The goal of the Cross-Lingual Lexical Substitu­tion (CLLS) task is to substitute a word in a lan­guage Ls, which occurs in a particular context, by providing the best substitutions in a different language Lt. In SemEval-2010 the source lan­guage Ls is English, while the target language Lt is Spanish. Clearly, this task is related to Lexical Giovanni Semeraro

Substitution (LS) (McCarthy and Navigli, 2007) which consists in selecting an alternative word for a given one in a particular context by preserving its meaning. The main difference between the LS task and the CLLS one is that in LS source and target languages are the same. CLLS is not a easy task since neither a list of candidate words nor a specific parallel corpus are supplied by the orga­nizers. However, this opens the possibility of us­ing several knowledge sources, instead of a single one fixed by the task organizers. Therefore, the system must identify a set of candidate words in Lt and then select only those words which fit the context. From another point of view, the cross-lingual nature of the task allows to exploit auto­matic machine translation methods, hence the goal is to find as many good translations as possible for the given target word. A thorough description of the task can be found in (Mihalcea et al., 2010; Sinhaetal.,2009).

To easily understand the task, an example fol­lows. Consider the sentence:

During the siege, George Robertson had appointed Shuja-ul-Mulk, who was a bright boy only 12 years old and the youngest surviving son of Aman-ul-Mulk, as the ruler of Chitral.

In the previous sentence the target word is "bright". Taking into account the meaning of the word "bright" in this particular context, the best substitutions in Spanish are: "inteligente", "brillante" and "listo".

We propose two systems to tackle the problem of CLLS: the first is based on an automatic trans­lation system which exploits the API of Google Translator, the second is based on a parallel cor­pus approach which relies on Wikipedia. In par­ticular, in the second approach we use a struc­tured version of Wikipedia called DBpedia (Bizer et al., 2009). Both systems adopt several lexical resources to select the list of possible substitutions for a given word. Specifically, we use three differ­ent dictionaries: Google Dictionary, Babylon Dic­tionary and Spanishdict. Then, we combine the dictionaries into a single one, as described in Sec­tion 2.1.

The paper is organized as follows: Section 2 de­scribes the strategy we adopted to tackle the CLLS task, while results of an experimental session we carried out in order to evaluate the proposed ap­proaches are presented in Section 3. Conclusions are discussed in Section 4.

Generally speaking, the problem of CLLS can be coped with a strategy which consists of two steps, as suggested in (Sinha et al., 2009):

• candidate collection: in this step several re­sources are queried to retrieve a list of po­tential translation candidates for each target word and part of speech;

• candidate selection: this step concerns the ranking of potential candidates, which are the most suitable ones for each instance, by using information about the context.

Regarding the candidate collection, we exploit three dictionaries: Google Dictionary, Babylon Dictionary and Spanishdict. Each dictionary is modeled using a strategy described in Section 2.1. We use the same approach to model each dictio­nary in order to make easy both the inclusion of fu­ture dictionaries and the integration with the can­didate selection step.

Candidate selection is performed in two dif­ferent ways. The first one relies on the auto­matic translation of the sentence in which the tar­get word occurs, in order to find the best substitu­tions. The second method uses a parallel corpus built on DBpedia to discover the number of doc­uments in which the target word is translated by one of the potential translation candidates. Details about both methods are reported in Section 2.2

• Sen'- The raw honesty of that basic crudeness makes you feel stronger in a way.

• Ses: La cruda honestidad de esa crudeza de base que te hace sentir mas fuerte en un camino.

2. no potential candidates occur into the trans­lation, but a correct translation of the target word is provided:

• Sen: Many institutional investors are now deciding that they are getting a raw deal from the company boards of Aus­tralia.

• Ses: Muchos inversores institucionales estan ahora decidiendo que estan recibi-endo un trato injusto de los directorios de las empresas de Australia.

In detail, the strategy can be split in several steps:

1. Retrieve the list M of potential translation candidates using the method described in Section 2.1.

2. Translate the target sentence Sen from En­glish to Spanish, using the google-api-translate-java, which results into the sentence

3. Enrich M by adding multiword expressions. To implement this step, the two bigrams which contain the target word and the only trigram in which the target word is the 2ndterm are taken into to account.

Coming back to the first sentence in the previ­ous example, the following n-grams are built: "the raw", "raw honesty" and "the raw hon­esty". For each n-gram, candidate transla­tions are looked for using Google Dictionary. If translations are found, they are added to M with an initial score equal to 0.

