In this paper a Word Sense Disambiguation sys­tem based on Specification Marks (SM) and Maximum Entropy probability models (ME) is presented. SM is an unsupervised knowledge-based method and has been applied to noun disambiguation. ME belongs to the statistical approach to WSD in NLP and uses a tagged cor­pus in order to learn a probability model that can be used to predict the correct sense of a word. SM does not need a previously tagged corpus, it uses the semantic information stored in WordNet.

The weakness of supervised corpus-based ap­proaches rely on availability of corpora and their dependency of the data which were used in the training phase. Knowledge-based approaches use previously acquire linguistic knowledge. This knowledge is extracted from human lex­icographers experience and can be in form of electronic dictionary or lexicon. While their success seems poorest than statistical methods, they don't need neither an existing corpus nor a training phase and they can be more domain independent.

* This paper has been partially supported by the Span­ish Government (CICYT) project number TIC2000-0664-C02-02.

So, the University of Alicante system per­forms the WSD task combining unsupervised with supervised methods. The whole system has been used in lexical sample English task and lexical sample Spanish task.

The method we present here consists basically of the automatic sense-disambiguating of nouns that appear within the context of a sentence and whose different possible senses are quite re­lated. Its context is the group of words that co-occur with it in the sentence and their rela­tionship to the norm to be disambiguated. The disambiguation is resolved with the use of the WordNet lexical knowledge base.

The intuition underlying this approach is that the more similar two words are, the more infor­mative the most specific concept that subsumes them both will be. In other words, their low­est upper bound in the taxonomy (A "con­cept" here, corresponds to a Specification Mark (SM)). In other words, the more information two concepts share in common, the more similax they obviously are, and the information com­monly shared by two concepts is indicated by the concept that subsumes them in the taxon­omy.

The input for the WSD module will be the group of words W = {W\, W2,Wn}. Each word wi is sought in WordNet, each one has an associated set Si = {Sn, S^,Sin} of pos­sible senses. Furthermore, each sense has a set of concepts in the IS-A taxonomy (hyper-nymy/Hyponymy relations). First, the concept that is common to all the senses of all the words that form the context is sought. We call this concept the Initial Specification Mark (ISM), and if it does not immediately resolve the ambi­guity of the word, we descend from one level to another through WordNet's hierarchy, as­signing new Specification Marks. The number of concepts that contain the subhierarchy will then be counted for each Specification Mark. The sense that corresponds to the Specification Mark with highest number of words will then be chosen as the sense disambiguation of the noun in question, within its given context.

At this point, we should like to point out that after having evaluated the method, we subse­quently discovered that it could be improved with a set of heuristics, providing even better results in disambiguation. The set of heuristics are Heuristic of Hypernym, Heuristic of Defini­tion, Heuristic of Common Specification Mark, Heuristic of Gloss Hypernym, Heuristic of Hy-ponym and Heuristic of Gloss Hyponym. De­tailed explanation and evaluation of the method and heuristics can be found in (Montoyo and Palomar, 2000; Montoyo and Palomar, 2001), while its application to NLP tasks are addressed in (Montoyo et al., 2001).

Maximum Entropy(ME) modeling is a frame­work for integrating information from many heterogeneous information sources for classifica­tion. ME probability models were successfully applied to some NLP tasks such as POS tagging or sentence boundary detection (Ratnaparkhi, 1998).

The WSD system presented in this paper is based on conditional ME probability mod­els (Saiz-Noeda et al., 2001). It implements a supervised learning method consisting of the building of word sense classifiers through train­ing on a semantically tagged corpus. A classifier obtained by means of a ME technique consists of a set of parameters or coefficients estimated by means of an optimization procedure. Each co­efficient is associated to one feature observed in training data. A feature is a function that gives a measure for some characteristic in a context associated to a class. The main purpose is to obtain the probability distribution that maxi­mizes the entropy, that is, maximum ignorance is assumed and nothing apart of training data is considered. As advantages of ME framework, knowledge-poor features applying and accuracy can be mentioned; ME framework allows a vir­tually unrestricted ability to represent problemspecific knowledge in the form of features (Rat­naparkhi, 1998).

Let us assume a set of contexts X and a set of classes C. The function cl : X -> C that performs the classification in a condi­tional probability model p chooses the class with the highest conditional probability: cl(x) = argmaxcp(c|a;), where a; is a context and c a class. The features have the form of (1), where cp(x) is some observable characteristic. The conditional probability p(c\x) is defined as (2) where ai are the parameters or weights of each feature, and Z(x) is a constant to ensure that the sum of probabilities for each possible class in this context is equal to 1.

The Spanish and English lexical sample tasks at the Senseval-2 workshop had been performed by our system in three phases. The first one is a naive multi-word detection; the second one, the disambiguation of nouns by means of the SM method, and the third one, the disambigua­tion of verbs and adjectives by means of the ME method.

In a previous step, training and test data had been tagged with Tree-Tagger(Schmid, 1994) for English files and Conexor's FDG Parser (Tapanainen and Järvinen, ) for Spanish files in order to get the part-of-speech information and identify nouns, verbs and adjectives.

Multi-words detection

The multi-word detection has been performed by combining the words around the target word in each sample and consulting WordNet for En­glish (examining the training data, we conclude that this is not necessary for Spanish data). If a multi-word is found in WordNet a multi-word instance is assigned and no further single word disambiguation will be done. This kind of in­stances has been disambiguated with the first sense of WordNet (even if it is a polysemous one).

