This small dataset contains 3 speech corpora collected using the Alex Translate telephone service (https://ufal.mff.cuni.cz/alex#alex-translate).
The "part1" and "part2" corpora contain English speech with transcriptions and Czech translations. These recordings were collected from users of the service. Part 1 contains earlier recordings, filtered to include only clean speech; Part 2 contains later recordings with no filtering applied.
The "cstest" corpus contains recordings of artificially created sentences, each containing one or more Czech names of places in the Czech Republic. These were recorded by a multinational group of students studying in Prague.
We present a test corpus of audio recordings and transcriptions of presentations of students' enterprises together with their slides and web-pages. The corpus is intended for evaluation of automatic speech recognition (ASR) systems, especially in conditions where the prior availability of in-domain vocabulary and named entities is benefitable.
The corpus consists of 39 presentations in English, each up to 90 seconds long, and slides and web-pages in Czech, Slovak, English, German, Romanian, Italian or Spanish.
The speakers are high school students from European countries with English as their second language.
We benchmark three baseline ASR systems on the corpus and show their imperfection.
Additional three Czech reference translations of the whole WMT 2011 data set (http://www.statmt.org/wmt11/test.tgz), translated from the German originals. Original segmentation of the WMT 2011 data is preserved. and This project has been sponsored by the grants GAČR P406/11/1499 and EuroMatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003+7E11051 of the Ministry of Education, Youth and Sports of the Czech Republic)
Data
-------
Bengali Visual Genome (BVG for short) 1.0 has similar goals as Hindi Visual Genome (HVG) 1.1: to support the Bengali language. Bengali Visual Genome 1.0 is the multi-modal dataset in Bengali for machine translation and image
captioning. Bengali Visual Genome is a multimodal dataset consisting of text and images suitable for English-to-Bengali multimodal machine translation tasks and multimodal research. We follow the same selection of short English segments (captions) and the associated images from Visual Genome as HGV 1.1 has. For BVG, we manually translated these captions from English to Bengali taking the associated images into account. The manual translation is performed by the native Bengali speakers without referring to any machine translation system.
The training set contains 29K segments. Further 1K and 1.6K segments are provided in development and test sets, respectively, which follow the same (random) sampling from the original Hindi Visual Genome. A third test set is
called the ``challenge test set'' and consists of 1.4K segments. The challenge test set was created for the WAT2019 multi-modal task by searching for (particularly) ambiguous English words based on the embedding similarity and
manually selecting those where the image helps to resolve the ambiguity. The surrounding words in the sentence however also often include sufficient cues to identify the correct meaning of the ambiguous word.
Dataset Formats
---------------
The multimodal dataset contains both text and images.
The text parts of the dataset (train and test sets) are in simple tab-delimited plain text files.
All the text files have seven columns as follows:
Column1 - image_id
Column2 - X
Column3 - Y
Column4 - Width
Column5 - Height
Column6 - English Text
Column7 - Bengali Text
The image part contains the full images with the corresponding image_id as the file name. The X, Y, Width and Height columns indicate the rectangular region in the image described by the caption.
Data Statistics
---------------
The statistics of the current release are given below.
Parallel Corpus Statistics
--------------------------
Dataset Segments English Words Bengali Words
---------- -------- ------------- -------------
Train 28930 143115 113978
Dev 998 4922 3936
Test 1595 7853 6408
Challenge Test 1400 8186 6657
---------- -------- ------------- -------------
Total 32923 164076 130979
The word counts are approximate, prior to tokenization.
Citation
--------
If you use this corpus, please cite the following paper:
@inproceedings{hindi-visual-genome:2022,
title= "{Bengali Visual Genome: A Multimodal Dataset for Machine Translation and Image Captioning}",
author={Sen, Arghyadeep
and Parida, Shantipriya
and Kotwal, Ketan
and Panda, Subhadarshi
and Bojar, Ond{\v{r}}ej
and Dash, Satya Ranjan},
editor={Satapathy, Suresh Chandra
and Peer, Peter
and Tang, Jinshan
and Bhateja, Vikrant
and Ghosh, Anumoy},
booktitle= {Intelligent Data Engineering and Analytics},
publisher= {Springer Nature Singapore},
address= {Singapore},
pages = {63--70},
isbn = {978-981-16-6624-7},
doi = {10.1007/978-981-16-6624-7_7},
}
COSTRA 1.0 is a dataset of Czech complex sentence transformations. The dataset is intended for the study of sentence-level embeddings beyond simple word alternations or standard paraphrasing.
The dataset consist of 4,262 unique sentences with average length of 10 words, illustrating 15 types of modifications such as simplification, generalization, or formal and informal language variation.
The hope is that with this dataset, we should be able to test semantic properties of sentence embeddings and perhaps even to find some topologically interesting “skeleton” in the sentence embedding space.
Costra 1.1 is a new dataset for testing geometric properties of sentence embeddings spaces. In particular, it concentrates on examining how well sentence embeddings capture complex phenomena such paraphrases, tense or generalization. The dataset is a direct expansion of Costra 1.0, which was extended with more sentences and sentence comparisons.
CsEnVi Pairwise Parallel Corpora consist of Vietnamese-Czech parallel corpus and Vietnamese-English parallel corpus. The corpora were assembled from the following sources:
- OPUS, the open parallel corpus is a growing multilingual corpus of translated open source documents.
The majority of Vi-En and Vi-Cs bitexts are subtitles from movies and television series.
The nature of the bitexts are paraphrasing of each other's meaning, rather than translations.
- TED talks, a collection of short talks on various topics, given primarily in English, transcribed and with transcripts translated to other languages. In our corpus, we use 1198 talks which had English and Vietnamese transcripts available and 784 talks which had Czech and Vietnamese transcripts available in January 2015.
The size of the original corpora collected from OPUS and TED talks is as follows:
CS/VI EN/VI
Sentence 1337199/1337199 2035624/2035624
Word 9128897/12073975 16638364/17565580
Unique word 224416/68237 91905/78333
We improve the quality of the corpora in two steps: normalizing and filtering.
In the normalizing step, the corpora are cleaned based on the general format of subtitles and transcripts. For instance, sequences of dots indicate explicit continuation of subtitles across multiple time frames. The sequences of dots are distributed differently in the source and the target side. Removing the sequence of dots, along with a number of other normalization rules, improves the quality of the alignment significantly.
In the filtering step, we adapt the CzEng filtering tool [1] to filter out bad sentence pairs.
