LDC January 2023 Newsletter

Renew your LDC membership today 

30th Anniversary Highlight: CSR  

New publications:

AIDA Ukrainian Broadcast and Telephone Speech Audio and Transcripts

LORELEI Swahili Representative Language Pack

_______________________________________________________________________

Renew your LDC membership today 
The importance of curated resources for language-related education, research and technology development drives LDC’s mission to create them, to accept data contributions from researchers across the globe and to broadly share such resources through the LDC Catalog. LDC members enjoy no-cost access to new corpora released annually, as well as the ability to license legacy data sets from among our 925+ holdings at reduced fees. Ensure that your data needs continue to be met by renewing your LDC membership or by joining the Consortium today.

Now through March 1, 2023, 2022 members receive a 10% discount on 2023 membership, and new or returning organizations receive a 5% discount. Membership remains the most economical way to access current and past LDC releases. Consult Join LDC for more details on membership options and benefits. 

30th Anniversary Highlight: CSR  
The CSR (continuous speech recognition) corpus series was developed in the early 1990s under DARPA’s Spoken Language Program to support research on large-vocabulary CSR systems. 

CSR-I (WSJ0) Complete (LDC93S6A) and CSR-II (WSJ1) Complete (LDC94S13A) contain speech from a machine-readable corpus of Wall Street Journal news text. They also include spontaneous dictation by journalists of hypothetical news articles as well as transcripts.

The text in CSR-I (WSJ0) was selected to fall within either a 5,000-word subset or a 20,000-word subset. Audio includes speaker-dependent and speaker-independent sections as well as sentences with verbalized and nonverbalized punctuation. (Doddington, 1992). CSR-II features “Hub and Spoke” test sets that include a 5,000-word subset and a 64,000-word subset. Both data sets were collected using two microphones – a close-talking Sennheiser HMD414 and a second microphone of varying type. 

WSJ0 Cambridge Read News (LDC95S24) was developed by Cambridge University and consists of native British English speakers reading CSR WSJ news text, specifically, sentences from the 5,000-word and 64,000-word subsets. All speakers also recorded a common set of 18 adaptation sentences.  

The CSR corpora continue to have value for the research community. CSR-I (WSJ0) target utterances were used in the CHiME2 and CHiME3 challenges which focused on distant-microphone automatic speech recognition in real-world environments. CHiME2 WSJ0 (LDC2017S10) and CHiME2 Grid (LDC2017S07) each contain over 120 hours of English speech from a noisy living room environment. CHiME3 (LDC2017S24) consists of 342 hours of English speech and transcripts from noisy environments and 50 hours of noisy environment audio. 

CSR-I target utterances were also used in the Distant-Speech Interaction for Robust Home Applications (DIRHA) Project which addressed natural spontaneous speech interaction with distant microphones in a domestic environment. DIRHA English WSJ Audio (LDC2018S01) is comprised of approximately 85 hours of real and simulated read speech from native American English speakers in an apartment setting with typical domestic background noises and inter/intra-room reverberation effects.

Multi-Channel WSJ Audio (LDC2014S03), designed to address the challenges of speech recognition in meetings, contains 100 hours of audio from British English speakers reading sentences from WSJ0 Cambridge Read News. There were three recording scenarios: a single stationary speaker, two stationary overlapping speakers, and one single moving speaker. 

All CSR corpora and their related data sets are available for licensing by Consortium members and non-members. Visit Obtaining Data for more information.

New publications:
 
AIDA Ukrainian Broadcast and Telephone Speech Audio and Transcripts and is comprised of approximately 156 hours of Ukrainian conversational telephone speech and broadcast news audio with 1.2 million words of corresponding orthographic transcripts. 
 
The news audio data was taken from 87 recordings broadcast by various Ukrainian sources. The telephone speech was generated from telephone calls by native Ukrainian speakers to acquaintances in their social network. Native Ukrainian speakers manually segmented the data into sentence-level units as part of the transcription process.
 
The broadcast recordings and transcripts were produced by LDC to support the DARPA AIDA (Active Interpretation of Disparate Alternatives) program which aimed to develop a multi-hypothesis semantic engine to generate explicit alternative interpretations of events, situations and trends from a variety of unstructured sources. The telephone speech audio recordings were collected by LDC to support the NIST 2011 Language Recognition Evaluation  and are also contained in Multi-Language Conversational Telephone Speech 2011 – Slavic Group LDC2016S11.

2023 members can access this corpus through their LDC accounts. Non-members may license this data for a fee.

