SemantiKLUE: Robust Semantic Similarity at Multiple Levels Using Maximum Weight Matching

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 532–540, Dublin, Ireland, August 23-24, 2014. SemantiKLUE: Robust Semantic Similarity at Multiple Levels Using Maximum Weight Matching Thomas Proisl and Stefan Evert and Paul Greiner and Besim Kabashi Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) Department Germanistik und Komparatistik Professur für Korpuslinguistik Bismarckstr. 6, 91054 Erlangen, Germany {thomas.proisl,stefan.evert,paul.greiner,besim.kabashi}@fau.de Abstract Being able to quantify the semantic similar- ity between two texts is important for many practical applications. SemantiKLUE com- bines unsupervised and supervised tech- niques into a robust system for measuring semantic similarity. At the core of the sys- tem is a word-to-word alignment of two texts using a maximum weight matching algorithm. The system participated in three SemEval-2014 shared tasks and the com- petitive results are evidence for its usability in that broad field of application. 1 Introduction Semantic similarity measures the semantic equiv- alence between two texts ranging from total dif- ference to complete semantic equivalence and is usually encoded as a number in a closed interval, e. g. [0,5]. Here is an example for interpreting the numeric similarity scores taken from Agirre et al. (2013, 33): 0. The two sentences are on different topics. 1. The two sentences are not equivalent, but are on the same topic. 2. The two sentences are not equivalent, but share some details. 3. The two sentences are roughly equivalent, but some important information differs/missing. 4. The two sentences are mostly equivalent, but some unimportant details differ. 5. The two sentences are completely equivalent, as they mean the same thing. Systems capable of reliably predicting the semantic similarity between two texts can be beneficial for a This work is licensed under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ broad range of NLP applications, e. g. paraphrasing, MT evaluation, information extraction, question answering and summarization. A general system for semantic similarity aiming at being applicable in such a broad scope has to be able to adapt to the use case at hand, because different use cases might, for example, require dif- ferent similarity scales: For one application, two texts dealing roughly with the same topic should get a high similarity score, whereas for another ap- plication being able to distinguish between subtle differences in meaning might be important. The three SemEval-2014 shared tasks focussing on se- mantic similarity (cf. Sections 3, 4 and 5 for more detailed task descriptions) provide a rich testbed for such a general system, as the individual tasks and subtasks have slightly different objectives. In the remainder of this paper, we describe SemantiKLUE, a general system for measuring se- mantic similarity between texts that we built based on our experience from participating in the *SEM 2013 shared task on “Semantic Textual Similarity” (Greiner et al., 2013). 2 System Description SemantiKLUE operates in two stages. In the first, unsupervised stage, a number of similarity mea- sures are computed. Those measures are the same for all tasks and range from …

Content truncated to 500 words