| Category: ||E-Announce |
| Subject: ||Semantic Similarity Experiment |
| From: || |
| Email: ||nseco_(on)_dei.uc.pt |
| Date received: ||17 Jul 2007 |
[Apologies for cross-postings]
[Please distribute to potentially interested parties]
In the context of joint research project we are asking fellow researchers to
contribute about 10 min of their time and collaborate in experiment that (we
hope) will help us gather a large dataset of similarity ratings for pairs of
words. Participation is quite simple, so if you are interested please read
the section HOW TO PARTICIPATE. If you want to learn more about the
experiment please read the section INTRODUCTION.
Thanks in advance,
Giuseppe Pirrò & Nuno Seco
Semantic similarity plays an important role in Information Retrieval,
Natural Language Processing, Ontology Mapping and other related fields of
In particular researchers have developed a variety of semantic similarity
and relatedness measures by exploiting information found in lexical
resources such as WordNet. Current similarity metrics based on WordNet can
be classified in one of the following categories:
Edge-Counting measures that are based on the number of links relating
two concepts that are being compared.
Information Content measures that are based on the idea that the
similarity of two concepts is related to the amount of information they have
Feature-Based measures that exploit the features (e.g., descriptions in
natural language) of a term while usually ignoring their location in the
Hybrid measures that combine ideas from previous categories.
In order to evaluate the suitability of the various similarity measures they
are usually compared against human judgements by calculating correlation
values. A typical reference, in terms of evaluation, are the results of the
Rubenstein and Goodenough (R&G) experiment. R&G in 1965 obtained "synonymy
judgments" of 51 human subjects on 65 pairs of words. The pairs ranged from
"highly synonymous" (gem-jewel) to "semantically unrelated" (noon-string).
Subjects were asked to rate them on the scale of 0.0 to 4.0 according to
their "similarity of meaning" and ignoring any other observed semantic
Even if from the R&G experiment, other similar experiments have been carried
out, we are not aware of similarity experiments aimed at showing how robust
the different measures are when compared against different versions of
WordNet. With this objective in mind we want to collect human similarity
estimations on the whole Rubsteing and Goodenough dataset and
subsequentially compare outputs of existing similarity measures. We chose to
adopt the R&G dataset since others have worked on it, thus permitting direct
comparison of results obtained by different experiments.
Moreover, we want to show the suitability of an Information Content metric
that solely relies on the WordNet taxonomy, without relying on external
collection of texts.
How to participate:
In order to participate in the similarity experiment point your browser to:
Then by clicking on the register link you can register and immediately
receive a password via email.
After logging in you should indicate similarity values for all the word
pairs by using the Slider provided for each pair. The estimated time
required is about 10 minutes including time for registering.
Results of the experiment and the data will be published as soon as we
collect a significant amount of ratings.
Elsnet-list mailing list