Sorokin, Daniil (2021)
Knowledge Graphs and Graph Neural Networks for Semantic Parsing.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00019187
Ph.D. Thesis, Primary publication, Publisher's Version
|
Text
Daniil-Sorokin-Knowledge-Graphs-and-Graph-Neural-Networks-for-Semantic-Parsing.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (2MB) | Preview |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Knowledge Graphs and Graph Neural Networks for Semantic Parsing | ||||
Language: | English | ||||
Referees: | Gurevych, Prof. Dr. Iryna ; Kersting, Prof. Dr. Kristian ; Roth, Prof. Dan | ||||
Date: | 2021 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xii, 212 Seiten | ||||
Date of oral examination: | 10 June 2021 | ||||
DOI: | 10.26083/tuprints-00019187 | ||||
Abstract: | Human communication is inevitably grounded in the real world. Existing work on natural language processing uses structured knowledge bases to ground language expressions. The process of linking entities and relations in a text to world knowledge and composing them into a single coherent structure constitutes a semantic parsing problem. The output of a semantic parser is a grounded semantic representation of a text, which can be universally used for downstream applications, such as fact checking or question answering. This dissertation is concerned with improving the accuracy of grounding methods and with incorporating the grounding of individual elements and the construction of the full structured representation into one unified method. We present three main contributions: - we develop new methods to link texts to a knowledge base that integrate context information; - we introduce Graph Neural Networks for encoding structured semantic representations; - we explore generalization potential of the developed knowledge-based methods and apply them on natural language understanding tasks. For our first contribution, we investigate two tasks that focus on linking elements of a text to external knowledge: relation extraction and entity linking. Relation extraction identifies relations in a text and classifies them into one of the types in a knowledge base schema. Traditionally, relations in a sentence are processed one-by-one. Instead, we propose an approach that considers multiple relations simultaneously and improves upon the previous work. The goal of entity linking is to find and disambiguate entity mentions in a text. A knowledge base contains millions of world entities, which span different categories from common concepts to famous people and place names. We present a new architecture for entity linking that is effective across diverse entity categories. Our second contribution is centered on a grounded semantic parser. Previous semantic parsing methods grounded individual elements in isolation and composed them later into a complete semantic representation. Such approaches struggle with semantic representations that include multiple grounded elements, world entities and semantic relations. We integrate the grounding step and the construction of a full semantic representation into a single architecture. To encode semantic representations, we adapt Gated Graph Neural Networks for this task for the first time. Our semantic parsing methods are less prone to error propagation and are more robust for constructing semantic representations with multiple relations. We prove the efficiency of our grounded semantic parser empirically on the challenging open-domain question answering task. In our third contribution, we cover the extrinsic evaluation of the developed methods on three applications: argumentative reasoning, fact verification and text comprehension. We show that our methods can be successfully transferred to other tasks and datasets that they were not trained on. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-191875 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science > Ubiquitous Knowledge Processing | ||||
Date Deposited: | 02 Dec 2021 13:07 | ||||
Last Modified: | 05 Sep 2022 08:27 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/19187 | ||||
PPN: | 489267777 | ||||
Export: |
View Item |