Evaluation of Distributional Semantic Models for the Extraction of Semantic Relations for Named Rivers from a Small Specialized Corpus

EcoLexicon (http://ecolexicon.ugr.es) is a terminological knowledge base on environmental science, whose design permits the geographic contextualization of data. For the geographic contextualization of landform concepts such as named rivers (e.g., Nile River), distributional semantic models (DSMs) w...

Full description

Saved in:
Bibliographic Details
Published in:Procesamiento del Lenguaje Natural 2019-09, Vol.63, p.51
Main Authors: Rojas-Garcia, Juan, Faber, Pamela
Format: Article
Language:English
Subjects:
Corpus analysis
Corpus linguistics
Datasets
Environmental science
Evaluation
Knowledge bases (artificial intelligence)
Landforms
Literature reviews
Parameter identification
Rivers
Semantic relations
Semantics
Toponymy
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:EcoLexicon (http://ecolexicon.ugr.es) is a terminological knowledge base on environmental science, whose design permits the geographic contextualization of data. For the geographic contextualization of landform concepts such as named rivers (e.g., Nile River), distributional semantic models (DSMs) were applied to a small-sized, specialized corpus to extract the terms related to each named river mentioned in it and their semantic relations. Since the construction of DSMs is highly parameterized and their evaluation in small specialized corpora has received little attention, this paper identified parameter combinations in DSMs suitable for the extraction of the semantic relations takes_place_in, affects, and located_at, frequently held by named rivers in the corpus. The models were thus evaluated using three gold standard datasets. The results showed that, for a small-sized corpus, count-based models outperformed prediction-based ones with the log-likelihood association measure, and the detection of a specific relation depended largely on the context window size.
ISSN:1135-5948
1989-7553
DOI:10.26342/2019-63-5