Ambiguity is the task of building up multiple alternative linguistic structures for a single input (Kozareva et al, 2007). Word Sense Disambiguation (WSD) is a key enabling-technology that automatically chooses the intended sense of a word in context. In this task, one of the most used lexical data base is WordNet (WN) (Fellbaum, 1998). WN is an online lexical reference system whose design is inspired by current psycholinguistic theories of human lexical memory. Due to the great popularity of WN in Natural Language Processing (NLP), several authors (Magnini and Cavaglia, 2000), (Niles and Pease, 2001), (Niles and Pease, 2003), (Valitutti, 2004) have proposed to incorporate to the semantic net of WN, some taxonomies that characterize, in one or several concepts, the senses of each word. In spite of the fact that there have been developed a lot of WordNet's mappings, there isn't one unique resource to integrate all of them in a single system approach. To solve this need we have developed a resource that joins WN, the SUMO Ontology, WordNet Domainsand WordNet Affect. Our purpose is to test the advantages of having all the resources together for the resolution of the WSD task.

The rest of the paper is organized as follows. In Section 2 we describe the architecture of the integrative resource. Our approach is shown in Section 3. Next section presents the obtained results and a discussion. And finally the conclusions in Section 5.

Our proposal consists in accomplishing three runs with different algorithms. Both first utilize the domain's vectors; the third method utilizes k-cliques' techniques.

This work is divided in several stages:

1. Pre-processing of the corpus (lemmatization with Freeling) (Atserias et al, 2006).

2. Context selection (For the first (3.1), and the third (3.3) run the context window was constituted by the sentence that contains the word to disambiguate; in the second run the context window was constituted by the sentence that contains the word to disambiguate, the previous sentence and the next one).

3. Obtaining the domain vector, this vector is used in first and the second runs (when the lemma of the words in the analyzed sentence is obtained, the integrative resource of WordNet's Mappings is used to get the respective senses from each lemma).

4. Obtaining the all resource vector: SUMO, Affects, and Domain resource. This is only for the third run (3.3).

5. Relevant Semantic Tree construction (Addition of concepts parents to the vectors. For the first (3.1) and second (3.2) runs only Domain resource is used; for the third (3.3) run all the resources are used).

6. Selection of the correct senses (the first and the second runs use the same way to do the selection; the third run is different. We make an exception: For the verb "be" we select the sense with the higher frequency according to Freeling.

The conducted experiments measure the influence of the aforementioned resources in the disambiguation task. We have evaluated them individually and as a whole. In the Table 3 it is represented each one of the inclusions and combinations experimented with the Relevant Semantic Tree method.

As it can be observed, in the evaluation for specific domain corpus the best results are reached when only domain resource is used. But this is not a conclusion about the resources inclusion because the use of this method for global domain, for example with the task English All words from adding all the resources showed good results. This is due to the fact that the global domain includes information of different contexts, exactly what

is representing in the mentioned resources. For 430 this reason, in the experiment with global domain and the inclusion of all the resource obtained better results than using this method with specific domain, 42% of recall and 45% of precision (Gutierrez, 2010).

For example, with the k-clique's technique, utilizing the English All word task from Senseval-2's corpus, the results for the test with global dominion were: with single domain inclusion 40 % of precision and recall; but with the three resources 41.7 % for both measures.

Table 4 shows the obtained results for the test data set. The average performance of our system is 32% and we ranked on 27-th position from 27 participating systems. Although, we have used different sources of information and various approximations, in the future we have to surmount a number of obstacles.

One of the limitations comes from the usage of the POS-tagger Freeling which introduces some errors in the grammatical discrimination. Representing a loss of 3.7% in the precision of our system.

The base of knowledge utilized in the task was WordNet 1.6; but the competition demanded the results with WordNet 3.0. In order to achieve this we utilized mappings among versions where 119 of 1398 resulting senses emitted by Semeval-2 were did not found. This represents an 8.5%.

In our proposal, the sense belonging to the Factotum Domain was eliminated, what disabled that the senses linked to this domain went candidates to be recovered. 777 senses of 1398 annotated like correct for Semeval-2 belong to domain Factotum, what represents that the 66% were not recovered by our system. Considering the senses that are not correlated to Factotum, that is, that correlate to another domains, we are speaking about 621 senses to define; The system would emit results of a 72,4%. Senses selected correctly were 450, representing a 32%. However, 189 kept on like second candidates to be elected. This represents a 13.5%. If a technique of more precise decision takes effect, the results of the system could be increased largely.

For our participation in the Semeval-2 task 17 (All-words Word Sense Disambiguation on Specific Domain), we presented three methods for disambiguation approach which uses an Methods I Precision I Recall I Attempted integrative resource of WordNet mappings. We conducted an experimental study with the trail data set, according to which the RelevantSemantic Tree reaches the best performance. Our current approach can be improved with the incorporation of more granularities in the hierarchy of WordNet Domains. Because it was demonstrated that to define correct senses associated to specific domains an improvement of 72.4% is obtained. At this moment, only domain information is used in our first and second method. Besides was demonstrated for specific domains, the inclusion of several resources worsened the results with the first and second proposal method, the third one has been not experimented yet. Despite the fact that we have knowledge of SUMO, WordNet-Affect and WordNet Domain in our third method we still not obtain a relevant result.

It would be convenient to enrich our resource with other resources like Frame-Net, Concept-Net or others with the objective of characterizing even more the senses of the words.

This paper has been supported partially by Ministerio de Ciencia e Innovation - Spanish Government (grant no. TIN2009-13391-C04-01), and Conselleria d'Educacio - Generalitat Valenciana (grant no. PROMETEO/2009/119 and ACOMP/2010/288).