Daniil Sorokin from UKP Lab, TU Darmstadt will be giving a talk this coming Wednesday at the Institute for Computational Linguistics on graph NNs and Knowledge-Informed NLU.
Time / location: Wednesday June 26, 16:00 / Andreasstrasse 15, room AND 2.23.
Graph Neural Networks for Knowledge-Informed Natural Language Understanding
To answer a natural language question, one may search for information in a structured knowledge base or in a set of unstructured documents. In this talk, I present a semantic parsing approach to question answering that relies on external world knowledge. The question meaning is modeled precisely through the entities and relations available in the knowledge base.
I address the problem of processing semantic structures that consist of multiple entities and relations. Previous work on knowledge base question answering has largely focused on selecting the correct semantic relations for the question and disregarded the connections between entities and the directions of the relations. I propose to use gated graph neural networks to encode the semantic structures.
First, I demonstrate on an open-domain knowledge base question answering data set that the gated graph neural networks lead to an improved performance against the baseline models that do not explicitly model the semantic structure. Second, I show that a knowledge based interpretation of the question can be used to improve results of a state-of-the-art document question answering system.
Daniil Sorokin is a final year PhD student in Natural Language Processing at the UKP Lab, Technische Universität Darmstadt. He works on neural network approaches to link texts to structured knowledge bases and has recently published on the topics of relation extraction, semantic parsing for question answering, graph neural networks and fact verification.
Before the PhD studies, Daniil had completed a master degree in Computational Linguistics at the University of Tübingen and had worked as a developer at a machine translation company.