Introduction
Automatic summarization of meetings is a critical task for information management in professional settings. Traditionally, systems have relied on extractive methods, which select and concatenate the most salient sentences from a transcript. However, spoken language is inherently fragmented, and when combined with errors from Automatic Speech Recognition (ASR), extractive summaries often become disjointed "word salads" lacking necessary context.
This study investigates an abstractive approach, where the system interprets the underlying messages of a meeting and generates entirely new, concise text. We evaluate whether these generated abstracts improve user navigation and comprehension compared to traditional extracts, particularly in the presence of ASR noise.
Methodology
The proposed system utilizes a "Conversation Abstraction" framework consisting of four primary stages:
- Ontology Mapping: Sentences are mapped to a general conversation ontology using supervised classifiers to identify "Dialogue Acts" such as decisions, actions, or problems.
- Message Generation: The system identifies recurring patterns across the meeting. For instance, if multiple participants express positive views on a specific entity, a "RepeatedPositiveMessage" is generated.
- Content Selection: Integer Linear Programming (ILP) is employed to select the most informative messages, balancing the weight of individual sentences against the significance of the higher-level messages they constitute.
- Summary Generation: Using the
simpleNLGrealizer, the system plans and generates natural language text, aggregating information to create a cohesive narrative.
Experimental Results
The system was evaluated using the AMI Meeting Corpus. Nineteen participants were tasked with browsing meeting transcripts under a strict 15-minute deadline using different summary types.
Performance on Manual Transcripts
On high-quality manual transcripts, automatic abstracts (AAM) demonstrated superior performance over human-selected extracts (EHM) in several key metrics. Specifically, automatic abstracts were rated significantly higher for coherence (Q3) and usefulness in navigating the transcript (Q5).
Resilience to ASR Noise
The advantages of abstractive summarization become even more pronounced when dealing with noisy ASR transcripts. While ASR errors often render extractive "snippets" incomprehensible, the abstractive system acts as a noise filter by mapping noisy text to high-level concepts before generating clean output.
As shown in the ASR condition, automatic abstracts were rated nearly twice as high for coherence compared to extracts ($p = 0.001$). Users reported that extractive summaries felt like "snippets of conversation out of context," whereas the abstractive summaries provided a clearer overview of the meeting's progression.
Discussion and Conclusion
The study highlights that for spoken dialogue, the relationship between sentences is often more important than the sentences themselves. Abstractive summarization effectively addresses the "snippet problem" by providing the necessary context that extractive methods lack. Furthermore, the system's ability to reduce the cognitive effort required to review a meeting makes it a highly effective tool for meeting navigation.
Future work will focus on improving entity recognition beyond simple noun phrases and optimizing the balance between sentence and message weights to further enhance summary specificity.
Summary updated