Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs (2104.08400v1)

Abstract: Abstractive conversation summarization has received much attention recently. However, these generated summaries often suffer from insufficient, redundant, or incorrect content, largely due to the unstructured and complex characteristics of human-human interactions. To this end, we propose to explicitly model the rich structures in conversations for more precise and accurate conversation summarization, by first incorporating discourse relations between utterances and action triples ("who-doing-what") in utterances through structured graphs to better encode conversations, and then designing a multi-granularity decoder to generate summaries by combining all levels of information. Experiments show that our proposed models outperform state-of-the-art methods and generalize well in other domains in terms of both automatic evaluations and human judgments. We have publicly released our code at https://github.com/GT-SALT/Structure-Aware-BART.

Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs

This paper presents a pioneering approach to tackle the challenges inherent in abstractive conversation summarization, specifically targeting the structural complexity and unstructured nature of human interactions. Unlike typical document summarization, conversations present unique difficulties such as interruptions, speaker variations, and scattered key information which impede the generation of concise and factual summaries. To address these issues, the authors propose a novel method that explicitly models discourse relations and actions within conversations using structured graphs.

Proposed Methodology

The proposed methodology revolves around the construction of two types of graphs:

1. Discourse Relation Graphs: These graphs are predicated on dependency-based discourse relations between utterances. Each utterance is considered as an Elementary Discourse Unit (EDU) and linked through 16 different types of discourse relations. This graph facilitates the identification of related utterances through relations like Question Answer Pairs, Comment, and Explanation, which aids in determining salient content crucial for generating meaningful summaries.
2. Action Graphs: The action graphs are constructed utilizing "who-doing-what" triplets derived from utterances. These graphs link actors (who), actions (doing), and objects (what), offering explicit associations between speakers and actions. This is instrumental in avoiding common factual inaccuracies such as incorrect references or misassociations within generated summaries.

The method employs a sequence-to-sequence model enhanced with these structured graphs. It uses a multi-granularity decoder to synthesize information from encoded utterances, discourse relations, and action graphs to produce summaries. The integration of structured information through two graph encoders allows for powerful representation and information synthesis.

Experimental Results and Implications

Empirical evaluations on the SAMSum corpus demonstrate that the proposed models significantly outperform existing methods, including state-of-the-art pre-trained models like BART. The major findings include:

• An improvement in ROUGE metrics over the base BART model indicates that structured graphs enhance the summarization process, demonstrating their effectiveness in handling the complexities of conversational data.
• Human evaluations confirm that summaries produced by the proposed methods excel in factualness, succinctness, and informativeness.

The models also exhibit strong generalizability, performing robustly in domain settings not utilized during training, as evidenced by results on the ADSC corpus.

Implications and Future Directions

The implications of this research are broad, offering significant advancements in the field of conversational AI. By successfully integrating structured discourse and action information, the paper paves the way for more reliable and context-aware summarization methods. The approach highlights the importance of structured representations in improving the faithfulness and informativeness of AI-generated text.

Future work could explore the application of this methodology to other conversational domains, such as emails, debates, and multi-party meetings, where the complexity of interactions is even greater. Additionally, research could focus on refining discourse analysis frameworks, improving graph construction methodologies, and enhancing model training techniques to further boost performance across various conversational contexts.

In summary, this paper contributes a substantial advancement to the field of conversation summarization, leveraging structured graphs to mitigate the inherent challenges posed by unstructured human interactions, achieving improved performance and generalization across diverse conversational domains.