The task for Spanish is a "lexical sample" for 39 words (17 nouns, 13 verbs, and 9 adjectives). See table 1 for the complete list of all words selected for the Spanish lexical sample task. The words can belong only to one of the syntactic categories. The fourteen words selected to be translation-equivalents to English has been:

• Nouns: arte (=art), autoridad (= authority), canal {— channel), circuito (= circuit), and nat-uraleza (= nature).

• Verbs: conducir (= drive), tratar {— treat), and usar (= use).

• Adjectives: ciego (= blind), local (— local), nat­ural (= natural), simple (= simple), verde (= green), and vital (= vital).

The corpus was collected from two different sources: "El Periödico" (a Spanish newspaper) and LexEsp(a balanced corpus of 5.5 million words). The length of corpus samples is the sentence.

The lexicon provided was created specifically for the task and it consists of a definition for each sense linked to the Spanish version of EuroWordNet and, thus, to the English WordNet 1.5. The syntactic category and, sometimes, examples and synonyms are also provided. The connections to EuroWord­Net have been provided in order to have a common language independent conceptual structure. Neither proper nouns nor multiwords has been considered. We have also provided the complete mapping be­tween WordNet 1.5 and 1.6 versions. Each dictio­nary entry have been constructed consulting the corpus and multiple Spanish dictionaries (including the Spanish WordNet).

The Spanish SENSEVAL annotation procedure was divided into three consecutive phases.

• Corpus and dictionary creation

• Annotation

• Referee process

All these processes have been possible thanks to the effort of volunteers from three NLP groups from Universität Politècnica de Catalunya (UPC), Uni­versität de Barcelona (UB) and Universidad Na-cional de Educaciön a Distancia (UNED).

The Spanish lexical sample is a selection of higr medium and low polysemy frequent nouns, verbs an adjectives. The dictionary has 5.10 senses per wor and the polysemy degree ranges from 2 to 13. Noun has 3.94 ranging from 2 to 10, verbs 7.23 from 4 t 13 and adjectives 4.22 from 2 to 9 (see table 1 fo further details).

The lexical entries of the dictionary have the fol lowing form:

< SENSENUMBER ># < SYNSETNUMBERs >#

For instance, the dictionary for noun head wor < arte (= art) is:

arte#NCMS#l#Actividad humana o producto d tal actividad que expresa simbolicamente un as pecto de la realidad: el arte de la musica; el art precolombino #SIN:?#00518008n/02980374n?

arte#NCMS#2#Sabiduria, destreza o habilida< de una persona en una actividad o con ducta determinada: tiene mucho arte bai lando; desplego todo su arte para convencerl #SIN:?#03850627n#

We adopted, when possible, the guidelines propose* by the SENSEVAL organisers (Edmonds, 2000). Fo each word selected having n senses we provided a least 75 -f 15n instances. For the adjective popular ; larger set of instances has been provided to test per formance improvement when increasing the numbe of examples. These data has been then ramdomh divided in a ratio of 2:1 between training and tes set.

The corpus was structured following the standan SENSEVAL XML format.

In this section we discuss the most frequent and reg ular types of disagreement between annotators.

In particular, the dictionary proved to be not suf ficiently representative of the selected words to b< annotated. Although the dictionary was built fo: the task, out of 48% of the problems during the sec ond phase of the annotation where due to the lacl of the appropriate sense in the corresponding dictio­nary entry. This portion includes 5% of metaphori­cal uses not explicitly described into the dictionary entry. Furthermore, 51% of the problems reported by the annotators were concentrated only on five words (pasaje, canal, bomba, usar, and saltar).

Selecting only one sentence as a context during annotation was the other main problem. Around 26% of the problems where attributed to insufficient context to determine the appropriate sense.

Other sources of minor problems included differ­ent Part-of-Speech from the one selected for the word to be annotated, and sentences with multiple meanings.

In general, disagreement between annotators (and sometimes the use of multiple tags) must be inter­preted as misleading problems in the definition of the dictionary entries. The inter-tagger agreement between UPC and UNED teams was 0.64% and the Kappa measure 0.44%.