Cross Language Word Sense Disambiguation (CL-WSD) is the problem of finding the correct target language translation of a word given the context in which it appears in the source language. In many cases a full disambiguation may not be necessary as it is common for different meanings of a word to have the same translation. This is es­pecially true in cases where the sense distinction is very fine and two or more senses of a word are closely related. For example, the two senses of the word letter, namely, "formal document' and "written/printed message " have the same French translation "lettre ". The problem is thus reduced to distinguishing between the coarser senses of a word and ignoring the finer sense distinctions which is known to be a common cause of errors in conventional WSD. CL-WSD can thus be seen as a slightly relaxed version of the conventional

WSD problem. However, CL-WSD has its own set of challenges as described below.

The translations learnt from a parallel corpus may contain a lot of errors. Such errors are hard to avoid due to the inherent noise associated with statistical alignment models. This problem can be overcome if good bilingual dictionaries are avail­able between the source and target language. Eu-roWordNet can be used to construct such a bilin­gual dictionary between English and French but it is not freely available. Instead, in this work, we use a noisy statistical dictionary learnt from the Europarl parallel corpus (Koehn, 2005) which is freely downloadable.

Another challenge arises in the form of match­ing the lexical choice of a native speaker. For ex­ample, the word coach (as in, vehicle) may get translated differently as autocar, autobus or bus even when it appears in very similar contexts. Such decisions depend on the native speaker's in­tuition and are very difficult for a machine to repli­cate due to their inconsistent usage in a parallel training corpus.

The above challenges are indeed hard to over­come, especially in an unsupervised setting, as ev­idenced by the lower accuracies reported by all systems participating in the SEMEVAL Shared Task on Cross-lingual Word Sense Disambigua­tion (Lefever and Hoste, 2010). Our system ranked second in the English French task (in the out-of-five evaluation). Even though its average performance was lower than the baseline by 3% it performed better than the baseline for 12 out of the 20 target nouns.

Our approach identifies the top-five translations of a word by taking a majority vote over the trans­lations appearing in the nearest neighbors of the test sentence as found in the training data. We use a pairwise similarity measure which finds the affinity between two sentences by calculating a weighted sum of the word overlap and the seman­tic overlap between them. The semantic overlap is calculated using standard Wordnet Similarity mea­sures.

The remainder of this paper is organized as fol­lows. In section 2 we describe related work on WSD. In section 3 we describe our approach. In Section 4 we present the results followed by con­clusion in section 5.

Knowledge based approaches to WSD such as Lesk's algorithm (Lesk, 1986), Walker's algorithm (Walker and Amsler, 1986), Conceptual Density (Agirre and Rigau, 1996) and Random Walk Algo­rithm (Mihalcea, 2005) are fundamentally overlap based algorithms which suffer from data sparsity. While these approaches do well in cases where there is a surface match (i.e., exact word match) between two occurrences of the target word (say, training and test sentence) they fail in cases where their is a semantic match between two occurrences of the target word even though there is no surface match between them. The main reason for this failure is that these approaches do not take into account semantic generalizations (e.g., train is-a vehicle).

On the other hand, WSD approaches which use Wordnet based semantic similarity measures (Pat-wardhan et al., 2003) account for such seman­tic generalizations and can be used in conjunc­tion with overlap based approaches. We there­fore propose a scoring function which combines the strength of overlap based approaches - fre­quently co-occurring words indeed provide strong clues - with semantic generalizations using Word-net based similarity measures. The disambigua­tion is then done using A;-NN (Ng and Lee, 1996) where the k nearest neighbors of the test sentence are identified using this scoring function. Once the nearest neighbors have been identified, the best translation is found by taking a majority vote over the translations of these nearest neighbors.

In this section we explain our approach for Cross Language Word Sense Disambiguation. The main emphasis is on disambiguation i.e. finding English sentences from the training data which are closely related to the test sentence.

We report results on the English-French Cross-Lingual Word Sense Disambiguation task. The test data contained 50 instances for 20 polysemous nouns, namely, coach, education, execution, fig­ure, job, letter, match, mission, mood, paper, post, pot, range, rest, ring, scene, side, soil, strain and test. We first extracted the sentences containing these words from the English-French portion of the Europarl corpus. These sentences served as the training data to be compared with each test sen­tence for identifying the nearest neighbors. The appropriate translations for the target word in the test sentence were then identified using the ap­proach outlined in section 3.2 and 3.3. For the best evaluation we submitted two runs: one con­taining only the top-1 translation and another con­taining top-2 translations. For the oof evaluation we submitted one run containing the top-5 trans­lations. The system was evaluated using Precision and Recall measures as described in the task pa­per (Lefever and Hoste, 2010). In the oof evalua­tion our system gave the second best performance among all the participants. However, the average precision was 3% lower than the baseline calcu­lated by simply identifying the five most frequent translations of a word according to GIZA++ word alignments. A detailed analysis showed that in the oof evaluation we did better than the baseline for 12 out of the 20 nouns and in the best evaluation we did better than the baseline for 5 out of the 20 nouns. Table 1 summarizes the performance of our system in the best evaluation and Table 2 gives the detailed performance of our system in the oof eval­uation. In both the evaluations our system pro­vided a translation for every word in the test data and hence the precision was same as recall in all cases. We refer to our system as OWNS (Overlap and WordNet Similarity).

Table 1 : Performance of our system in best evalu­ation

We described our system for English French Cross-Lingual Word Sense Disambiguation which calculates the affinity between two sentences by combining the weighted word overlap counts with semantic similarity measures. This similarity score is used to find the nearest neighbors of the test sentence from the training data. Once the nearest neighbors have been identified, the best translation is found by taking a majority vote over the translations of these nearest neighbors. Our system gave the second best performance in the oofevaluation among all the systems that partic­ipated in the English French Cross-Lingual Word Sense Disambiguation task. Even though the av­erage performance of our system was less than the baseline by around 3%, it outperformed the base­line system for 12 out of the 20 nouns.