Seminar in Computational Linguistics

Knowledge representation has been a cornerstone in traditional AI and there have been massive efforts in developing structured knowledge bases that encode various facts about the world. Knowledge bases are also widely used in many specialised domains (such as biology, medicine).

More recently, language models (LMs) trained on raw text have been proposed as an alternative to hand-built knowledge resources and such models have performed quite well on benchmarks for a number of tasks where knowledge-based systems have traditionally been used. In this talk, we will discuss a sample of recent work in this area: evaluations of LMs in knowledge-intensive tasks, limitations of LMs in this context, and finally some hybrid approaches that use a combined training objective based on raw text and structured knowledge.