Nouns with Specification Marks

The second phase consist of noun classification, and has been performed by the SM method de­scribed previously.

Verbs and adjectives with Maximum Entropy

The third and final phase, the verbs and ad­jectives disambiguation, has been performed by the ME method. The Senseval-2 training data has been used in order to obtain the classifica­tion functions to be applied on the test data. The set of features defined for ME training is described below and it is based on features se­lection made in (Ng and Lee, 1996) and (Escud-ero et al., 2000).

The set of features corresponds to words around the word to classify and POS la­bels at positions related to the target word in each sentence: wq, w_25 w_35 p+i, p+2? P+3- Each W{ is the lemma of the word at position i in the context (in collocations, at least one of the words must be a content word). Each pi is the POS label at position i.

Other set of features consists of a surround­ing nouns selection. This selection is doing by means of frequency information of norms co-occurring with a sense. Nouns co-occurring with a class in a K% of examples of that class in the corpus or more are selected to build a feature for each possible class.

Analyzing the first evaluation results of the English lexical sample task (fine-grained scoring) reported by Senseval-2 committee (precision = 0.421 and recall = 0.411) , some conclusions can be extracted from them.

The nouns disambiguation obtains the worst results (see table 1). We can mostly assure that the reason is the kind of method used: knowledge-based for nouns and corpus-based for verbs and adjectives.

The results of the Spanish lexical sample task (fine-grained scoring) reported by Senseval-2 committee are precision = 0.514 and recall = 0.503. Nevertheless, the nouns results rise to 56% of precision (table 2). It seems that the set of nouns selected for this task is easier to Specification Marks than English ones, maybe related to lexical resources used and the lan­guage itself. However, the recall of nouns is too low because a implementation error causes that the accented words had not been recognize (corazon, operacion and organo).

The preprocessing of the train and test data are relevant. Some errors of the POS-tagger had been detected and they affect some answer in­stances. Multi-words are a not resolved prob­lem. The detection and disambiguation method is too simple and causes too much errors. More preprocessing is necessary, as well: the con­text information can be enriched and accuracy increased with entity recognition, full-parsing, and so on.

The University of Alicante system presented at Senseval-2 workshop joins the two general ap­proaches to the WSD task: knowledge-based and corpus-based methods. The Specification Marks method belongs to the first one and Max­imum Entropy-based method to the second one.

At this moment, the two methods presented here are being improved with new knowledge sources like full parsing information and domain categories that in order to decrease the Word-Net granularity. The WSD system will be com­pleted with other NLP software like Name En­tity recognition and multi-words detection mod­ules.

Recent work in our research group indicates that it is possible to combine the two methods in a hybrid method that assign a sense to a context combining the answers of both methods with a relevant improvement of accuracy (Suârez and Montoyo, 2001). Our intention is to extent this combination with the help of other well known WSD methods and to establish a voting method or some other manner of cooperation.

Our main objective is to develop a complete WSD system in order to help other NLP activ­ities in our research group. The work presented here is our first attempt to participate at Sen-seval and we hope to get the proper conclusions in order to improve our system and compete in the next Senseval. References

Gerard Escudero, Lluis Marquez, and Ger­man Rigau. 2000. Boosting applied to word sense disambiguation. In Proceedings of the 12th Conference on Machine Learning ECML2000, Barcelona, Spain.

A. Montoyo and M. Palomar. 2000. Word Sense Disambiguation with Specification Marks in Unrestricted Texts, pages 103-107.

A. Montoyo and M. Palomar. 2001. Specifi­cation Marks for Word Sense Disambigua­tion: New Development. In Proceedings of 2nd International conference on Intelligent Text Processing and Computational Linguis­tics (CICLing-2001), pages 182-191.

A. Montoyo, M. Palomar, and G. Rigau. 2001. WordNet Enrichment with Classification Sys­tems. In ACL, editor, Proceedings of NA ACL Workshop WordNet and Other Lexical Re­sources: Applications, Extensions and Cus­tomization^ Pittsburgh, PA, USA.

Hwee Tou Ng and Hian Beng Lee. 1996. In­tegrating multiple knowledge sources to dis­ambiguate word senses: An exemplar-based approach. In Proceedings 34th Annual Meet­ing of the ACL-1996., San Francisco, USA.

Adwait Ratnaparkhi. 1998. Maximum Entropy Models for Natural Language Ambiguity Res­olution. Ph.D. thesis, University of Pennsyl­vania.

Maximiliano Saiz-Noeda, Armando Suârez, and Manuel Palomar. 2001. Semantic pattern learning through maximum entropy-based wsd technique. In Proceedings of CoNLL-2001, pages 23-29. Toulouse, France.

Helmut Schmid. 1994. Probabilistic part-of-speech tagging using decision trees. In Pro­ceedings International Conference on New Methods in Language Processing., pages 44­49, Manchester, UK.

Armando Suârez and Andrés Montoyo. 2001. Estudio de cooperaciön entre métodos de desambiguaciön léxica: Marcas de especifi-cidad vs. maxima entropia. Procesamiento Lenguaje Natural, 27(1):207-214, September.

Pasi Tapanainen and Timo Järvinen. A non-projective dependency parser. In Proceedings

» of the Fifth Conference on Applied Natural Language Processing, pages 64-71.