Latent relational search is a novel search paradigm based on the analogy between two entity pairs.
For example, given the query {(Google, Mountain View), (Microsoft, ?)}, a latent relational search engine is expected to retrieve and rank the entity "Redmond" at the top of the result list
because the relation between Google and Mountain View is highly similar to that between Microsoft and Redmond.
The idea of Latent Relational Search has been discussed in
(T. Veale 2003), (Turney 2006)
and clearly stated in (Bollegala et al. 2009).
Many latent relational search systems have been developed (WWW2REL,
Kato et al. CIKM 2009, Goto 2010).
In Milresh, we extend the latent relational search paradigm to be able to process cross-lingual queries (i.e., the input pair and the target pair are in different languages).
Moreover, we propose a method for building an index to perform latent relational search in high speed. Our system is the first system that allows latent relational search in
a normal user's search session (i.e., less than 10 seconds).
- Monolingual latent relational search demo video:
English,
Japanese
- Sample monolingual queries and results:
sample-mono-en.csv
(CSV file: right click, "Save As...", then open with Excel)
- Live demo of monolingual latent relational search.
- Cross-language latent relational search demo video:
clrsdemo.wmv (WMV file, 4.1MB)
- Sample cross-language queries:
sample-clrs.csv
(warning: this file contains Japanese characters)
- Live demo of Japanese - English cross-language latent relational search on a very small corpus, you need a Japanese Input Method Editor to input Japanese.
Back to Duc's home page