*

LORELEI Swahili Representative Language Pack was developed by LDC and is comprised of approximately 4.3 million words of Swahili monolingual text, 90,000 Swahili words translated from English data, and 545,000 words of found Swahili-English parallel text. Approximately 100,000 words were annotated for named entities and up to 26,000 words were annotated for entity discovery and linking and situation frames (identifying entities, needs and issues). Data was collected from discussion forum, news, reference, social network, and weblogs.

The LORELEI (Low Resource Languages for Emergent Incidents) program was concerned with building human language technology for low resource languages in the context of emergent situations. Representative languages were selected to provide broad typological coverage.

The knowledge base for entity linking annotation is available separately as LORELEI Entity Detection and Linking Knowledge Base (LDC2020T10).

2023 members can access this corpus through their LDC accounts. Non-members may license this data for a fee. 

 

LDC June 2017 Newsletter

New publications:

Abstract Meaning Representation (AMR) Annotation Release 2.0

CHiME2 WSJ0

UCLA High-Speed Laryngeal Video and Audio

______________________________________________________

(1) Abstract Meaning Representation (AMR) Annotation Release 2.0 was developed by LDC, SDL/Language Weaver, Inc., the University of Colorado's Computational Language and Educational Research group and the Information Sciences Institute at the University of Southern California. It contains a sembank (semantic treebank) of over 39,260 English natural language sentences from broadcast conversations, newswire, weblogs and web discussion forums.

AMR captures “who is doing what to whom” in a sentence. Each sentence is paired with a graph that represents its whole-sentence meaning in a tree-structure. AMR utilizes PropBank frames, non-core semantic roles, within-sentence coreference, named entity annotation, modality, negation, questions, quantities, and so on to represent the semantic structure of a sentence largely independent of its syntax.

LDC also released Abstract Meaning Representation (AMR) Annotation Release 1.0 (LDC2014T12).

Abstract Meaning Representation (AMR) Annotation Release 2.0 is distributed via web download.

2017 Subscription Members will automatically receive copies of this corpus. 2017 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for a fee.

*

(2) CHiME2 WSJ0 was developed as part of The 2nd CHiME Speech Separation and Recognition Challenge and contains approximately 166 hours of English speech from a noisy living room environment. The CHiME Challenges focus on distant-microphone automatic speech recognition (ASR) in real-world environments.

CHiME2 WSJ0 reflects the medium vocabulary track of the CHiME2 Challenge. The target utterances were taken from CSR-I (WSJ0) Complete (LDC93S6A), specifically, the 5,000 word subset of read speech from Wall Street Journal news text. Data is divided into training, development and test sets and includes baseline scoring, decoding and retraining tools. 

CHiME2 WSJ0 is distributed via web download.

2017 Subscription Members will automatically receive copies of this corpus. 2017 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for a fee.

*

(3) UCLA High-Speed Laryngeal Video and Audio was developed by UCLA Speech Processing and Auditory Perception Laboratory and is comprised of high-speed laryngeal video recordings of the vocal folds and synchronized audio recordings form nine subjects collected between April 2012 and April 2013. Speakers were asked to sustain the vowel /i/ for approximately ten seconds while holding voice quality, fundamental frequency, and loudness as steady as possible.

In the field of speech production theory, data such as contained in this release may be used to study the relationship between vocal folds vibration and resulting voice quality.

None of the subjects had a history of a voice disorder. There was no native language requirement for recruiting subjects; participants were native speakers of various languages, including English, Mandarin Chinese, Taiwanese Mandarin, Cantonese and German.

UCLA High-Speed Laryngeal Video and Audio is distributed via hard drive.

2017 Subscription Members will automatically receive copies of this corpus. 2017 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for a fee. 

LDC April 2014 Newsletter

New publications:

Domain-Specific Hyponym Relations
GALE Arabic-English Parallel Aligned Treebank -- WebTraining
Multi-Channel WSJ Audio 

---

(1) Domain-Specific Hyponym Relations was developed by the Shaanxi Province Key Laboratory of Satellite and Terrestrial Network Technology at Xi’an Jiaotung University, Xi’an, Shaanxi, China. It provides more than 5,000 English hyponym relations in five domains including data mining, computer networks, data structures, Euclidean geometry and microbiology. All hypernym and hyponym words were taken from Wikipedia article titles.

A hyponym relation is a word sense relation that is an IS-A relation. For example, dog is a hyponym of animal and binary tree is a hyponym of tree structure. Among the applications for domain-specific hyponym relations are taxonomy and ontology learning, query result organization in a faceted search and knowledge organization and automated reasoning in knowledge-rich applications.

