Spatial vs. graphical representation of distributional semantic knowledge

Mao, Shufan

Permalink

https://hdl.handle.net/2142/110573 Copy

Description

Title

Spatial vs. graphical representation of distributional semantic knowledge

Author(s)

Mao, Shufan

Issue Date

2021-04-27

Director of Research (if dissertation) or Advisor (if thesis)

Willits, Jon Anthony

Department of Study

Psychology

Discipline

Psychology

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Date of Ingest

2021-09-17T01:11:18Z

Keyword(s)

• Semantic memory
• distributional models
• spreading activation

Abstract

Distributional semantic models represent words in a vector space and are competent in various semantic tasks. They are also limited in certain aspects such as representing different types of relations (e.g. syntagmatic vs. paradigmatic) in the same space and form indirect semantic relations, resulting in difficulty of advanced tasks such as inference meaning of unseen phrases and analogy. In this article, we propose a hybrid semantic model encoding distributional data (word co-occurrence) with graphical structure and measure lexical semantic relatedness by a spreading activation algorithm, which addresses the issue of spatial models. We systematically investigated the modeling parameters contributing to the representational capability, by manipulating and controlling the hyperparameters, and testing the models on a selectional preference task. The models are trained on an artificial corpus generated to describe ordered events happened in a toy world simulation, which embeds verb-noun selectional preference, and the task require the models to recover the verb-noun co-occurrence information and making inference on the selectional preference of verb-noun pairs absent in the corpus. We showed that both the graphical data structure with the spreading activation measure and the co-occurrence (information) encoding type attributing to the better performance and the capability to encode both syntagmatic and paradigmatic relations to form indirect semantic relationship and infer on unseen word pairs. As the hybrid graphical model is trained on corpus data, it is a semantic network from linguistic distributional statistics, and a new way of learning and representing semantic knowledge.

Graduation Semester

2021-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/110573

Copyright and License Information

Copyright 2021 Shufan Mao

Owning Collections