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).
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},
}
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)