The size of cleaned corpora as published is as follows:
CS/VI EN/VI
Sentence 1091058/1091058 1113177/1091058
Word 6718184/7646701 8518711/8140876
Unique word 195446/59737 69513/58286
The corpora are used as training data in [2].
References:
[1] Ondřej Bojar, Zdeněk Žabokrtský, et al. 2012. The Joy of Parallelism with CzEng 1.0. Proceedings of LREC2012. ELRA. Istanbul, Turkey.
[2] Duc Tam Hoang and Ondřej Bojar, The Prague Bulletin of Mathematical Linguistics. Volume 104, Issue 1, Pages 75–86, ISSN 1804-0462. 9/2015
CUBBITT En-Cs translation models, exported via TensorFlow Serving, available in the Lindat translation service (https://lindat.mff.cuni.cz/services/translation/).
Models are compatible with Tensor2tensor version 1.6.6.
For details about the model training (data, model hyper-parameters), please contact the archive maintainer.
Evaluation on newstest2014 (BLEU):
en->cs: 27.6
cs->en: 34.4
(Evaluated using multeval: https://github.com/jhclark/multeval)
CUBBITT En-Fr translation models, exported via TensorFlow Serving, available in the Lindat translation service (https://lindat.mff.cuni.cz/services/translation/).
Models are compatible with Tensor2tensor version 1.6.6.
For details about the model training (data, model hyper-parameters), please contact the archive maintainer.
Evaluation on newstest2014 (BLEU):
en->fr: 38.2
fr->en: 36.7
(Evaluated using multeval: https://github.com/jhclark/multeval)
CUBBITT En-Pl translation models, exported via TensorFlow Serving, available in the Lindat translation service (https://lindat.mff.cuni.cz/services/translation/).
Models are compatible with Tensor2tensor version 1.6.6.
For details about the model training (data, model hyper-parameters), please contact the archive maintainer.
Evaluation on newstest2020 (BLEU):
en->pl: 12.3
pl->en: 20.0
(Evaluated using multeval: https://github.com/jhclark/multeval)
Czech subjectivity lexicon, i.e. a list of subjectivity clues for sentiment analysis in Czech. The list contains 4626 evaluative items (1672 positive and 2954 negative) together with their part of speech tags, polarity orientation and source information.
The core of the Czech subjectivity lexicon has been gained by automatic translation of a freely available English subjectivity lexicon downloaded from http://www.cs.pitt.edu/mpqa/subj_lexicon.html. For translating the data into Czech, we used parallel corpus CzEng 1.0 containing 15 million parallel sentences (233 million English and 206 million Czech tokens) from seven different types of sources automatically annotated at surface and deep layers of syntactic representation. Afterwards, the lexicon has been manually refined by an experienced annotator. and The work on this project has been supported by the GAUK 3537/2011 grant and by SVV project number 267 314.
CzEng 1.0 is the fourth release of a sentence-parallel Czech-English corpus compiled at the Institute of Formal and Applied Linguistics (ÚFAL) freely available for non-commercial research purposes.
CzEng 1.0 contains 15 million parallel sentences (233 million English and 206 million Czech tokens) from seven different types of sources automatically annotated at surface and deep (a- and t-) layers of syntactic representation. and EuroMatrix Plus (FP7-ICT-2007-3-231720 of the EU and 7E09003+7E11051 of the Ministry of Education, Youth and Sports of the Czech Republic),
Faust (FP7-ICT-2009-4-247762 of the EU and 7E11041 of the Ministry of Education, Youth and Sports of the Czech Republic),
GAČR P406/10/P259,
GAUK 116310,
GAUK 4226/2011
Czech-Slovak parallel corpus consisting of several freely available corpora (Acquis [1], Europarl [2], Official Journal of the European Union [3] and part of OPUS corpus [4] – EMEA, EUConst, KDE4 and PHP) and downloaded website of European Commission [5]. Corpus is published in both in plaintext format and with an automatic morphological annotation.
References:
[1] http://langtech.jrc.it/JRC-Acquis.html/
[2] http://www.statmt.org/europarl/
[3] http://apertium.eu/data
[4] http://opus.lingfil.uu.se/
[5] http://ec.europa.eu/ and This work has been supported by the grant Euro-MatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003 of the Czech Republic)
CzEng 0.7 is a Czech-English parallel corpus compiled at the Institute of Formal and Applied Linguistics (ÚFAL), Charles University, Prague. The corpus contains no manual annotation. It is limited only to texts which have been already available in an electronic form and which are not protected by authors' rights in the Czech Republic. The main purpose of the corpus is to support Czech-English and English-Czech machine translation research with the necessary data. CzEng 0.7 consists of a large set of parallel textual documents mainly from the fields of European law, information technology, and fiction, all of them converted into a uniform XML-based file format and provided with automatic sentence alignment.
ELITR Minuting Corpus consists of transcripts of meetings in Czech and English, their manually created summaries ("minutes") and manual alignments between the two.
Czech meetings are in the computer science and public administration domains and English meetings are in the computer science domain.
Each transcript has one or multiple corresponding minutes files. Alignments are only provided for a portion of the data.
This corpus contains 59 Czech and 120 English meeting transcripts, consisting of 71097 and 87322 dialogue turns respectively. For Czech meetings, we provide 147 total minutes with 55 of them aligned. For English meetings, it is 256 total minutes with 111 of them aligned.
Please find a more detailed description of the data in the included README and stats.tsv files.
If you use this corpus, please cite:
Nedoluzhko, A., Singh, M., Hledíková, M., Ghosal, T., and Bojar, O.
(2022). ELITR Minuting Corpus: A novel dataset for automatic minuting
from multi-party meetings in English and Czech. In Proceedings of the
13th International Conference on Language Resources and Evaluation
(LREC-2022), Marseille, France, June. European Language Resources
Association (ELRA). In print.
@inproceedings{elitr-minuting-corpus:2022,
author = {Anna Nedoluzhko and Muskaan Singh and Marie
Hled{\'{\i}}kov{\'{a}} and Tirthankar Ghosal and Ond{\v{r}}ej Bojar},
title = {{ELITR} {M}inuting {C}orpus: {A} Novel Dataset for
Automatic Minuting from Multi-Party Meetings in {E}nglish and {C}zech},
booktitle = {Proceedings of the 13th International Conference
on Language Resources and Evaluation (LREC-2022)},
year = 2022,
month = {June},
address = {Marseille, France},
publisher = {European Language Resources Association (ELRA)},
note = {In print.}
}
Eyetracked Multi-Modal Translation (EMMT) is a simultaneous eye-tracking, 4-electrode EEG and audio corpus for multi-modal reading and translation scenarios. It contains monocular eye movement recordings, audio data and 4-electrode wearable electroencephalogram (EEG) data of 43 participants while engaged in sight translation supported by an image.
