Generating Semantic Graphs for Natural Language- [electronic resource]
Generating Semantic Graphs for Natural Language- [electronic resource]
- 자료유형
- 학위논문파일 국외
- 최종처리일시
- 20240214100457
- ISBN
- 9798379604158
- DDC
- 004
- 저자명
- Zhou, Jiawei.
- 서명/저자
- Generating Semantic Graphs for Natural Language - [electronic resource]
- 발행사항
- [S.l.]: : Harvard University., 2023
- 발행사항
- Ann Arbor : : ProQuest Dissertations & Theses,, 2023
- 형태사항
- 1 online resource(189 p.)
- 주기사항
- Source: Dissertations Abstracts International, Volume: 84-12, Section: B.
- 주기사항
- Advisor: Rush, Alexander M. ;Yu, Minlan.
- 학위논문주기
- Thesis (Ph.D.)--Harvard University, 2023.
- 사용제한주기
- This item must not be sold to any third party vendors.
- 초록/해제
- 요약Natural language understanding is a critical capability in achieving advanced artificial intelligent language processing systems such as reading comprehension, question answering, and interactive dialogues. Despite the remarkable progress made by modern deep learning techniques for natural language processing (NLP) in the past decade, machines still lag behind human capacity in deep language understanding. This requires machines to represent and comprehend the underlying meaning, or semantics, from the surface form of language despite its variations and intricacies. Explicit semantic representations of language provide a systematic way of building interpretable and controllable agents with language understanding ability, especially with versatile graph structures. Abstracting away from the surface form of language, the semantic graphs can capture complex semantic phenomena, and provide a structured and consistent way of presenting the underlying meaning of language, which can be utilized for applications that require accurate semantic interpretation. Depending on applications, certain semantic graphs such as functional programs can also be executable for direct machine processing, paving ways for interactive and efficient human-machine communication. However, the complexity of the structured graphs and the expensiveness of expert data annotation pose unique challenges in automating the generation process of these graphs.In this thesis, we develop techniques using machine learning models to generate such semantic graphs for natural language, as well as exploring efficient utilization of these graphs in real applications such as dialogue systems. We first formulate a general framework for text-to-graph generation with an autoregressive process through a carefully designed sequence of actions, and then devise a principled approach that combines the general graph construction process and neural models such as sequence-to-sequence Transformer models and pointer networks with synergy. We apply the proposed method to interpret natural language sentences into abstract semantic graphs, where the end-to-end deep learning model is guided by carefully designed logic-based state machines that manage the graph and action transduction. The hybrid approach injects an effective form of structured inductive bias in the model computation, resulting in high-quality graph generation without complex modeling pipelines. With recent advances of pre-trained language models benefiting from large amounts of unlabeled data, we further study the effective way of merging the benefits of these unstructured models with structured generation of semantic graphs to increase data efficiency. Furthermore, we extend our text-to-graph generation framework for executable semantic graphs that are programs serving as essential building blocks of a reliable task-oriented dialogue system. We propose a novel online semantic parsing paradigm which aims to generate and execute partial semantic graphs simultaneously as the sentence is being revealed. The application enables real-time interpretation of human utterances to accelerate machine response, making human-machine interaction experience more natural. We hope our research on text-to-graph generation and application not only sheds some light on natural language understanding and reliable semantic-aware system building, but also creates further opportunities for interdisciplinary research beyond NLP where symbolic graph-structured data modeling and generation are of vital importance.
- 일반주제명
- Computer science.
- 일반주제명
- Linguistics.
- 일반주제명
- Information science.
- 키워드
- Dialogue system
- 키워드
- Graph generation
- 키워드
- Semantic graphs
- 기타저자
- Harvard University Engineering and Applied Sciences - Engineering Sciences
- 기본자료저록
- Dissertations Abstracts International. 84-12B.
- 기본자료저록
- Dissertation Abstract International
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.