Relational Learning over Temporal Knowledge Graphs under Realistic Settings
Zifeng Ding (Ph.D. Student)
There has been an increasing interest in performing prediction and learning tasks on temporal knowledge graphs (TKGs). TKGs are widely used resources for studying multi-relational data in the form of a directed graph, where each labeled edge describes a factual statement, such as (Olaf Scholz, is chancellor of, Germany, 2023). Since TKGs are known to suffer from incompleteness, recently, various methods have been developed for automatically completing TKGs. Though these methods achieve superior performance on temporal knowledge graph completion (TKGC), they have their limitations. (1) In real-world scenarios, TKGs evolve over time, indicating that new (unseen) entities and relations may emerge constantly. Besides, real-world TKGs exhibit long-tail distributions, where a large portion of entities and relations only have few associated edges. Traditional TKGC methods learn the representations of a fixed set of observed (seen) entities and relations, and perform link prediction over them. To learn the optimal representations, these methods require a large number of training examples associated with each of them, making them unable to deal with newly-emerged, yet unseen entities and relations. In this PhD project, we aim to focus on using few-shot learning to solve the data scarcity problem brought by this realistic setting. (2) Some recent breakthroughs in artificial intelligence (AI) target on developing AI agents that are able to communicate with humans in natural language, e.g., a number of large language models such as GPT4. TKGs serve as strong knowledge sources that can be used to aid human decisions by leveraging relevant knowledge. This calls for the need to integrate TKG reasoning into a series of tasks of natural language processing (NLP), e.g., question answering (QA). Traditional TKGC methods only focus on the graphical reasoning over TKGs and cannot be generalized to NLP tasks, making them unable to be directly implemented on the new wave of AI agents. In this PhD project, we aim to propose a new task, i.e., forecasting TKGQA, to bridge the gap among TKG reasoning, future forecasting, and natural language QA. We also aim to develop the first forecasting TKGQA method to solve this novel task. The motivation behind it is that we want to draw attention to developing new AI agents that help humans in decision making and future planning by first understanding natural language and then using the abundant knowledge stored in TKGs for answer inference.
|Volker Tresp (LMU Munich & Siemens)
|Michael Bronstein (Imperial College London)
|01 June 2021 - 31 May 2024
|01 December 2023 - 31 May 2024 - Ongoing