The details about the experiment and the dataset can be found in the README file.
English-Hindi parallel corpus collected from several sources. Tokenized and sentence-aligned. A part of the data is our patch for the Emille parallel corpus. and FP7-ICT-2007-3-231720 (EuroMatrix Plus) 7E09003 (Czech part of EM+)
English-Slovak parallel corpus consisting of several freely available corpora (Acquis [1], Europarl [2], Official Journal of the European Union [3] and part of OPUS corpus [4] – EMEA, EUConst, KDE4 and PHP) and downloaded website of European Commission [5]. Corpus is published in both in plaintext format and with an automatic morphological annotation.
References:
[1] http://langtech.jrc.it/JRC-Acquis.html/
[2] http://www.statmt.org/europarl/
[3] http://apertium.eu/data
[4] http://opus.lingfil.uu.se/
[5] http://ec.europa.eu/ and This work has been supported by the grant Euro-MatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003 of the Czech Republic)
EnTam is a sentence aligned English-Tamil bilingual corpus from some of the publicly available websites that we have collected for NLP research involving Tamil. The standard set of processing has been applied on the the raw web data before the data became available in sentence aligned English-Tamil parallel corpus suitable for various NLP tasks. The parallel corpus includes texts from bible, cinema and news domains.
ESIC (Europarl Simultaneous Interpreting Corpus) is a corpus of 370 speeches (10 hours) in English, with manual transcripts, transcribed simultaneous interpreting into Czech and German, and parallel translations.
The corpus contains source English videos and audios. The interpreters' voices are not published within the corpus, but there is a tool that downloads them from the web of European Parliament, where they are publicly avaiable.
The transcripts are equipped with metadata (disfluencies, mixing voices and languages, read or spontaneous speech, etc.), punctuated, and with word-level timestamps.
The speeches in the corpus come from the European Parliament plenary sessions, from the period 2008-11. Most of the speakers are MEP, both native and non-native speakers of English. The corpus contains metadata about the speakers (name, surname, id, fraction) and about the speech (date, topic, read or spontaneous).
The current version of ESIC is v1.0. It has validation and evaluation parts.
ESIC (Europarl Simultaneous Interpreting Corpus) is a corpus of 370 speeches (10 hours) in English, with manual transcripts, transcribed simultaneous interpreting into Czech and German, and parallel translations.
The corpus contains source English videos and audios. The interpreters' voices are not published within the corpus, but there is a tool that downloads them from the web of European Parliament, where they are publicly avaiable.
The transcripts are equipped with metadata (disfluencies, mixing voices and languages, read or spontaneous speech, etc.), punctuated, and with word-level timestamps.
The speeches in the corpus come from the European Parliament plenary sessions, from the period 2008-11. Most of the speakers are MEP, both native and non-native speakers of English. The corpus contains metadata about the speakers (name, surname, id, fraction) and about the speech (date, topic, read or spontaneous).
ESIC has validation and evaluation parts.
The current version is ESIC v1.1, it extends v1.0 with manual sentence alignment of the tri-parallel texts, and with bi-parallel sentence alignment of English original transcripts and German interpreting.
We have created test set for syntactic questions presented in the paper [1] which is more general than Mikolov's [2]. Since we were interested in morphosyntactic relations, we extended only the questions of the syntactic type with exception of nationality adjectives which is already covered completely in Mikolov's test set.
We constructed the pairs more or less manually, taking inspiration in the Czech side of the CzEng corpus [3], where explicit morphological annotation allows to identify various pairs of Czech words (different grades of adjectives, words and their negations, etc.). The word-aligned English words often shared the same properties. Another sources of pairs were acquired from various webpages usually written for learners of English. For example for verb tense, we relied on a freely available list of English verbs and their morphological variations.
We have included 100-1000 different pairs for each question set. The questions were constructed from the pairs similarly as by Mikolov: generating all possible pairs of pairs. This leads to millions of questions, so we randomly selected 1000 instances per question set, to keep the test set in the same order of magnitude. Additionally, we decided to extend set of questions on opposites to cover not only opposites of adjectives but also of nouns and verbs.
This package contains the eye-tracker recordings of 8 subjects evaluating English-to-Czech machine translation quality using the WMT-style ranking of sentences.
We provide the set of sentences evaluated, the exact screens presented to the annotators (including bounding box information for every area of interest and even for individual letters in the text) and finally the raw EyeLink II files with gaze trajectories.
The description of the experiment can be found in the paper:
Ondřej Bojar, Filip Děchtěrenko, Maria Zelenina. A Pilot Eye-Tracking Study of WMT-Style Ranking Evaluation.
Proceedings of the LREC 2016 Workshop “Translation Evaluation – From Fragmented Tools
and Data Sets to an Integrated Ecosystem”, Georg Rehm, Aljoscha Burchardt et al. (eds.). pp. 20-26. May 2016, Portorož, Slovenia.
This work has received funding from the European Union's Horizon 2020 research
and innovation programme under grant agreement no. 645452 (QT21). This work was
partially financially supported by the Government of Russian Federation, Grant
074-U01.
This work has been using language resources developed, stored and distributed
by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of
the Czech Republic (project LM2010013).
Data
-------
Hausa Visual Genome 1.0, a multimodal dataset consisting of text and images suitable for English-to-Hausa multimodal machine translation tasks and multimodal research. We follow the same selection of short English segments (captions) and the associated images from Visual Genome as the dataset Hindi Visual Genome 1.1 has. We automatically translated the English captions to Hausa and manually post-edited, taking the associated images into account.
The training set contains 29K segments. Further 1K and 1.6K segments are provided in development and test sets, respectively, which follow the same (random) sampling from the original Hindi Visual Genome.
Additionally, a challenge test set of 1400 segments is available for the multi-modal task. This challenge test set was created in Hindi Visual Genome by searching for (particularly) ambiguous English words based on the embedding similarity and manually selecting those where the image helps to resolve the ambiguity.
