News & Events

27.06.2014

A data-driven approach to commonsense reasoning based on qualitative spatial representations from the web


Steven Schockaert, Cardiff University

 

Conceptual spaces (sometimes also called semantic spaces) are high-dimensional geometric spaces in which entities are modelled as points and categories as convex regions. Given a relevant text corpus, conceptual spaces can be induced from data by using dimensionality reduction methods such as multi-dimensional scaling. To date, conceptual spaces are mainly used for measuring similarity. In this talk, I will argue that many aspects of semantic relatedness can be identified with qualitative spatial relations in conceptual spaces, and that this allows us to learn symbolic knowledge about categories in a purely data-driven way.

 

As a simple example, the spatial part-of relation corresponds to the conceptual is-a relation, which is useful for automatically refining taxonomies. Second, spatial betweenness in conceptual spaces is useful for identifying “intermediary concepts”. For example, we can think of a tapas bar as being intermediate between a pub and a restaurant, and accordingly we can expect the representation of “tapas bar” to be geometrically between the representations of “restaurant” and “pub” in a conceptual space of places. Finally, relative attributes such as “more violent than” (for films) or “more tannic than” (for wines) can be associated with direction relations in a conceptual space.

 

The aforementioned spatial relations are useful for implementing various patterns of commonsense reasoning. Betweenness, for example, can be used to implement a symbolic counterpart to numerical interpolation, e.g. from the knowledge that bars and restaurants sell drinks, we derive that tapas-bars are also likely to sell drinks. Direction relations are useful for implementing a fortiori inference (and other forms of analogical reasoning). For example, given that Die Hard has received an 18 certificate from the British Board of Film Classification and that Drive is more violent than Die Hard, we can plausibly derive that Drive has also received an 18 certificate.

 

I will report the result of experiments that show the usefulness of the aforementioned patterns of commonsense reasoning. For example, we have found that a betweenness based classifier, using a conceptual space of place types induced from Flickr, outperforms humans in categorising places from the Foursquare and Geonames taxonomies. Similarly, by identifying qualitative direction relations in a conceptual space of films (derived from a large corpus of film reviews), we can implement a form of analogical reasoning which outperforms standard classifiers such as SVMs, kNN and C4.5 in a variety of film related categorisation tasks.

 

 

Bio:

Steven Schockaert studied computer science at Ghent University, where he also defended his PhD thesis in 2008 on "Reasoning about fuzzy temporal and spatial information from the web". This thesis was awarded with the 2008 ECCAI Artificial Intelligence Dissertation Award and the IBM Belgium Prize for Computer Science. After his PhD, he obtained a postdoctoral fellowship from the Research Foundation - Flanders. In September 2011 he became a lecturer at Cardiff University, where he currently works.

His current research interests are largely centred around three areas: multi-valued logics, commonsense reasoning, and geographic information retrieval. In the area of multi-valued logics, he is mainly working on Łukasiewicz logic (e.g. automated reasoning and complexity) and on extensions of answer set programming, a form of logic programming based on the notion of stable models. In the area of commonsense reasoning, he looks at qualitative techniques for managing imperfect knowledge bases, including possibilistic logic and qualitative approaches to similarity-based and analogical reasoning. Finally, in the area of geographic information retrieval, his main focus is on acquiring geographic information from semi-structured web resources

Date: 27.06.2014

Time: 15:30 h

Location: Cartesium, Bremen

Speaker-URL: users.cs.cf.ac.uk/S.Schockaert/Home.html

 


Files:
Steven_Schockaert.pdf118 K