Application Showcase

Our application vision of sentiment analysis from semantic parses provides a longer-term goal. For this project, we propose anapplication that can showcase the utility of Web-scale structural semantics directly at the level of MRS representations: content retrieval taking advantage of semantic relations.

Imagine a user searching for smartphones that support the Scalable Vector Graphics (SVG) Web extension. In part owed to ambiguity of the terms support and SVG, a keyword-based Internet search yields unsatisfactory results. There are relevant entries from Wikipedia, technology blogs, and review sites, of course—but few make it into the top few pages of standard search engine results. In contrast, posing a structured query against semantically parsed Web content would retrieve fewer but more relevant matches, i.e. higher search precision at (potentially) reduced recall. In MRS, a relational query could essentially take the following form: { _smartphone_ n(x1), _named_(x2, SVG), _support_v( _ , x1, x2) }.

Schäfer, Uszkoreit, Federmann, Marek, & Zhang (2008) present a search infrastructure based on MRS relations developed at DFKI, and this is the technology we plan to adapt for this track. DFKI is currently engaged in a closely related project (also using the ERG, PET, and MRS) and have agreed to the exchange of experience and software, and to jointly developing a comprehensive relational search showcase for ICT-related information. Reflecting the scale of our seach on resource creation and semantic parsing, the demonstrator will index many millions of IT-related documents (though not billions, as general Web search engines do), with a focus on English and enabling information access across UGC types and services. To make the showcase more attractive, recall can be increased through integration of query expansion (cell phone ⊂ smartphone or support ⊂ implement, for example) and coreference resolution of pronouns, both R&D activities already underway at UiO and DFKI.

By Stephan Oepen, Lilja Øvrelid
Published Sep. 27, 2012 1:20 PM - Last modified Sep. 27, 2012 1:23 PM