Dataset Formats
-----------------------
The multimodal dataset contains both text and images.
The text parts of the dataset (train and test sets) are in simple tab-delimited plain text files.
All the text files have seven columns as follows:
Column1 - image_id
Column2 - X
Column3 - Y
Column4 - Width
Column5 - Height
Column6 - English Text
Column7 - Hausa Text
The image part contains the full images with the corresponding image_id as the file name. The X, Y, Width, and Height columns indicate the rectangular region in the image described by the caption.
Data Statistics
--------------------
The statistics of the current release are given below.
Parallel Corpus Statistics
-----------------------------------
Dataset Segments English Words Hausa Words
---------- -------- ------------- -----------
Train 28930 143106 140981
Dev 998 4922 4857
Test 1595 7853 7736
Challenge Test 1400 8186 8752
---------- -------- ------------- -----------
Total 32923 164067 162326
The word counts are approximate, prior to tokenization.
Citation
-----------
If you use this corpus, please cite the following paper:
@InProceedings{abdulmumin-EtAl:2022:LREC,
author = {Abdulmumin, Idris
and Dash, Satya Ranjan
and Dawud, Musa Abdullahi
and Parida, Shantipriya
and Muhammad, Shamsuddeen
and Ahmad, Ibrahim Sa'id
and Panda, Subhadarshi
and Bojar, Ond{\v{r}}ej
and Galadanci, Bashir Shehu
and Bello, Bello Shehu},
title = "{Hausa Visual Genome: A Dataset for Multi-Modal English to Hausa Machine Translation}",
booktitle = {Proceedings of the Language Resources and Evaluation Conference},
month = {June},
year = {2022},
address = {Marseille, France},
publisher = {European Language Resources Association},
pages = {6471--6479},
url = {https://aclanthology.org/2022.lrec-1.694}
}
HindEnCorp parallel texts (sentence-aligned) come from the following sources:
Tides, which contains 50K sentence pairs taken mainly from news articles. This dataset was originally col- lected for the DARPA-TIDES surprise-language con- test in 2002, later refined at IIIT Hyderabad and provided for the NLP Tools Contest at ICON 2008 (Venkatapathy, 2008).
Commentaries by Daniel Pipes contain 322 articles in English written by a journalist Daniel Pipes and translated into Hindi.
EMILLE. This corpus (Baker et al., 2002) consists of three components: monolingual, parallel and annotated corpora. There are fourteen monolingual sub- corpora, including both written and (for some lan- guages) spoken data for fourteen South Asian lan- guages. The EMILLE monolingual corpora contain in total 92,799,000 words (including 2,627,000 words of transcribed spoken data for Bengali, Gujarati, Hindi, Punjabi and Urdu). The parallel corpus consists of 200,000 words of text in English and its accompanying translations into Hindi and other languages.
Smaller datasets as collected by Bojar et al. (2010) include the corpus used at ACL 2005 (a subcorpus of EMILLE), a corpus of named entities from Wikipedia (crawled in 2009), and Agriculture domain parallel corpus.

For the current release, we are extending the parallel corpus using these sources:
Intercorp (Čermák and Rosen,2012) is a large multilingual parallel corpus of 32 languages including Hindi. The central language used for alignment is Czech. Intercorp’s core texts amount to 202 million words. These core texts are most suitable for us because their sentence alignment is manually checked and therefore very reliable. They cover predominately short sto- ries and novels. There are seven Hindi texts in Inter- corp. Unfortunately, only for three of them the English translation is available; the other four are aligned only with Czech texts. The Hindi subcorpus of Intercorp contains 118,000 words in Hindi.
TED talks 3 held in various languages, primarily English, are equipped with transcripts and these are translated into 102 languages. There are 179 talks for which Hindi translation is available.
The Indic multi-parallel corpus (Birch et al., 2011; Post et al., 2012) is a corpus of texts from Wikipedia translated from the respective Indian language into English by non-expert translators hired over Mechanical Turk. The quality is thus somewhat mixed in many respects starting from typesetting and punctuation over capi- talization, spelling, word choice to sentence structure. A little bit of control could be in principle obtained from the fact that every input sentence was translated 4 times. We used the 2012 release of the corpus.
Launchpad.net is a software collaboration platform that hosts many open-source projects and facilitates also collaborative localization of the tools. We downloaded all revisions of all the hosted projects and extracted the localization (.po) files.
Other smaller datasets. This time, we added Wikipedia entities as crawled in 2013 (including any morphological variants of the named entitity that appears on the Hindi variant of the Wikipedia page) and words, word examples and quotes from the Shabdkosh online dictionary. and LM2010013,
Data
----
Hindi Visual Genome 1.0, a multimodal dataset consisting of text and images suitable for English-to-Hindi multimodal machine translation task and multimodal research. We have selected short English segments (captions) from Visual Genome along with associated images and automatically translated them to Hindi with manual post-editing, taking the associated images into account. The training set contains 29K segments. Further 1K and 1.6K segments are provided in a development and test sets, respectively, which follow the same (random) sampling from the original Hindi Visual Genome.
Additionally, a challenge test set of 1400 segments will be released for the WAT2019 multi-modal task. This challenge test set was created by searching for (particularly) ambiguous English words based on the embedding similarity and manually selecting those where the image helps to resolve the ambiguity.
Dataset Formats
--------------
The multimodal dataset contains both text and images.
The text parts of the dataset (train and test sets) are in simple tab-delimited plain text files.
All the text files have seven columns as follows:
Column1 - image_id
Column2 - X
Column3 - Y
Column4 - Width
Column5 - Height
Column6 - English Text
Column7 - Hindi Text
The image part contains the full images with the corresponding image_id as the file name. The X, Y, Width and Height columns indicate the rectangular region in the image described by the caption.
Data Statistics
----------------
The statistics of the current release is given below.
Parallel Corpus Statistics
---------------------------
Dataset Segments English Words Hindi Words
------- --------- ---------------- -------------
Train 28932 143178 136722
Dev 998 4922 4695
Test 1595 7852 7535
Challenge Test 1400 8185 8665 (Released separately)
------- --------- ---------------- -------------
Total 32925 164137 157617
The word counts are approximate, prior to tokenization.