The data is presented in XML format, and each file provides hyponym relations in one domain. Within each file, the term, Wikipedia URL, hyponym relation and the names of the hyponym and hypernym words are included. The distribution of terms and relations is set forth in the table below:

Dataset	Terms	Hyponym Relations
Data Mining	278	364
Computer Network	336	399
Data Structure	315	578
Euclidean Geometry	455	690
Microbiology	1,028	3,533

Domain-Specific Hyponym Relations is distributed via web download.

2014 Subscription Members will automatically receive two copies of this data on disc. 2014 Standard Members may request a copy as part of their 16 free membership corpora.  This data is made available at no-cost to LDC members and non-members under the Creative Commons Attribution-Noncommercial Share Alike 3.0 license. 

*

(2) GALE Arabic-English Parallel Aligned Treebank -- Web Training was developed by LDC and contains 69,766 tokens of word aligned Arabic and English parallel text with treebank annotations. This material was used as training data in the DARPA GALE (Global Autonomous Language Exploitation) program.

Parallel aligned treebanks are treebanks annotated with morphological and syntactic structures aligned at the sentence level and the sub-sentence level. Such data sets are useful for natural language processing and related fields, including automatic word alignment system training and evaluation, transfer-rule extraction, word sense disambiguation, translation lexicon extraction and cultural heritage and cross-linguistic studies. With respect to machine translation system development, parallel aligned treebanks may improve system performance with enhanced syntactic parsers, better rules and knowledge about language pairs and reduced word error rate.

In this release, the source Arabic data was translated into English. Arabic and English treebank annotations were performed independently. The parallel texts were then word aligned.

LDC previously released Arabic-English Parallel Aligned Treebanks as follows:

• Newswire
• Broadcast News Part 1
• Broadcast News Part 2

This release consists of Arabic source web data (newsgroups, weblogs) collected by LDC in 2004 and 2005. All data is encoded as UTF-8. A count of files, words, tokens and segments is below.

Language	Files	Words	Tokens	Segments
Arabic	162	46,710	69,766	3,178

Note: Word count is based on the untokenized Arabic source, token count is based on the ATB-tokenized Arabic source.

The purpose of the GALE word alignment task was to find correspondences between words, phrases or groups of words in a set of parallel texts. Arabic-English word alignment annotation consisted of the following tasks:

• Identifying different types of links: translated (correct or incorrect) and not translated (correct or incorrect)
• Identifying sentence segments not suitable for annotation, e.g., blank segments, incorrectly-segmented segments, segments with foreign languages
• Tagging unmatched words attached to other words or phrases

GALE Arabic-English Parallel Aligned Treebank -- Web Training is distributed via web download.

2014 Subscription Members will automatically receive two copies of this data on disc. 2014 Standard Members may request a copy as part of their 16 free membership corpora.  Non-members may license this data for a fee.

*

(3) Multi-Channel WSJ Audio (MCWSJ) was developed by the Centre for Speech Technology Research at the University of Edinburgh and contains approximately 100 hours of recorded speech from 45 British English speakers. Participants read Wall Street Journal texts published in 1987-1989 in three recording scenarios: a single stationary speaker, two stationary overlapping speakers and one single moving speaker.

This corpus was designed to address the challenges of speech recognition in meetings, which often occur in rooms with non-ideal acoustic conditions and significant background noise, and may contain large sections of overlapping speech. Using headset microphones represents one approach, but meeting participants may be reluctant to wear them. Microphone arrays are another option. MCWSJ supports research in large vocabulary tasks using microphone arrays. The news sentences read by speakers are taken from WSJCAM0 Cambridge Read News, a corpus originally developed for large vocabulary continuous speech recognition experiments, which in turn was based on CSR-I (WSJ0) Complete, made available by LDC to support large vocabulary continuous speech recognition initiatives.

Speakers reading news text from prompts were recorded using a headset microphone, a lapel microphone and an eight-channel microphone array. In the single speaker scenario, participants read from six fixed positions. Fixed positions were assigned for the entire recording in the overlapping scenario. For the moving scenario, participants moved from one position to the next while reading.

Fifteen speakers were recorded for the single scenario, nine pairs for the overlapping scenario and nine individuals for the moving scenario. Each read approximately 90 sentences.

Multi-Channel WSJ Audio is distributed on 2 DVD-ROM.

2014 Subscription Members will receive a copy of this data provided that they have completed the User License Agreement for Multi-Channel WSJ Audio LDC2014S03. 2014 Standard Members may request a copy as part of their 16 free membership corpora.  Non-members may license this data for a fee.