2022 COLING Outstanding Paper Award

Posted: 

CSE is very delighted to share that Prof. Yu Su’s paper, titled “ArcaneQA: Dynamic Program Induction and Contextualized Encoding for Knowledge Base Question Answering” has won the 2022 COLING Outstanding Paper Award (top 15 out of 2253 submissions.). This paper was first-authored by Yu Gu (currently a 4th-year Ph.D. student advised by Prof. Su).

Here is a brief introduction of the ArcaneQA paper:

ArcaneQA uses a novel unified framework to address two fundamental challenges in knowledge base question answering (KBQA), namely, large search space and schema linking. Modern knowledge bases (KBs) like Freebase or Google Knowledge Graph contain millions of entities and billions of facts about these entities (as edges in the graph). Finding answers to a natural language question from such massive KBs can induce an extremely large search space. In addition, the broad semantic space defined by the schema of the KB leads to pervasive ambiguities. For example, for the question “Flaminia Brasini was the designer of what game?” does it refer to the designer of games (denoted as games.game.designer) or golf courses (denoted as sports.golf_course.designer). While it is easy for humans to disambiguate between the two because we understand the semantics of the entire question, it is non-trivial for AI models, especially when there are many more of such similar schema items competing with each other. Existing methods tackle KBQA with a pipelined design, which are limited in both scalability and accuracy (due to error propagation). In this paper, Yu Gu and Prof. Su propose a unified model that addresses both the aforementioned challenges with two mutually-boosting techniques, i.e., a dynamic program induction module and a dynamic contextualized encoding module. The former can effectively prune the search space on the fly, while the latter further intelligently guides the search process for the right answer by jointly encoding the question and the KB schema using pre-trained language models like BERT. ArcaneQA achieves a high accuracy on multiple standard KBQA benchmarks and is faster than existing methods by an order of magnitude. It can be used to answer questions for domains it has never seen before, i.e., zero-shot generalization, which is a key capability for such systems to be used in practice.

COLING, the International Conference on Computational Linguistics, is one of the premier conferences for the natural language processing and computational linguistics. COLING 2022 provides a leading forum for disseminating the latest research in computational linguistics and natural language processing. The Outstanding Paper Award of COLING 2022 is highly selective. Only 15 papers out of 2253 submissions have won this award.

Yu Gu and Prof. Su are part of the OSU NLP group.

 

Faculty Profile

Related News

Ohio State leads new global climate center on AI for biodiversity change, funded by NSF
TALK: Towards a Distributed Learning Model for Teaching at Scale
5G-OH, CSE, TDAI Talk Featuring Dr. David Joyner: Towards a Distributed Learning Model for Teaching at Scale