README FILE FOR LDC CATALOG ID: LDC2020T11
TITLE: LORELEI Oromo Incident Language Pack
AUTHORS: Jennifer Tracey, Dave Graff, Stephanie Strassel, Michael Arrigo,
Jonathan Wright, Ann Bies
1.0 Introduction
This corpus contains all the text data, annotations and supplemental resources
for the Oromo language that were used in the DARPA LORELEI / LoReHLT 2017
Evaluation, which was conducted by NIST in August of that year.
Detailed information about the corpus content is provided in section 3
for each of the partitions ("sets") in the corpus. Combining all sets,
the corpus contains approximately 3.9 million words of monoligual text
in Oromo, 25,000 words of monolingual text in English, 135,000 words
of parallel and comparable Oromo-English text, and 50,000 words of
data annotated for Entity Discovery and Linking and Situation Frames.
The LORELEI (Low Resource Languages for Emergent Incidents) Program is
concerned with building Human Language Technology for low resource
languages in the context of emergent situations like natural disasters
or disease outbreaks. Linguistic resources for LORELEI include
Representative Language Packs for over 2 dozen low resource languages,
comprising data, annotations, basic natural language processing tools,
lexicons and grammatical resources. Representative languages are
selected to provide broad typological coverage, while Incident
Languages are selected to evaluate system performance on a language
whose identity is disclosed at the start of the evaluation, and for
which no training data has been provided. This package comprises all
of the resources and test set references for Oromo, which was one of
the Program's Incident Languages.
The evaluation protocol is based on a scenario in which some unforeseen event
(the "incident") triggers a need for humanitarian and logistical support in a
region where the predominant language (the "incident language") is one that
has received little or no attention as yet in NLP research. The objective for
evaluation participants is to provide NLP solutions, including information
extraction and machine translation, based only on limited resources and with
very little time for development.
For more information about LORELEI language resources, see
https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/lrec2016-lorelei-language-packs.pdf.
Each incident language pack has one or more focal incidents (a natural
disaster or other event which might trigger humanitarian needs). To
support the evaluation scenario, the evaluation package contents are
divided into the following subsets:
set0 : "pre-incident" text data and reference resources for the
language, including monolingual text, dictionaries, grammars,
and parallel or comparable text (in English and the incident
language); monolingual and parallel data in this set includes
documents published prior to the beginning of the earliest
focal incident and/or reference materials for which
publication date is not relevant, such as religious
materials
setE : "post-incident" text data that forms the basis for scoring
NLP system performance (using the scoring protocol and
software developed by NIST); set E consists of monolingual
text, along with reference translations and annotations
setS : "post-incident" text data in English, including information
that pertains to the incident itself; this was made available
to systems after the initial set of scorable outputs had been
submitted
set1 : supplemental "post-incident" text data, made available after
the initial set of scorable outputs had been submitted
set2 : a larger set of supplemental "post-incident" text data, made
available after the second set of scorable outputs had been
submitted
Each subset is presented as a directory within the data folder at the
top-level of the release package. Tools for data processing are provided
as part of set0 only, but are applicable to all sets.
2.0 Corpus organization
2.1 Directory Structure
The directory structure and contents of the package are summarized
below -- paths shown are relative to the base (root) directory of the
package:
./docs/README.txt -- this file
./data/set0/
./data/set0/tools/ -- software for data file format conversion
./data/set0/data/ -- monolingual and parallel text directories
./data/set0/dtds/ -- DTDs for all .xml data formats
./data/set0/docs/ -- lexical and grammatical resources, information
about various set0 components and properties
./data/set1/data/ -- monolingual text
./data/set1/docs/ -- information about various set1 components and properties
./data/set2/data/ -- monolingual text
./data/set2/docs/ -- information about various set2 components and properties
./data/setS/data/ -- monolingual text
./data/setS/docs/ -- information about various setS components and properties
./data/setE/data/monolingual_text/ -- monolingual text directory
./data/setE/data/annotation/
il6_edl.tab -- table of Entity-Detection-Linking annotations
situation_frame/ -- subdirectories for entity mentions, needs, and issues tables
./data/setE/data/translation/
eng/
ltf/ -- ltf.xml files (*.eng_A and *.eng_B versions for each Oromo doc)
psm/ -- psm.xml files
il6/
ltf/ -- ltf.xml files
psm/ -- psm.xml files
./data/setE/docs/ -- information about various setE components and properties
2.2 File Name Conventions
All monolingual text documents are presented as distinct files with unique
file names. For convenience, each file name provides a consistent set of
information about the content of the file via a set of fixed-width fields,
as follows:
- Language (3 letters)
- Genre (2 letters)
- Source (6-digit numeric)
- Date (8-digit numeric)
- Unique Index Number (9 alpha-numeric characters)
The language field for all Oromo documents uses "IL6" instead of the
ISO code for the language, as the practice in LORELEI was to refer to
incident languages by numeric identifiers to preserve the secrecy of
the language name until the start of the evaluation.
