My MsC project

Starting Time: 1998

download the full corpus

download the journalistic texts     download the literary texts     download the didactic texts

download the tagset

2) Four taggers available in the WWW have been trained on this 100,000-word corpus, namely Unigram (Treetagger), N-gram (Treetagger), transformation-based (TBL) and Maximum-Entropy tagging (MXPOST). The latter displayed the best accuracy (88.73%), which is still much lower than the state-of-the-art accuracy for English. The low accuracy is attributed to the reduced size of the training corpus. 

3) A symbolic tagger has been developed _ PoSiTagger (Aires, 2000). download rules files

4) Twelve methods of combination were used, four of which led to an improvement over the MXPOST accuracy. The best result (89.42%) was obtained with a majority-wins voting strategy.

Team
Rachel Virgínia Xavier Aires - MSc Student

Sandra Maria Aluísio - Supervisor

Marcio Luis Barse Andreeta - a student who worked on the tagging of the training-test corpus and on the codification of several tools for combining the taggers and evaluate them, [1998-2000]

Ronaldo Teixeira Martins - The linguist who wrote the version of the NILC tagset used on this project

Denise Khun - The linguist who wrote most of the PoSiTagger rules, [June 2000]

Ana Raquel Marchi - The linguist who worked on the correction of the training-test corpus, [June 2000]

• Since November 2000, I have made available a Portuguese model for MXPOST that was trained using a much simpler tagset than the one used on my MsC project. We considered only 27 tags plus punctuation marks tags, achieving 97% accuracy. Even though we have used a 10 folders cross-validation test strategy, the accuracy should not be generalized to texts in general. We must remember that the corpus used during the training is small ~ 100,000 words, and for this reason it is not a representative model of the Portuguese language in general. It was showed on the MsC project that the precision is different for each of the three genres studied and that the journalistic genre is the one with less ambiguity and the easiest one to tag. Since November 2000, I have made available a Portuguese model for MXPOST that was trained using a much simpler tagset than the one used on this MsC project. We considered only 27 tags plus punctuation marks tags, achieving 97% accuracy. Even though we have used a 10 folders cross-validation test strategy, the accuracy should not be generalized to texts in general. We must remember that the corpus used during the training is small _ 100,000 words, and for that reason it is not a representative model of the Portuguese language in general. It was showed on the MsC project that the precision is different for each of the tree genres studied, and that the journalistic used on most tests on the literature, is the one with less ambiguity and the easiest one to tag. You can download the model, the tagset and the evaluation results per tag. 

• In another project _  LacioWeb, MXPOST, TreeTagger  and TBL have been trained with the MAC-MORPHO Corpus, a corpus with 1,221,468 words. The best precision achieved for MXPOST was 96.98% with 22 tags plus contraction and punctuation tags.

Aires, R.V.X.; Aluísio, S.M.(1999). Um Etiquetador para o Português do Brasil. In IV Workshop de Teses e Dissertações em Andamento do ICMC/USP. São Carlos, 1999. p.57-58.