Citation
--------
If you use this corpus, please cite the following paper:
@article{hindi-visual-genome:2019,
title={{Hindi Visual Genome: A Dataset for Multimodal English-to-Hindi Machine Translation}},
author={Parida, Shantipriya and Bojar, Ond{\v{r}}ej and Dash, Satya Ranjan},
journal={Computaci{\'o}n y Sistemas},
note={In print. Presented at CICLing 2019, La Rochelle, France},
year={2019},
}
Data
----
Hindi Visual Genome 1.1 is an updated version of Hindi Visual Genome 1.0. The update concerns primarily the text part of Hindi Visual Genome, fixing translation issues reported during WAT 2019 multimodal task. In the image part, only one segment and thus one image were removed from the dataset.
Hindi Visual Genome 1.1 serves in "WAT 2020 Multi-Modal Machine Translation Task".
Hindi Visual Genome is a multimodal dataset consisting of text and images suitable for English-to-Hindi multimodal machine translation task and multimodal research. We have selected short English segments (captions) from Visual Genome along with associated images and automatically translated them to Hindi with manual post-editing, taking the associated images into account.
The training set contains 29K segments. Further 1K and 1.6K segments are provided in a development and test sets, respectively, which follow the same (random) sampling from the original Hindi Visual Genome.
A third test set is called ``challenge test set'' consists of 1.4K segments and it was released for WAT2019 multi-modal task. The challenge test set was created by searching for (particularly) ambiguous English words based on the embedding similarity and manually selecting those where the image helps to resolve the ambiguity. The surrounding words in the sentence however also often include sufficient cues to identify the correct meaning of the ambiguous word.
Dataset Formats
--------------
The multimodal dataset contains both text and images.
The text parts of the dataset (train and test sets) are in simple
tab-delimited plain text files.
All the text files have seven columns as follows:
Column1 - image_id
Column2 - X
Column3 - Y
Column4 - Width
Column5 - Height
Column6 - English Text
Column7 - Hindi Text
The image part contains the full images with the corresponding image_id as the file name. The X, Y, Width and Height columns indicate the rectangular region in the image described by the caption.
Data Statistics
----------------
The statistics of the current release is given below.
Parallel Corpus Statistics
---------------------------
Dataset Segments English Words Hindi Words
------- --------- ---------------- -------------
Train 28930 143164 145448
Dev 998 4922 4978
Test 1595 7853 7852
Challenge Test 1400 8186 8639
------- --------- ---------------- -------------
Total 32923 164125 166917
The word counts are approximate, prior to tokenization.
Citation
--------
If you use this corpus, please cite the following paper:
@article{hindi-visual-genome:2019,
title={{Hindi Visual Genome: A Dataset for Multimodal English-to-Hindi Machine Translation}},
author={Parida, Shantipriya and Bojar, Ond{\v{r}}ej and Dash, Satya Ranjan},
journal={Computaci{\'o}n y Sistemas},
volume={23},
number={4},
pages={1499--1505},
year={2019}
}
A Hindi corpus of texts downloaded mostly from news sites. Contains both the original raw texts and an extensively cleaned-up and tokenized version suitable for language modeling. 18M sentences, 308M tokens and FP7-ICT-2007-3-231720 (EuroMatrix Plus), 7E09003 (Czech part of EM+)
Hindi monolingual corpus. It is based primarily on web crawls performed using various tools and at various times. Since the web is a living data source, we treat these crawls as completely separate sources, despite they may overlap. To estimate the magnitude of this overlap, we compared the total number of segments if we concatenate the individual sources (each source being deduplicated on its own) with the number of segments if we de-duplicate all sources to- gether. The difference is just around 1%, confirming, that various web crawls (or their subsequent processings) differ significantly.
HindMonoCorp contains data from:
Hindi web texts, a monolingual corpus containing mainly Hindi news articles has already been collected and released by Bojar et al. (2008). We use the HTML files as crawled for this corpus in 2010 and we add a small crawl performed in 2013 and re-process them with the current pipeline. These sources are denoted HWT 2010 and HWT 2013 in the following.
Hindi corpora in W2C have been collected by Martin Majliš during his project to automatically collect corpora in many languages (Majliš and Žabokrtský, 2012). There are in fact two corpora of Hindi available—one from web harvest (W2C Web) and one from the Wikipedia (W2C Wiki).
SpiderLing is a web crawl carried out during November and December 2013 using SpiderLing (Suchomel and Pomikálek, 2012). The pipeline includes extraction of plain texts and deduplication at the level of documents, see below.
CommonCrawl is a non-profit organization that regu- larly crawls the web and provides anyone with the data. We are grateful to Christian Buck for extracting plain text Hindi segments from the 2012 and 2013-fall crawls for us.
Intercorp – 7 books with their translations scanned and manually alligned per paragraph
RSS Feeds from Webdunia.com and the Hindi version of BBC International followed by our custom crawler from September 2013 till January 2014. and LM2010013,
We present a large corpus of Czech parliament plenary sessions. The corpus
consists of approximately 444 hours of speech data and corresponding text
transcriptions. The whole corpus has been segmented to short audio snippets
making it suitable for both training and evaluation of automatic speech
recognition (ASR) systems. The source language of the corpus is Czech, which
makes it a valuable resource for future research as only a few public datasets
are available for the Czech language.
Data
-------
Malayalam Visual Genome (MVG for short) 1.0 has similar goals as Hindi Visual Genome (HVG) 1.1: to support the Malayalam language. Malayalam Visual Genome 1.0 is the first multi-modal dataset in Malayalam for machine translation and image captioning.
Malayalam Visual Genome 1.0 serves in "WAT 2021 Multi-Modal Machine Translation Task".
Malayalam Visual Genome is a multimodal dataset consisting of text and images suitable for English-to-Malayalam multimodal machine translation task and multimodal research. We follow the same selection of short English segments (captions) and the associated images from Visual Genome as HGV 1.1 has. For MVG, we automatically translated these captions from English to Malayalam and manually corrected them, taking the associated images into account.
The training set contains 29K segments. Further 1K and 1.6K segments are provided in development and test sets, respectively, which follow the same (random) sampling from the original Hindi Visual Genome.
A third test set is called ``challenge test set'' and consists of 1.4K segments. The challenge test set was created for the WAT2019 multi-modal task by searching for (particularly) ambiguous English words based on the embedding similarity and manually selecting those where the image helps to resolve the ambiguity. The surrounding words in the sentence however also often include sufficient cues to identify the correct meaning of the ambiguous word. For MVG, we simply translated the English side of the test sets to Malayalam, again utilizing machine translation to speed up the process.