The date field for news reports represents the date of original publication
for the report. Where possible, discussion forum material uses the date when
a given discussion thread was initiated. When date information is not
available or meaningful for a given document, the date field will reflect
(roughly) the time at which the content was initially collected by the LDC,
and may be left "incomplete" by setting the "day" field (last two digits) to
zero (e.g. "20140900").
Files containing translations from a source language have the source language
identified in the "Language Code" field of the file name, and the translation
language as a 3-letter extension that immediately follows the main part of the
file name.
Pairs of corresponding files in "found" translation may have distinct
identifier strings (one with IL6 in the initial file name field, and one with
ENG in that field), if they were harvested independently of each other and
were later found to contain parallel content. Alternately, some sources of
found translation data present their own source and translated text as a
single unit, in which case the corresponding pair of files will have a single
identifier string, and the English member of the pair will have ".eng"
appended. In the former case, the alignment data specifies how the IL6 and
ENG files are paired.
2.3 Genres
Five genres are represented in this data set, as follows:
NW - news and general text harvested from news sites
SN - "social network" data (i.e. Twitter)
WL - weblog and newsgroup data
DF - discussion forum data
RF - data from "reference" materials, including religious text, government/NGO information sites, etc.
Note that the SN (Twitter) data cannot be distributed directly by LDC, due to
the Twitter Terms of Use. The file "docs/twitter_info.tab" (described in
Section 7.0 below) provides the necessary information for users to fetch the
particular tweets directly from Twitter.
3.0 Content Summary
3.1 Set 0
3.1.1 Monolingual text
Document and token counts of monolingual text by genre:
Genre N_Docs N_Tokens
NW 1624 803690
SN 8192 111012
WL 4028 2254231
3.1.2 Parallel and comparable text
Parallel text document and token count by genre (counts based on Oromo documents):
Genre N_Docs N_Tokens
RF 87 59358
Comparable text document and token count by genre (counts based on Oromo documents):
Genre N_Docs N_Tokens
NW 36 8317
WL 19 14594
All parallel text is aligned at the sentence level, while comparable
text is aligned into clusters of documents based on topic
similarity. Parallel and comparable text for Oromo and Enlgish can be
found in set0/data/translation/, which contains the following
structure of subdirectories:
found/
sentence_alignment/
eng/{ltf,psm}/
il6/{ltf,psm}/
comparable/
clusters/
eng/{ltf,psm}/
il6/{ltf,psm}/
The "found" data set consists of files from web data sources that had
parallel text content in Oromo and English. Each "leaf" directory in
the tree (*/ltf, */psm, sentence_alignment) contains a matched set of
data files. Parallel file pairs were identified and harvested
automatically, processed into LTF.xml format, and then aligned at the
level of "segments" (putative sentences). The alignment files
(*.align.xml) contain one or more "alignment" elements, in which one
or more "source" (English) segments is associated with one or more
"translation" (Oromo) segments. It's not assured that all segments in
a given (Oromo or English) data file are accounted for in a given set
of alignments. The sentence alignment files contain references to the
source document and the translation document (both files can be found
in their respective directories), and multiple "alignment" elements,
each of which contains one source element and one translation element.
The "segments" attribute of the source and translation element
contains space delimited segment ids referring to SEG IDs in the
corresponding ltf files.
NB: We refer to English as the "source" purely as a matter of
convenience and consistency across language packs; we do not have
confirmable evidence as to the true original language of a given data
file. In fact, for some web data sources, it may be the case that
documents were translated from some third language into both English
and Oromo.
