ILC-CNR Sample Clauses

ILC-CNR. As stated in section 2.1.3 ILC-CNR plans to develop a subcategorization acquisition system. We plan to adapt UCAM technologies for lexical builder and filtering methods to remove noisy SCFs. Evaluation for the domain-based exercise will be conducted as stated in D7.1. In the second trimester of 2010 we plan to start the development of a general domain SCF acquisition system. This system will be evaluated against dictionaries and manual inspection. For the domain based exercise for SCF, ILC-CNR will develop an evaluation corpus from two domains (20-30 verbs from each domain, with manually annotated examples from a relevant corpus). In line with UCAM workplan, human annotation time will be measured in order to determine how much benefit can be obtained by automatic acquisition of SCFs. Furthermore, in the encoding format of SCF we will work in the perspective of the LMF standard, which will use to facilitate the merging of dictionaries. Tentative timeline: end of 2010 have a general domain SCF acquisition system and lexicon evaluated. Domain adaptation task will be performed as soon as the crawled monolingual corpora will be available and after the creation of the domain specific data set.

ILC-CNR. The ultimate goal is to build a component for acquiring MWes from domain corpora for building or enriching lexica with collocational information. Since MWE is a wide area including various types of structures, work will be focussed on those MWEs which may benefit multilingual applications more and for which more robust methods exist: namely, nominal collocations (i.e. noun compounds, complex nominals, and adjective noun pairs). Thus, target MWEs will have the form NN (which in Italian is not very productive, but is still salient in domain terminologies), AdjN or Nadj, and N prep N. First a system following the n-gram with POS filtering corpora will be built as baseline and various AMs will be used for ranking the candidates as in most common state-of-the-art methods described above. Then the system will be adapted to work on chunked coupus data in order to reduce noise in the candidate list and experiments will be done also with dependency parsed data to assess improvements in performance. In fact, it must be assessed on the specific case of PANACEA whether a syntactic approach is better, as the potential errors in the parsed input corpus may affect MWE extraction (cfr. Seretan and Wehrli, 2009:78). Given that no ready-made tool will be used in this task, it design will try to take into account directly the requirements of the platform and in particular the fact that it should run as a webservice, which is no trivial issue given that MWE methods require processing of large quantities of data. Work will start on Italian, using first general purpose corpora for developing and the acquisition component based on the state-of-the-art methods, and then the monolingual domain corpora obtained in WP4 for domain tuning. The official evaluation will be performed on domain data only. Finally, given that the methods are relatively language independent (although in the literature it is reported that AM ranking works differently in different languages), the possibility of applying the same technology to English (and possibly other languages for which domain corpora will be crawled) will be assessed.