Cascina Elisa Politecnico di Torino [email protected]
Pellegrino Andrea Politecnico di Torino [email protected]
Tozzi Lorenzo Politecnico di Torino [email protected]
GitHub repository:
https://github.com/andry2327/Emotions-and-Topics-Injection-for-Abstractive-Dialogue-Summarization
PoliTo research group:
Abstract
<aside> <img src="/icons/news_purple.svg" alt="/icons/news_purple.svg" width="40px" /> In this study, we introduce novel extensions to enhance abstractive chat summarization, built on the SICK framework that underscored the advantages of integrating commonsense knowledge among dialogues participants into the task. The first novelty involves extracting emotions from the dialogue and incorporating them into the model; the second extension regards the injection of predicted relevant custom topics into the summarization process. Our framework exhibits promising results and set the stage for future investigations. The outcomes suggest that injecting commonsense knowledge enriched by extracted topics has a major impact when compared to emotion injection. Nevertheless both the extensions enhance the model, contributing to the generation of more informative and consistent summaries.
</aside>
PDF:
Summarization_Team_14_s318105_s309855_s317330.pdf
Abstractive dialogue summarization is a type of artificial intelligence (AI) technique that aims to automatically generate concise summaries of conversations. Unlike simpler methods that just pick out existing sentences, abstractive summarization creates a new, shortened version of the dialogue that captures the most important information.
We aimed to make improvements over an already existing architecture for abstractive dialogue summarization, the SICK framework:
SICK/SICK++ architecture
Mind the Gap! Injecting Commonsense Knowledge for Abstractive...
Our study introduces two key novelties:
Emotions injection: Leveraging a fine-tuned BERT model, we aim to predict the emotion expressed in each utterance of the dialogue. The emotion labels are then concatenated to the utterances.
Topics injection: The process involves the use of a zero- shot classifier to identify potential topics covered by the dialogue and append them to the last utterance of the dialogue.
