A Study of Summarization and Keyword Extraction Function in Meeting Note Generation System from Voice Records
Amma Liesvarastranta Haz, Nobuo Funabiki, Evianita Dewi Fajrianti, Sritrusta Sukaridhoto · 2023
Currently, we are studying a meeting note generation system to automatically make a note from voice records. Previously, we implemented the first function of the system to convert voice records to texts using the Whisper model. However, the current implementation suffers from a lack of coherency due to the vast text from audio. Therefore summary and keywords should be derived from the text to improve comprehension. In this paper, we propose the second function of the system to make summaries and extract keywords from texts. The BART model is used to generate human-like summaries from the semantic idea of texts. The keyword extraction method ranks the importance of each word or phrase in texts. For evaluations, we apply the proposal to three voice records with varied durations and compare the generated summary and keywords with manually made ones. The results show that for the summary, the average ROUGE-1 is 50.02%, ROUGE-2 is 24.04%, and ROUGE-L is 43.14%, and for keywords, the average cosine similarity is 54.09%. Thus, the validity of the proposal is confirmed. In future works, we will focus on incorporating additional methods to extract the memo from audio and text from the documents.