Transformer Based Multimodal Summarization and Highlight Abstraction Approach for Texts and Speech Audios

Multimodal summarization is a kind of summarization application in which its inputs and/or outputs can be in different data types like text, video, and audio. In this study, a new approach based on fine tuning of different pre-trained transformers was developed for abstractive and extractive summarization of audio and text data. In the proposed method, abstractive and extractive summaries of text data are provided only as text, while extractive summaries of audio data are presented as both text and audio data. Abstractive summaries of the audio data are presented as text only. Transformers with text2text input-output relationship were used in both extractive and abstractive summarization processes of the proposed method. For the training and inference processes of audio this type of data to be handled in transformers, an ASR step was followed before the summarization step. The experimental results obtained were given in detail and compared with similar approaches in the literature. As a result of the comparison, it was seen that the proposed method achieved better performance than similar prior approaches.