Introduction
Generating meeting minutes is a complex and subjective natural language processing (NLP) task. Traditional supervised learning approaches often struggle because human-written minutes are highly inconsistent; different annotators may prioritize different information from the same meeting. Furthermore, a significant linguistic gap exists between meeting transcripts, which are typically first-person conversational speech, and minutes, which are structured third-person reports.
Team Hitachi’s submission to the AutoMin 2021 challenge addresses these issues by proposing a reference-free pipeline. Instead of training on noisy "ground-truth" minutes, the system focuses on topic-based segmentation and general dialogue summarization, followed by a novel argument mining step to ensure structural clarity.
Methodology: The Task A Pipeline
The primary generation pipeline (Task A) consists of three distinct stages designed to transform long, unstructured transcripts into readable, itemized minutes.
1. Topic Segmentation
Meeting transcripts are often too lengthy for standard transformer models. To manage this, the team utilized a Longformer model combined with a Bi-LSTM to perform "BIO-tagging." This identifies topic boundaries within the transcript. The segmenter was trained on a custom-annotated subset of the AutoMin data, where human annotators showed high agreement (0.818) on topic boundaries, suggesting that while content is subjective, the structure of a meeting is generally consistent.
2. Abstractive Summarization
Each identified topic block is processed by a BART model. To bridge the gap between dialogue and summary, the model was fine-tuned on the SAMSum corpus, a dataset of chat dialogues. This allows the system to naturally translate conversational "I/you" statements into third-person summaries without requiring direct supervision from the inconsistent AutoMin training minutes.
3. Argument Structure Construction
To avoid producing a dense "wall of text," the team introduced an argument mining parser trained on the Cornell eRulemaking Corpus (CDCP). This stage labels sentences as Tasks (subjective policies), Facts (objective information), or Discussions (opinions).
By identifying "Reason" or "Evidence" links between these labels, the system constructs a hierarchical, itemized report. This structure significantly improves readability by creating logical parent-child relationships between different points of the meeting.
Classification Tasks (B & C)
For Task B (linking minutes to transcripts) and Task C (determining if minutes originate from the same meeting), the team employed a traditional machine learning approach. They used Support Vector Machines (SVMs) and Multi-Layer Perceptrons (MLPs) fed by surface-oriented features, including:
- TF-IDF cosine similarity
- Jaccard coefficients
- Named Entity (NE) overlap ratios
- BERTScore
Key Results and Insights
Team Hitachi’s approach proved highly effective, particularly in terms of content quality. In the human evaluation for Task A, the system achieved the highest adequacy score (4.25/5) among all participants. It also performed competitively in grammatical correctness (4.34) and fluency (3.93).
For the classification tasks, the feature-based models achieved an impressive accuracy of 0.977 on Task B (English). However, the authors noted that automatic metrics like ROUGE showed weak correlation with human scores, reinforcing the idea that standard overlap-based metrics are insufficient for evaluating the quality of meeting minutes.
Conclusion and Future Work
The success of the Hitachi pipeline suggests that for highly subjective tasks like minuting, a reference-free approach focusing on structure and transfer learning (via SAMSum) is more robust than direct supervision on noisy data. Future work will focus on improving multilingual performance, as the current translation-based approach for Czech data struggled to capture the nuances of spoken language, leading to less natural results.
Summary updated