4. Fix a window W of n words to the right and to the left of the target word, and perform the following steps:

(a) for each candidate in M, try to find Cfc in W. If Cfc occurs in W, then add 2 to the score of in M;

(b) if no exact match is found in the previ­ous step, perform a new search by com­paring Cfc with the words in W using the Levenshtein distance(Levenshtein, 1966). If the Levenshtein distance is greater than 0.8, then add 2 to the score of Cfc in M.

5. If no exact/partial match is found in the pre­vious steps, probably the target word is trans­lated with a word which does not belong to M. To overcome this problem, we implement a strategy able to discover a possible transla­tion in Ses which is not in M. This approach involves three steps:

(a) for each word Wi in Sen, a list of poten­tial translations Pi is retrieved;

(b) if a word in Pi is found in Ses, the word is removed from Ses ;

(c) at this point, Ses contains a list R of words with no candidate translations. A score is assigned to those words by tak­ing into account their position in Seswith respect to the position of the target word in Sen, using the following equa­tion:

where posc is the translation candidate position in Ses, post is the target word position in Sen and Lmax is the maxi­mum length between the length of Senand Ses-

Moreover, the words not semanti-cally related to the potential candidates (found using Spanish WordNet) are re­moved from R. In detail, for each candi­date in M a list of semantically related words in Spanish WordNet is retrieved which results in a set WN of related words. Words in R but not in WN are removed from R. In the final step, the list R is sorted and the first word in R is added to M assigning a score equal to 2.

6. In the last step, the list M is sorted. The out­put of this process is the ranked list of poten­tial candidates.

It is important to underline that both Sen and Ses are tokenized, part-of-speech tagged and lem-matized. Lemmatization plays a key role in the matching step, while part-of-speech tagging is needed to query both the dictionaries and the Spanish WordNet. We adopt META (Basile et al., 2008) and FreeLing (Atserias et al., 2006) to per­form text processing for English and Spanish re­spectively.

The second proposed system, called unibaWiki, is based on the idea of automati­cally building a parallel corpus from Wikipedia. We use a structured version of Wikipedia called DBpedia (Bizer et al., 2009). The main idea be­hind DBpedia is to extract structured information from Wikipedia and then to make this information available. The main goal is to have access easily to the large amount of information in Wikipedia. DBpedia opens new and interesting ways to use Wikipedia in NLP applications.

In CLLS task, we use the extended abstracts of English and Spanish provided by DBpedia. For each extended abstract in Spanish which has the corresponding extended abstract in English, we build a document composed by two fields: the for­mer contains the English text (texten) and the lat­ter contains the Spanish text (textes)- We adopt Lucene as storage and retrieval engine to make the documents access fast and easy.

The idea behind unibaWiki is to count, for each potential candidate, the number of docu­ments in which the target word occurs in textenand the potential candidate occurs in textes- A score equal to the number of retrieved documents is assigned, then the candidates are sorted accord­ing to that score.

Given the list M of potential candidates and the target word t, for each e M we perform the following query:

texten '■ t AND textes '■ Ck where the field name is followed by a colon and by the term you are looking for.

It is important to underline here that multiword expressions require a specific kind of query. For each multiword expression we adopt the Phrase-Query which is able to retrieve documents that contain a specific sequence of words instead of a single keyword.

The goal of the evaluation is to measure the sys­tems' ability to find correct Spanish substitutions for a given word. The dataset supplied by the or­ganizers contains 1,000 instances in XML format.

We described our participation at SemEval-2 Cross-Lingual Lexical Substitution Task, propos­ing two systems called UBA-T and UBA-W. The first relies on Google Translator, the second is based on DBpedia, a structured version of Wikipedia. Moreover, we exploited several dictio­naries to retrieve the list of candidate substitutions.

UBA-T achieves the highest recall among all the participants to the task. Moreover, the results proved that the method based on Google Transla­tor is more effective than the one based on DBpe­dia.

This research was partially funded by Regione Puglia under the contract POR PUGLIA 2007-2013 - Asse I Linea 1.1 Azione 1.1.2 - Bando "Aiuti agli Investimenti in Ricerca per le PMI" -Fondo per le Agevolazioni alia Ricerca, project ti­tle: "Natural Browsing".