Lhis paper describes a graph-based system for Cross-Lingual Word Sense Disambiguation, i.e. the task of disambiguating a word in context by providing its most appropriate translations in dif­ferent languages (Lefever and Hoste, 2010, CL-WSD henceforth). Our goal at SemEval-2010 was to assess whether graph-based approaches, which have been successfully developed for monolingual Word Sense Disambiguation, represent a valid framework for CL-WSD. Lhese typically trans­form a knowledge resource such as WordNet (Fell-baum, 1998) into a graph and apply graph algo­rithms to perform WSD. In our work, we follow this line of research and apply graph-based meth­ods to multilingual co-occurrence graphs which are automatically created from parallel corpora.

Our method is heavily inspired by previous pro­posals from Véronis (2004, Hyperlex) and Agirre et al. (2006). Hyperlex performs graph-based

WSD based on co-occurrence graphs: given a monolingual corpus, for each target word a graph is built where nodes represent content words co-occurring with the target word in context, and edges connect the words which co-occur in these contexts. Lhe second step iteratively selects the node with highest degree in the graph (root hub) and removes it along with its adjacent nodes. Each such selection corresponds to isolating a high-density component of the graph, in order to select a sense of the target word. In the last step the root hubs are linked to the target word and the Mini­mum Spanning Lree (MSL) of the graph is com­puted to disambiguate the target word in context. Agirre et al. (2006) compare Hyperlex with an al­ternative method to detect the root hubs based on PageRank (Brin and Page, 1998). PageRank has the advantage of requiring less parameters than Hyperlex, whereas the authors ascertain equal per­formance of the two methods.

We start by building for each target word a mul­tilingual co-occurrence graph based on the target word's aligned contexts found in parallel corpora (Sections 3.1 and 3.2). Multilingual nodes are linked by translation edges, labeled with the target word's translations observed in the corresponding contexts. We then use an adapted PageRank al­gorithm to select the nodes which represent the target word's different senses (Section 3.3) and, given these nodes, we compute the MSL, which is used to select the most relevant words in con­text to disambiguate a given test instance (Section 3.4). Lranslations are finally given by the incom­ing translation edges of the selected context words.

Experimental Setting. We submitted two runs for the task (UHD-1 and UHD-2 henceforth). Since we were interested in assessing the impact of using different resources with our methodology, we automatically built multilingual graphs from different sentence-aligned corpora, i.e. Europarl (Koehn, 2005) for UHD-1, augmented with the JRC-Acquis corpus (Steinberger et al., 2006) for UHD-2. Both corpora were tagged and lemma-tized with LreeLagger (Schmid, 1994) and word aligned using GIZA++ (Och and Ney, 2003). For German, in order to avoid the sparseness deriving from the high productivity of compounds, we per­formed a morphological analysis using Morphisto (Zielinski et al., 2009).

To build the multilingual graph (Section 3.2), we used a minimum frequency threshold of 2 oc­currences for a word to be inserted as a node, and retained only those edges with a weight less or equal to 0.7. After constructing the multilin­gual graph, we additionally removed those trans­lations with a frequency count lower than 10 (7 in the case of German, due to the large amount of compounds). Finally, the translations gener­ated for the Best evaluation setting were ob­tained by applying the following rule onto the ranked answer translations: add translation tri while count(tri) > count(tri-\)/?>, where i is the i-th ranked translation.

Results and discussion. The results for the Best and out-of-five (OOF) evaluations are pre­sented in Tables 1 and 2 respectively. Results are computed using the official scorer (Lefever and Hoste, 2010) and no post-processing is applied to the system's output, i.e. we do not back-off to the baseline most frequent translation in case the sys­tem fails to provide an answer for a test instance. For the sake of brevity, we present the results for UHD-1, since we found no statistically significant difference in the performance of the two systems (e.g. UHD-2 outperforms UHD-1 only by +0.7% on the best evaluation for French).

We presented in this paper a graph-based system to perform CL-WSD. Key to our approach is the use of a co-occurrence graph built from multilin­gual parallel corpora, and the application of well-studied graph algorithms for monolingual WSD (Véronis, 2004; Agirre et al., 2006). Future work will concentrate on extensions of the algorithms, e.g. computing hubs in each language indepen­dently and combining them as a joint problem, as well as developing robust techniques for unsuper­vised tuning of the graph weights, given the obser­vation that the most frequent translations tend to receive too much weight and accordingly crowd out more appropriate translations. Finally, we plan to investigate the application of our approach directly to multilingual lexical resources such as PanDictionary (Mausam et al., 2009) and Babel-Net (Navigli and Ponzetto, 2010).