Leveraging Eventuality-Centric Implicit Knowledge for Natural Language Understanding

Data-driven techniques have significantly advanced the state-of-the-art in a wide range of natural language processing (NLP) tasks. However, their primary strength lies in pattern discovery within observed signals. As a result, these techniques often struggle to capture implicit knowledge–information that is implicitly assumed in human communication and not explicitly present in textual data. Human communication is full of semantic gaps due to reliance on such implicit knowledge. Although the importance of implicit knowledge has been widely recognized in the research community, it still remains challenging to represent implicit knowledge for target applications, align it with language expressions, and incorporate it into NLP systems for high-level semantic analysis. 

To achieve accurate and robust language understanding, this thesis proposes a series of techniques to identify and operationalize implicit knowledge, demonstrating that eventuality-centric knowledge is essential for understanding language phenomena across various real-world NLP application scenarios. The first part of the thesis focuses on the use of psychological and social implicit knowledge, showing that human basic motives underlie the sentiment expressed in customer reviews and that knowledge about human activities enhances computational representations of user-generated To-Do task descriptions. The thesis then argues that open-domain conversation systems necessitate an understanding of implicit situational context, encompassing psychological, social, and physical aspects. Empirical analyses reveal that implicit information can benefit NLP systems, but at the same time introduces the challenge of filtering out irrelevant information, for which potential solutions are explored. Finally, the thesis discusses implicit knowledge of the physical world, proposing frame-based representations of cooking recipes to capture unspoken knowledge about physical states, actions, and effects. Through these studies, this thesis introduces novel methods for defining tasks to achieve end goals, representing and acquiring implicit knowledge, and incorporating it into computational models as input signals and predictive targets.