Automatic Minuting: A Pipeline Method for Generating Minutes from Multi-Party Meeting Proceedings

Automatic Minuting: A Pipeline Method for Generating Minutes from Multi-Party Meeting Transcripts Kartik Shinde†, Tirthankar Ghosal‡, Muskaan Singh*, and Ondˇrej Bojar‡ †Indian Institute of Technology Patna, Bihta, Bihar, India ‡Charles University, Faculty of Mathematics and Physics, ´UFAL, Czech Republic *Speech and Audio Processing Group, IDIAP Research Institute, Martigny, Switzerland kartik 1901ce16@iitp.ac.in, msingh@idiap.ch (ghosal,bojar)@ufal.mff.cuni.cz Abstract Automatically generating meeting minutes is a challenging yet time-relevant problem in speech and natural language processing. Nowadays, meeting minutes seem more cru- cial than ever due to the manifold rise of on- line meetings. However, automatic minut- ing is not straightforward for various reasons: obtaining transcriptions of sufficient quality, summarizing long dialogue discourse, retain- ing topical relevance and coverage, handling redundancies and small talk, etc. This pa- per presents our investigations on a pipelined approach to automatically generate meeting minutes using a BART model (Bidirectional and Auto-Regressive Transformers) trained on multi-party dialogue summarization datasets. We achieve comparable results with our sim- ple yet intuitive method with respect to pre- vious large and computationally heavy state- of-the-art models. We make our code avail- able at https://github.com/ELITR/ minuting-pipeline. 1 Introduction Ever since most of our interactions went virtual, the need for automatic support to run online meetings became essential. Due to frequent meetings and the resulting context switching, people are experienc- ing an information overload (Fauville et al., 2021) of epic proportions. Hence a tool to automatically summarize a meeting transcript would be a valuable addition to the virtual workplace. Automatic Minut- ing is the task of generating bullet-point meeting minutes from multi-party meeting transcripts. The AutoMin shared task at Interspeech 2021 (Ghosal et al., 2021) is a community-wide effort in this direction. Organizers of AutoMin (Ghosal et al., 2022a) released a medium-scale annotated corpus (Nedoluzhko et al., 2022) of transcript-minute pairs for conducting the shared task. Automatic Minuting is close to summarization but not the same; subtle differences exist. Summariza- tion aims at generating a concise and coherent text summary. It often purposely removes some less crit- ical information; minuting is more inclined towards adequately capturing the entire contents of the meet- ing (coverage is probably more significant than co- herence and conciseness). Summarizing spoken multi-party dialogues comes with challenges: incorrect or noisy auto- mated speech recognition (ASR) outputs, long discourse, topical shifts, the dialogue turns, re- dundancies and small talk, etc. Hence we deem automatic minuting to be more difficult than text summarization. Due to the variety of sub-problems associated with this task, we adopt a pipelined approach. Our method encompasses (i) pre-processing the ASR- generated meeting transcripts to drop redundancies and noise, followed by (ii) unsupervised topical seg- mentation, and finally (iii) summarizing each seg- ment of the discourse with a BART model (Raffel et al., 2019) pre-trained on a large-scale dialogue sum- marization dataset. Our initial investigation yields encouraging results. The obtained minutes resemble the human gold standard in terms of readability and coverage. Our main contribution lies in developing a lightweight, easy-to-implement, and efficient au- tomatic minuting pipeline by leveraging pre-trained Transformer-based language …

Content truncated to 500 words