PhD student - Department of Applied mathematics, Computer Science and Statistics

3/30622
label Burse autorenew 2025-09-29, 17:01 history_edu Maria Dumitru
A long standing goal of artificial intelligence is to design systems that can learn and acquire knowledge by interacting with their environment. A common and effective way for humans to acquire knowledge is through consulting readily available textual sources, such as books, newspapers, web pages and so on. Processing and learning from this type of data requires the abilities to automatically extract information from text, to learn rules that fuse together different pieces of information to make novel inferences and to reason about which extracted and inferred facts are true.

The offered PhD position is part of a larger project in cooperation with the KU Leuven, aimed at developing a machine reading system in which the algorithms for extracting, learning and inference are tightly integrated. In this project you will work together with other researchers. The focus of your research will be on the inference part, which is crucial to deal with inconsistencies that might arise when different sources (possibly with varying levels of trustworthiness) contradict each other, or due to errors in the interpretation of the sources (e.g., because they contain vague or ambiguous language).



One avenue to explore is the use of a stratified possibilistic knowledge base where formulas and facts are organized into equivalence classes and ranked from most to least certain. This may confer several advantages in the context of the project. For example, it can alleviate the problem of calibrating probability estimates from different algorithms used for extraction of facts, as we no longer would need to worry about minor discrepancies in probability estimates. Furthermore, it will allow to reduce many inference tasks to satisfiability problems. Therefore, it will be possible to leverage the recent advances in satifiability solvers and avoid some of the computation burdens of more traditional probabilistic inference.

All developed techniques and algorithms will be evaluated on information extracted from text found in biological and medical web pages.