Work

Graph-based Knowledge Representation for Natural Language Understanding Relatrice:

Year: 2022
Type: book
Abstract: In recent years the use of neural-based models trained on just textual data for specific tasks has become the go-to option in Natural Language Processing, both in academia and the industry.While these models show good results in specific frameworks, they present several issues, particularly in relation to data and model maintenance.These issues are made all the more difficult to solve given the inherent black-box aspect of neural models, which makes them particularly difficult to improve and update.With regards to natural language understanding, an alternative to the textualdistributional representation implemented by neural models is represented by the use of structural, symbolic forms of representation for natural language utterances.These models are easier to implement and generally less data-hungry than neural frameworks, and can be implemented on top of neural models in order to solve some of their issues.This work shows the implementation of such symbolic representation, in particular by implementing graph-based knowledge representation in three main tasks: semantic enrichment, relation extraction and question answering.In particular, such representations are shown to be useful in connecting different sources of knowledge, namely texts and knowledge graphs available on the Web.By doing so, inherent contextual information extracted from texts can be integrated with common-sense and domain-specific knowledge in an explicit way.This is also beneficial since it creates an interpretable area that can act as a buffer against the black-box nature of neural models.
Language: English
Open Access status: green
OpenAlex ID: W7217341051
FWCI:
Cites: 376
Cited by:
Related to: 10