Dataset Formats
----------------------
The multimodal dataset contains both text and images.
The text parts of the dataset (train and test sets) are in simple tab-delimited plain text files.
All the text files have seven columns as follows:
Column1 - image_id
Column2 - X
Column3 - Y
Column4 - Width
Column5 - Height
Column6 - English Text
Column7 - Malayalam Text
The image part contains the full images with the corresponding image_id as the file name. The X, Y, Width and Height columns indicate the rectangular region in the image described by the caption.
Data Statistics
-------------------
The statistics of the current release are given below.
Parallel Corpus Statistics
---------------------------------
Dataset Segments English Words Malayalam Words
---------- -------------- -------------------- -----------------
Train 28930 143112 107126
Dev 998 4922 3619
Test 1595 7853 5689
Challenge Test 1400 8186 6044
-------------------- ------------ ------------------ ------------------
Total 32923 164073 122478
The word counts are approximate, prior to tokenization.
Citation
-----------
If you use this corpus, please cite the following paper:
@article{hindi-visual-genome:2019, title={{Hindi Visual Genome: A Dataset for Multimodal English-to-Hindi Machine Translation}}, author={Parida, Shantipriya and Bojar, Ond{\v{r}}ej and Dash, Satya Ranjan}, journal={Computaci{\'o}n y Sistemas}, volume={23}, number={4}, pages={1499--1505}, year={2019} }
Manual classification of errors of Czech-Slovak translation according to the classification introduced by Vilar et al. [1]. First 50 sentences from WMT 2010 test set were translated by 5 MT systems (Česílko, Česílko2, Google Translate and two Moses setups) and MT errors were manually marked and classified. Classification was applied in MT systems comparison [3]. Reference translation is included.
References:
[1] David Vilar, Jia Xu, Luis Fernando D’Haro and Hermann Ney. Error Analysis of Machine Translation Output. In International Conference on Language Resources and Evaluation, pages 697-702. Genoa, Italy, May 2006.
[2] http://matrix.statmt.org/test_sets/list
[3] Ondřej Bojar, Petra Galuščáková, and Miroslav Týnovský. Evaluating Quality of Machine Translation from Czech to Slovak. In Markéta Lopatková, editor, Information Technologies - Applications and Theory, pages 3-9, September 2011 and This work has been supported by the grants Euro-MatrixPlus (FP7-ICT-2007-3-231720 of the EU and
7E09003 of the Czech Republic)
Manual classification of errors of English-Slovak translation according to the classification introduced by Vilar et al. [1]. 50 sentences randomly selected from WMT 2011 test set [2] were translated by 3 MT systems described in [3] and MT errors were manually marked and classified. Reference translation is included.
References:
[1] David Vilar, Jia Xu, Luis Fernando D’Haro and Hermann Ney. Error Analysis of Machine Translation Output. In International Conference on Language Resources and Evaluation, pages 697-702. Genoa, Italy, May 2006.
[2] http://www.statmt.org/wmt11/evaluation-task.html
[3] Petra Galuščáková and Ondřej Bojar. Improving SMT by Using Parallel Data of a Closely Related Language. In Human Language Technologies - The Baltic Perspective - Proceedings of the Fifth International Conference Baltic HLT 2012, volume 247 of Frontiers in AI and Applications, pages 58-65, Amsterdam, Netherlands, October 2012. IOS Press. and This work has been supported by the grant Euro-MatrixPlus (FP7-ICT-2007-3-231720 of the EU and
7E09003 of the Czech Republic)
Manually ranked outputs of Czech-Slovak translations. Three annotators manually ranked outputs of five MT systems (Česílko, Česílko2, Google Translate and two Moses setups) on three data sets (100 sentences randomly selected from books, 100 sentences randomly selected from Acquis corpus and 50 first sentences from WMT 2010 test set). Ranking was applied in MT systems comparison in [1].
References:
[1] Ondřej Bojar, Petra Galuščáková, and Miroslav Týnovský. Evaluating Quality of Machine Translation from Czech to Slovak. In Markéta Lopatková, editor, Information Technologies - Applications and Theory, pages 3-9, September 2011 and This work has been supported by the grant Euro-MatrixPlus (FP7-ICT-2007-3-231720 of the EU and
7E09003 of the Czech Republic)
Data
----
We have collected English-Odia parallel and monolingual data from the
available public websites for NLP research in Odia.
The parallel corpus consists of English-Odia parallel Bible, Odia
digital library, and Odisha Goverment websites. It covers bible,
literature, goverment of Odisha and its policies. We have processed the
raw data collected from the websites, performed alignments (a mix of
manual and automatic alignments) and release the corpus in a form ready
for various NLP tasks.
The Odia monolingual data consists of Odia-Wikipedia and Odia e-magazine
websites. Because the major portion of data is extracted from
Odia-Wikipedia, it covers all kinds of domains. The e-magazines data
mostly cover the literature domain. We have preprocessed the monolingual
data including de-duplication, text normalization, and sentence
segmentation to make it ready for various NLP tasks.
Corpus Formats
--------------
Both corpora are in simple tab-delimited plain text files.
The parallel corpus files have three columns:
- the original book/source of the sentence pair
- the English sentence
- the corresponding Odia sentence
The monolingual corpus has a varying number of columns:
- each line corresponds to one *paragraph* (or related unit) of the
original source
- each tab-delimited unit corresponds to one *sentence* in the paragraph
Data Statistics
----------------
The statistics of the current release is given below.
Parallel Corpus Statistics
---------------------------
Dataset Sentences #English tokens #Odia tokens
------- --------- ---------------- -------------
Train 27136 706567 604147
Dev 948 21912 19513
Test 1262 28488 24365
------- --------- ---------------- -------------
Total 29346 756967 648025
Domain Level Statistics
------------------------
Domain Sentences #English tokens #Odia tokens
------------------ --------- ---------------- -------------
Bible 29069 756861 640157
Literature 424 7977 6611
Goverment policies 204 1411 1257
------------------ --------- ---------------- -------------
Total 29697 766249 648025
Monolingual Corpus Statistics
-----------------------------
Paragraphs Sentences #Odia tokens
---------- --------- ------------
71698 221546 2641308
Domain Level Statistics
-----------------------
Domain Paragraphs Sentences #Odia tokens
-------------- -------------- --------- -------------
General (wiki) 30468 (42.49%) 102085 1320367
Literature 41230 (57.50%) 119461 1320941
-------------- -------------- --------- -------------
Total 71698 221546 2641308
Citation
--------
If you use this corpus, please cite it directly (see above), but please cite also the following paper:
Title: OdiEnCorp: Odia-English and Odia-Only Corpus for Machine Translation
Author: Shantipriya Parida, Ondrej Bojar, and Satya Ranjan Dash
Proceedings of the Third International Conference on Smart Computing & Informatics (SCI) 2018
Series: Smart Innovation, Systems and Technologies (SIST)
Publisher: Springer Singapore
Data
-----
We have collected English-Odia parallel data for the purposes of NLP
research of the Odia language.
