Towards Reliable and Inclusive Natural Language Generation
Public Deposited- Creator
-
Zhang, Shiyue
- College of Arts and Sciences, Department of Computer Science
-
Zhang, Shiyue
- Abstract
- Natural language generation (NLG) is an important subfield of natural language processing (NLP) that produces natural language output. Despite notable advancements made by large-scale pre-trained language models in NLG, there remain several unresolved challenges. This thesis aims to enhance NLG from two significant aspects: reliability and inclusiveness. For reliability, on the one hand, we introduce novel training objectives that improve the alignment of language generation models with desired model behaviors. To improve the answerability of model-generated questions, we use a question answering model to provide additional rewards to a question generation model, encouraging the production of more answerable questions. In addition, we propose to train language models with a mixture of forward and reverse cross-entropies, demonstrating that the resulting models yield better generated text without complex decoding strategies. On the other hand, we propose novel evaluation methods to assess the performance of NLG models accurately and comprehensively. By combining human and automatic evaluations, we strike a balance between reliability and reproducibility. We delve into the unexplored issue of unfaithfulness in extractive summaries and conclude that extractive summarization does not guarantee faithfulness. For inclusiveness, we extend the coverage of NLG techniques to low-resource or endangered languages. We develop the first machine translation system for supporting translation between Cherokee, an endangered Native American language, and English, and we propose a roadmap for utilizing NLP to support language revitalization efforts. Additionally, we investigate the underrepresentation of low-resource languages during multilingual tokenization, a crucial data preprocessing step in training multilingual NLG models, and we present best practices for training multilingual tokenizers. Overall, this thesis works towards enhancing the trustworthiness of NLG models in practice and facilitating support for a more diverse range of languages worldwide.
- Date of publication
- 2023
- Keyword
- DOI
- Resource type
- Rights statement
- In Copyright - Educational Use Permitted
- Advisor
- Bansal, Mohit
- Chaturvedi, Snigdha
- Celikyilmaz, Asli
- Dagan, Ido
- Oliva, Junier
- Bansal, Mohit
- Degree
- Doctor of Philosophy
- Degree granting institution
- University of North Carolina at Chapel Hill Graduate School
- Graduation year
- 2023
- Language
Relations
- Parents:
This work has no parents.
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Zhang_unc_0153D_21968.pdf | 2023-08-24 | Public | Download |