Chared is a software tool which can detect character encoding of a text document provided the language of the document is known. The language of the text has to be specified as an input parameter so that the corresponding language model can be used. The package contains models for a wide range of languages (currently 57 --- covering all major languages). Furthermore, it provides a training script to learn models for additional languages using a set of user supplied sample html pages in the given language. The detection algorithm is based on determining similarity of byte trigrams vectors. In general, chared should be more accurate than other character encoding detection tools with no language constraints. This is an important advantage allowing precise character decoding needed for building large textual corpora. The tool has been used for building corpora in American Spanish, Arabic, Czech, French, Japanese, Russian, Tajik, and six Turkic languages consisting of 70 billions tokens altogether. Chared is an open source software, licensed under New BSD License and available for download (including the source code) at http://code.google.com/p/chared/. The research leading to this piece of software was published in POMIKÁLEK, Jan a Vít SUCHOMEL. chared: Character Encoding Detection with a Known Language. In Aleš Horák, Pavel Rychlý. RASLAN 2011. 5. vyd. Brno, Czech Republic: Tribun EU, 2011. od s. 125-129, 5 s. ISBN 978-80-263-0077-9. and PRESEMT, Lexical Computing Ltd
In NLP Centre, dividing text into sentences is currently done with
a tool which uses rule-based system. In order to make enough training
data for machine learning, annotators manually split the corpus of contemporary text
CBB.blog (1 million tokens) into sentences.
Each file contains one hundredth of the whole corpus and all data were
processed in parallel by two annotators.
The corpus was created from ten contemporary blogs:
hintzu.otaku.cz
modnipeklo.cz
bloc.cz
aleneprokopova.blogspot.com
blog.aktualne.cz
fuchsova.blog.onaidnes.cz
havlik.blog.idnes.cz
blog.aktualne.centrum.cz
klusak.blogspot.cz
myego.cz/welldone
AGREE is a dataset and task for evaluation of language models based on grammar agreement in Czech. The dataset consists of sentences with marked suffixes of past tense verbs. The task is to choose the right verb suffix which depends on gender, number and animacy of subject. It is challenging for language models because 1) Czech is morphologically rich, 2) it has relatively free word order, 3) high out-of-vocabulary (OOV) ratio, 4) predicate and subject can be far from each other, 5) subjects can be unexpressed and 6) various semantic rules may apply. The task provides a straightforward and easily reproducible way of evaluating language models on a morphologically rich language.
Sentence-parallel corpus made from English and Czech Wikipedias based on translated articles from English into Czech.
The work done is described in the paper: ŠTROMAJEROVÁ, Adéla, Vít BAISA a Marek BLAHUŠ. Between Comparable and Parallel: English-Czech Corpus from Wikipedia. In RASLAN 2016 Recent Advances in Slavonic Natural Language Processing. Brno: Tribun EU, 2016. s. 3-8, 6 s. ISBN 978-80-263-1095-2.
jusText is a heuristic based boilerplate removal tool useful for cleaning documents in large textual corpora. The tool has been implemented in Python, licensed under New BSD License and made an open source software (available for download including the source code at http://code.google.com/p/justext/). It is successfully used for cleaning large textual corpora at Natural language processing centre at Faculty of informatics, Masaryk university Brno and it's industry partners. The research leading to this piece of software was published in author's Ph.D. thesis "Removing Boilerplate and Duplicate Content from Web Corpora". The boilerplate removal algorithm is able to remove most of non-grammatical sentences from a web page like navigation, advertisements, tables, short notes and so on. It has been shown it overperforms or at least keeps up with it's competitors (according to comparison with participants of Cleaneval competition in author's Ph.D. thesis). The precise removal of unwanted content and scalability of the algorithm has been demonstrated while building corpora of American Spanish, Arabic, Czech, French, Japanese, Russian, Tajik, and six Turkic languages consisting --- over 20 TB of HTML pages were processed resulting in corpora of 70 billions tokens altogether. and PRESEMT, Lexical Computing Ltd