The data for the parallel corpus was extracted from existing parallel
corpora such as OdiEnCorp 1.0 and PMIndia, and books which contain both
English and Odia text such as grammar and bilingual literature books. We
also included parallel text from multiple public websites such as Odia
Wikipedia, Odia digital library, and Odisha Government websites.
The parallel corpus covers many domains: the Bible, other literature,
Wiki data relating to many topics, Government policies, and general
conversation. We have processed the raw data collected from the books,
websites, performed sentence alignments (a mix of manual and automatic
alignments) and released the corpus in a form suitable for various NLP
tasks.
Corpus Format
-------------
OdiEnCorp 2.0 is stored in simple tab-delimited plain text files, each
with three tab-delimited columns:
- a coarse indication of the domain
- the English sentence
- the corresponding Odia sentence
The corpus is shuffled at the level of sentence pairs.
The coarse domains are:
books ... prose text
dict ... dictionaries and phrasebooks
govt ... partially formal text
odiencorp10 ... OdiEnCorp 1.0 (mix of domains)
pmindia ... PMIndia (the original corpus)
wikipedia ... sentences and phrases from Wikipedia
Data Statistics
---------------
The statistics of the current release are given below.
Note that the statistics differ from those reported in the paper due to
deduplication at the level of sentence pairs. The deduplication was
performed within each of the dev set, test set and training set and
taking the coarse domain indication into account. It is still possible
that the same sentence pair appears more than once within the same set
(dev/test/train) if it came from different domains, and it is also
possible that a sentence pair appears in several sets (dev/test/train).
Parallel Corpus Statistics
--------------------------
Dev Dev Dev Test Test Test Train Train Train
Sents # EN # OD Sents # EN # OD Sents # EN # OD
books 3523 42011 36723 3895 52808 45383 3129 40461 35300
dict 3342 14580 13838 3437 14807 14110 5900 21591 20246
govt - - - - - - 761 15227 13132
odiencorp10 947 21905 19509 1259 28473 24350 26963 704114 602005
pmindia 3836 70282 61099 3836 68695 59876 30687 551657 486636
wikipedia 1896 9388 9385 1917 21381 20951 1930 7087 7122
Total 13544 158166 140554 14344 186164 164670 69370 1340137 1164441
"Sents" are the counts of the sentence pairs in the given set (dev/test/train)
and domain (books/dict/...).
"# EN" and "# OD" are approximate counts of words (simply space-delimited,
without tokenization) in English and Odia
The total number of sentence pairs (lines) is 13544+14344+69370=97258. Ignoring
the set and domain and deduplicating again, this number drops to 94857.
Citation
--------
If you use this corpus, please cite the following paper:
@inproceedings{parida2020odiencorp,
title={OdiEnCorp 2.0: Odia-English Parallel Corpus for Machine Translation},
author={Parida, Shantipriya and Dash, Satya Ranjan and Bojar, Ond{\v{r}}ej and Motlicek, Petr and Pattnaik, Priyanka and Mallick, Debasish Kumar},
booktitle={Proceedings of the WILDRE5--5th Workshop on Indian Language Data: Resources and Evaluation},
pages={14--19},
year={2020}
}
We define "optimal reference translation" as a translation thought to be the best possible that can be achieved by a team of human translators. Optimal reference translations can be used in assessments of excellent machine translations.
We selected 50 documents (online news articles, with 579 paragraphs in total) from the 130 English documents included in the WMT2020 news test (http://www.statmt.org/wmt20/) with the aim to preserve diversity (style, genre etc.) of the selection. In addition to the official Czech reference translation provided by the WMT organizers (P1), we hired two additional translators (P2 and P3, native Czech speakers) via a professional translation agency, resulting in three independent translations. The main contribution of this dataset are two additional translations (i.e. optimal reference translations N1 and N2), done jointly by two translators-cum-theoreticians with an extreme care for various aspects of translation quality, while taking into account the translations P1-P3. We publish also internal comments (in Czech) for some of the segments.
Translation N1 should be closer to the English original (with regards to the meaning and linguistic structure) and female surnames use the Czech feminine suffix (e.g. "Mai" is translated as "Maiová"). Translation N2 is more free, trying to be more creative, idiomatic and entertaining for the readers and following the typical style used in Czech media, while still preserving the rules of functional equivalence. Translation N2 is missing for the segments where it was not deemed necessary to provide two alternative translations. For applications/analyses needing translation of all segments, this should be interpreted as if N2 is the same as N1 for a given segment.
We provide the dataset in two formats: OpenDocument spreadsheet (odt) and plain text (one file for each translation and the English original). Some words were highlighted using different colors during the creation of optimal reference translations; this highlighting and comments are present only in the odt format (some comments refer to row numbers in the odt file). Documents are separated by empty lines and each document starts with a special line containing the document name (e.g. "# upi.205735"), which allows alignment with the original WMT2020 news test. For the segments where N2 translations are missing in the odt format, the respective N1 segments are used instead in the plain-text format.
This corpus contains annotations of translation quality from English to Czech in seven categories on both segment- and document-level. There are 20 documents in total, each with 4 translations (evaluated by each annotator in paralel) of 8 segments (can be longer than one sentence). Apart from the evaluation, the annotators also proposed their own, improved versions of the translations.
There were 11 annotators in total, on expertise levels ranging from non-experts to professional translators.
The ParCzech 3.0 corpus is the third version of ParCzech consisting of stenographic protocols that record the Chamber of Deputies’ meetings held in the 7th term (2013-2017) and the current 8th term (2017-Mar 2021). The protocols are provided in their original HTML format, Parla-CLARIN TEI format, and the format suitable for Automatic Speech Recognition. The corpus is automatically enriched with the morphological, syntactic, and named-entity annotations using the procedures UDPipe 2 and NameTag 2. The audio files are aligned with the texts in the annotated TEI files.
