Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.contributor.advisor | Li, Wenjie Maggie (COMP) | en_US |
| dc.creator | Mu, Feiteng | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13577 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Causality-centric narratives reasoning | en_US |
| dcterms.abstract | Narratives are one of the foundational concepts of human society. It is an account of the development of human events, along with explanations of how and why these events happened. Narrative events serve as mirrors reflecting the intricate causality inherent in human activities, rendering them indispensable tools for comprehending the complexities of social dynamics. | en_US |
| dcterms.abstract | Recently, artificial intelligence (AI) has spurred a new era of advancement. However, despite the crucial role narratives plays, a critical bottleneck confronting AI systems lies in enabling machines to comprehend narrative events and leverage them for commonsense narrative reasoning. Specifically, we identify at least three key research problems that must be addressed in this domain of commonsense reasoning within narratives. Research Problem 1: How to automatically obtain diverse and high-quality commonsense event knowledge to solve the knowledge bottleneck problem in commonsense reasoning? Research Problem 2: How to effectively utilize narrative knowledge for commonsense reasoning to mitigate low-quality issues like dullness and repetition in AI-generated narrative texts, while ensuring the content aligns with human commonsense? More importantly, how to teach AI systems to grasp the causal relationships within narrative events, enabling them to effectively address high-level counterfactual questions? Research Problem 3: Given the fact that narrative coherence evaluation is a notoriously difficult thing in the generation community, how can we devise robust quantitative methods to evaluate the coherence of AI-generated narrative content, thereby furnishing valuable tools for the community? | en_US |
| dcterms.abstract | To solve these challenges, we focus on developing comprehensive narrative reasoning systems from the following three aspects: automatically causality mining, causal-knowledge enhanced narrative reasoning, and hard-negatives mining for narrative coherence learning. Overall, in this thesis, we organize our research works into the following three parts. | en_US |
| dcterms.abstract | In the first part (work 1 and work 2), we explore the research problem 1. Specifically, we explore the rule-based causality extraction method and possible de-biasing approach to harvest causal knowledge from text. In work 1, we manually create causal rules to extract cause-effect pairs from text. And we further construct the event causality network and demonstrate its use in the task of narrative effect generation. In addition, to mitigate the false-positive problem introduced by our rule-based system, in work 2, we explore possible de-biasing approach to obtain high-quality causal knowledge. We inaugurate counterfactual thinking for Event Causality Identification (ECI) to solve the context-keywords bias and event pairs bias problems in existing work. This allows us to obtain high-precision causal event pairs. | en_US |
| dcterms.abstract | In the second part (work 3 and work 4), we explore the solution for research problem 2 with the aim of developing a causal knowledge enhanced reasoning system with stronger causal perception capabilities. In work 3, we delve into causality centric narrative reasoning and push forward the existing knowledge-aware narrative reasoning to a new frontier. We thoroughly leverage multi-level causal knowledge for narrative reasoning, employing a two-stage framework designed to fully exploit the unique characteristics of knowledge across various granularities. Experimental results have shown that our work is effective and can improve the quality of generated narrative effect text. In work 4, we are trying to endow AI systems with more advanced counterfactual reasoning capabilities. One major challenge of counterfactual narrative reasoning is to maintain the causality between the counterfactual condition and the generated counterfactual outcome. Previous works simply utilize supervised datasets to train conditional generation models, but face the risk of exploiting artifacts of the dataset, instead of learning to robustly reason about counterfactuals. We propose a basic variational approach for counterfactual narrative reasoning. We further introduce a pre-trained classifier and external commonsense event causality to mitigate the model collapse problem in the variational approach, and hence improve the causality between the counterfactual condition and the generated counterfactual outcome. We assess the efficacy of our approach using real-world public benchmarks. Experimental results demonstrate its effectiveness. | en_US |
| dcterms.abstract | In the third part (work 5), we target research problem 3 and propose novel hard-negatives mining strategies for self-supervised narrative coherence learning. Existing works mainly follow the contrastive learning paradigm. However, the negative samples in their methods can be easily distinguished, which makes their methods unsatisfactory. We devise two strategies for mining hard negatives, including (1) crisscrossing a narrative and its contrastive variants; and (2) event-level replacement. To obtain contrastive variants, we utilize the Brownian Bridge process to guarantee the quality of generated contrastive narratives. We assess our model across multiple tasks, confirming its effectiveness and demonstrating its applicability to various use cases. | en_US |
| dcterms.abstract | To sum up, we conduct a comprehensive study on narrative reasoning. Through the use of our proposed methods to real-world datasets, we have illustrated the significant improvements that can be achieved in existing narrative reasoning models. We believe that our works will have a profound impact on the field of narrative reasoning. Although this thesis presents novel methods for this topic, it still has many open problems. We list some future research directions at the end of this thesis. | en_US |
| dcterms.extent | xxi, 184 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2025 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Artificial intelligence | en_US |
| dcterms.LCSH | Narration (Rhetoric) | en_US |
| dcterms.LCSH | Reasoning | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13577