts into batches, apply padding/truncation, and generate attention masks that tell the transformer to ignore padding tokens. This enables efficient GPU utilization by minimizing wasted computation on padding while maintaining correctness of token-level predictions.

Enables selective compression of meeting transcripts by filtering tokens based on their importance scores, with configurable compression ratios (e.g., keep top 50% of tokens, remove bottom 50%). The model outputs importance scores for each token, which are then used to rank and filter tokens, producing a compressed transcript that retains high-importance content. This can be applied at different compression levels (aggressive: 30% of tokens, moderate: 60%, conservative: 80%) to trade off between compression and information retention.