Statistical component of Chimera, a state-of-the-art MT system. and Project DF12P01OVV022 of the Ministry of Culture of the Czech Republic (NAKI -- Amalach).
The dataset used for the Ptakopět experiment on outbound machine translation. It consists of screenshots of web forms with user queries entered. The queries are available also in a text form. The dataset comprises two language versions: English and Czech. Whereas the English version has been fully post-processed (screenshots cropped, queries within the screenshots highlighted, dataset split based on its quality etc.), the Czech version is raw as it was collected by the annotators.
The THEaiTRobot 1.0 tool allows the user to interactively generate scripts for individual theatre play scenes.
The tool is based on GPT-2 XL generative language model, using the model without any fine-tuning, as we found that with a prompt formatted as a part of a theatre play script, the model usually generates continuation that retains the format.
We encountered numerous problems when generating the script in this way. We managed to tackle some of the problems with various adjustments, but some of them remain to be solved in a future version.
THEaiTRobot 1.0 was used to generate the first THEaiTRE play, "AI: Když robot píše hru" ("AI: When a robot writes a play").
The THEaiTRobot 2.0 tool allows the user to interactively generate scripts for individual theatre play scenes.
The previous version of the tool (http://hdl.handle.net/11234/1-3507) was based on GPT-2 XL generative language model, using the model without any fine-tuning, as we found that with a prompt formatted as a part of a theatre play script, the model usually generates continuation that retains the format.
The current version also uses vanilla GPT-2 by default, but can also instead use a GPT-2 medium model fine-tuned on theatre play scripts (as well as film and TV series scripts). Apart from the basic "flat" generation using a theatrical starting prompt and the script model, the tool also features a second, hierarchical variant, where in the first step, a play synopsis is generated from its title using a synopsis model (GPT-2 medium fine-tuned on synopses of theatre plays, as well as film, TV series and book synopses). The synopsis is then used as input for the second stage, which uses the script model.
The choice of models to use is done by setting the MODEL variable in start_server.sh and start_syn_server.sh
THEaiTRobot 2.0 was used to generate the second THEaiTRE play, "Permeation/Prostoupení".
AMALACH project component TMODS:ENG-CZE; machine translation of queries from Czech to English. This archive contains models for the Moses decoder (binarized, pruned to allow for real-time translation) and configuration files for the MTMonkey toolkit. The aim of this package is to provide a full service for Czech->English translation which can be easily utilized as a component in a larger software solution. (The required tools are freely available and an installation guide is included in the package.)
The translation models were trained on CzEng 1.0 corpus and Europarl. Monolingual data for LM estimation additionally contains WMT news crawls until 2013.
UMC 0.1 Czech-English-Russian is a multilingual parallel corpus of texts in Czech, Russian and English languages with automatic pairwise sentence alignments. The primary aim of UMC is to extend the set of languages covered by the corpus CzEng mainly for the purposes of machine translation.
All the texts were downloaded from a single source — The Project Syndicate (Copyright: Project Syndicate 1995-2008), which contains a huge collection of high-quality news articles and commentaries. We were given the permission to use the texts for research and non-commercial purposes. and FP6-IST-5-034291-STP (EuroMatrix)
We release a sizeable monolingual Urdu corpus automatically tagged with part-of-speech tags. We extend the work of Jawaid and Bojar (2012) who use three different taggers and then apply a voting scheme to disambiguate among the different choices suggested by each tagger. We run this complex ensemble on a large monolingual corpus and release the both plain and tagged corpora. and it is supported by the MosesCore project sponsored by the European Commission’s Seventh Framework Programme (Grant Number 288487).
We provide the Vietnamese version of the multi-lingual test set from WMT 2013 [1] competition. The Vietnamese version was manually translated from English. For completeness, this record contains the 3000 sentences in all the WMT 2013 original languages (Czech, English, French, German, Russian and Spanish), extended with our Vietnamese version. Test set is used in [2] to evaluate translation between Czech, English and Vietnamese.
References
1. http://www.statmt.org/wmt13/evaluation-task.html
2. Duc Tam Hoang and Ondřej Bojar, The Prague Bulletin of Mathematical Linguistics. Volume 104, Issue 1, Pages 75--86, ISSN 1804-0462. 9/2015
Testing set from WMT 2011 [1] competition, manually translated from Czech and English into Slovak. Test set contains 3003 sentences in Czech, Slovak and English. Test set is described in [2].
References:
[1] http://www.statmt.org/wmt11/evaluation-task.html
[2] Petra Galuščáková and Ondřej Bojar. Improving SMT by Using Parallel Data of a Closely Related Language. In Human Language Technologies - The Baltic Perspective - Proceedings of the Fifth International Conference Baltic HLT 2012, volume 247 of Frontiers in AI and Applications, pages 58-65, Amsterdam, Netherlands, October 2012. IOS Press. and The work on this project was supported by the grant EuroMatrixPlus (FP7-ICT-
2007-3-231720 of the EU and 7E09003 of the Czech Republic)
The item contains models to tune for the WMT16 Tuning shared task for Czech-to-English.
CzEng 1.6pre (http://ufal.mff.cuni.cz/czeng/czeng16pre) corpus is used for the training of the translation models. The data is tokenized (using Moses tokenizer), lowercased and sentences longer than 60 words and shorter than 4 words are removed before training. Alignment is done using fast_align (https://github.com/clab/fast_align) and the standard Moses pipeline is used for training.
Two 5-gram language models are trained using KenLM: one only using the CzEng English data and the other is trained using all available English mono data for WMT except Common Crawl.
Also included are two lexicalized bidirectional reordering models, word based and hierarchical, with msd conditioned on both source and target of processed CzEng.
This item contains models to tune for the WMT16 Tuning shared task for English-to-Czech.
CzEng 1.6pre (http://ufal.mff.cuni.cz/czeng/czeng16pre) corpus is used for the training of the translation models. The data is tokenized (using Moses tokenizer), lowercased and sentences longer than 60 words and shorter than 4 words are removed before training. Alignment is done using fast_align (https://github.com/clab/fast_align) and the standard Moses pipeline is used for training.
Two 5-gram language models are trained using KenLM: one only using the CzEng Czech data and the other is trained using all available Czech mono data for WMT except Common Crawl.
Also included are two lexicalized bidirectional reordering models, word based and hierarchical, with msd conditioned on both source and target of processed CzEng.