September 14th, 2 pm, the Thesis Defense: "Semantic Similarity of Concepts for a Human-Centered Idea Recommendation Feature in the Clustering Application Orchard" by Luka Stärk will take place at HCC Lab.

Abstract

The task of bringing many ideas into spacial-relation, called clustering, is part of a more extensive ideation process and is considered highly creative and beneficial to humans in developing ideas [2]. Hence, taking away this task with an automated cluster generation might harm the ideation process using the Orchard clustering application. This thesis implements a recommendation feature that supports users in clustering many ideas to synthesize and evolve them into more advanced and promising ideas. However, through interaction they can control which ideas are explored next when building their representation of ideas by clustering them. This is done by repeatedly specifying a target for which the recommendation feature provides similar ideas. The similarity measurements of recommendation feature applies the information of knowledge graphs (KGs) that describe resources and encode relations and facts of concepts. This work includes experiments on the performance of five variants of concept similarity methods and one idea similarity method using the Wikidata KG to ensure the accuracy of the used similarity measures. The similarity method path uses the path-length between concepts of the KG, the methods res, lin, jcn differently apply a graph-based Information Content measurement and wpath combines the two [3]. The evaluation points out the potential of Wikidata as a knowledge source for semantic similarity. The semantic similarity approach proved its suitability by a high correlation to human perceptions of word similarity for five well-studied datasets, namely R&G, M&C, WS353, WS353-Sim, and SimLex. wpath performed best and thus was applied in the recommendation feature. The idea similarity method is a variant of the Word Mover’s Distance [1] and proved its accuracy in further experiments. A Think-Aloud user study with five participants conducted for this thesis shows that the recommendation feature achieves a targeted and more efficient iteration through the ideas in Orchard. With the novel feature, the participants reported they were able to explore the ideas for clusters of their current interest. These observations could indicate an improvement in the idea synthesis, which could be looked at in further research.

References

[1] Matt J Kusner et al. “From Word Embeddings To Document Distances”. en. In: 2015, p. 10.

[2] Marc Tassoul and Jan Buijs. “Clustering: An Essential Step from Diverging to Converging”. en. In: Creativity and Innovation Management 16.1 (2007), pp. 16–26. issn: 1467-8691. doi: 10.1111/j.1467-8691.2007.00413.x. url: https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467- 8691.2007.00413.x (visited on 09/19/2019).

[3] G. Zhu and C. A. Iglesias. “Computing Semantic Similarity of Concepts in Knowledge Graphs”. In: IEEE Transactions on Knowledge and Data Engineering 29.1 (Jan. 2017), pp. 72–85. issn: 1041-4347. doi: 10.1109/ TKDE.2016.2610428.

The defence will be held in english.

First assesor: Prof. Dr. Sebastian Möller
Second assesor: Prof. Dr. Claudia Müller-Birn

If you want to join the defense, please contact Phil Wernberger.