| Author | Worachot Nakduk |
| Call Number | AIT Thesis no.DSAI-23-11 |
| Subject(s) | Natural language processing (Computer science)
|
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Data Science and Artificial Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | Dialogue summarizing compresses the essential information from a conversation into a concise paragraph, enabling individuals to efficiently understand the main ideas without needing to look into the surrounding context. Recently, pre-trained language models have the potential to be helpful for dialogue summaries. Due to the engagement of sev eral participants, topic deviations, frequent references to previous statements, diverse interaction messages, and specific vocabulary, these variables contribute to increased complexity for summarizers. It can generate misinformation and generate misleading content that only covers partial facts of a conversation. Furthermore, a single discourse has the ability to incorporate multiple topics without a clearly defined boundary between them. The realistic importance of the dialogue summarization model can be better char acterized by assessing numerous aspects, such as the speaker’s goal or the specific topic under discussion. However, current summarization systems produce generic summaries that lack personalization and fail to align with customer preferences and expectations. To address this limitation, create customized different aspects of the produced summaries by engaging with the summarization process through textual submission in a format of descriptive prompts. In this paper, we present Contrastive Learning with a Topic Length Template Prompt for Dialogue Summarization. We select a topic either from the goal summary or attractive topics. These topics serve as a control signal, guided by a template’s prompt. We utilize contrastive learning methods to enhance the diversity of prompts in the template by using synonym replacement and random topics. These methods enable us to generate both positive and negative topics, thereby increasing the quantity of meaningful information available for training. Additionally, we use special tokens to highlight words and prompts to focus on important keywords. Experimentally, we show that our model can increase the ROUGE score in the DialogSum testing dataset. Our models are available at https://github.com/worachot-n/topic-length.git |
| Year | 2023 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Chaklam Silpasuwanchai |
| Examination Committee(s) | Mongkol Ekpanyapong;Dailey, Matthew N. |
| Scholarship Donor(s) | His Majesty the King’s Scholarships (Thailand) |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2023 |