The "comparable" data set is a more loosely structured inventory of
data files in which particular topics appear to be present in
documents in both languages during roughly the same period of time.
LDC used the results from two clustering techniques:
(1) Kutuzov et al. (https://arxiv.org/abs/1604.05372) for
multilingual document clustering on English and Oromo.
(2) Cosine similarity for monolingual document clustering on
English that was later augmented with Oromo documents.
For both approaches, the data was run on the tokenization of the
documents LTF.xml. The documents were divided into different sets,
where each set includes all documents with dates that span two weeks
(the weeks do not overlap). The final comparable text clusters
consist of English and IL6 documents that were clustered from both
approaches that fall within the same time period.
Note that some documents appear in multiple clusters.
The cluster files have names patterned as follows:
GN_clusters_YYYY-MM-DD_YYYY-MM-DD.xml
where "GN" is either "dfwl" or "nw" (lower-case), representing the
genre of the cluster (newswire or discussion forum/weblog). The third
and fourth fields are the beginning and end dates of the time span
during which the data files in that cluster were authored. The xml
structure in each cluster file consists of one or more "cluster"
elements, each of which contains some quantity of "doc" elements from
each language.
3.1.3 Lexical and grammatical resources
The docs/ directory contains two subdirectories:
categoryI_dictionary/
This directory contains the file IL6_categoryI_dictionary.pdf, which
is a bilingual Oromo-English dictionary ("An Afan Oromo-English,
English-Afan Oromo Dictionary" by E.C. Foot, published in 1913 and
included in the corpus on the understanding that it is in the public
domain.)
categoryII/
LORELEI Incident Language packs were required to contain (pointers
to) at least 5 of the following 8 "category II" resources:
-- bilingual IL-non-English dictionary
-- monolingual IL dictionary
-- bilingual grammar (reference grammar of the IL in English)
-- monolingual grammar in the IL
-- monolingual primer (grammar in the IL of the type used by school children)
-- bilingual gazetteer
-- monolingual gazetteer in the IL
-- monolingual gazetteer in English covering the incident region
The categoryII directory contains a pdf file (CategoryII_list.pdf)
with additional information and URLs for the resources identified. The
bilingual_gazetteer.txt is from Geonames (www.geonames.org) and is a
gazetteer for the country of Ethiopia.
3.2 Set 1
All data in this set is monolingual text in Oromo from the date of the
incident that serves as the focus of the evaluation and later. It may
contain some information about the incident, but also contains
documents whose content is not relevant to the incident in any way.
Genre N_Docs N_Tokens
DF 1 349
NW 645 257741
SN 4670 55056
WL 502 209413
3.3 Set 2
All data in this set is monolingual text in Oromo from the date of the
incident that serves as the focus of the evaluation and later. It may
contain some information about the incident, but also contains
documents whose content is not relevant to the incident in any way.
Genre N_Docs N_Tokens
DF 5 1254
NW 1278 508738
SN 9341 110495
WL 921 401469
3.4 Set S
All data in this set is monolingual text in English from the date of
the incident that serves as the focus of the evaluation and later. It
may contain some information about the incident, but also contains
documents whose content is not relevant to the incident in any way.
Genre N_Docs N_Tokens
NW 23 26445
3.5 Set E
3.5.1 Monolingual Text
This data set provides monolingual source data for the LORELEI 2017
Evaluation Test Set in Oromo. All data in this set is from the date
of the incident that serves as the focus of the evaluation and later.
Genre N_Docs N_Tokens
DF 2 987
NW 244 99769
SN 2432 31944
WL 132 67591
total 2810 200291
Because annotations obey the "full-token rule", meaning that all
reference annotation extents coincide with token boundaries as
provided by the automatic tokenization process, it was deemed to be
important for participants in the evaluation to be able to match the
LDC's tokenization for Twitter documents that they retrieved directly
from the Twitter API. For this reason, in set E only, the
monolingual_text directory contains "scrubbed" ltf for Twitter
documents. These ltf documents contain none of the actual tweet
content, but instead contain a series of underscores and whitespace
which allow users to match the tokenization of the tweet via the
character offsets provided in the ltf file.
3.5.2 Translation
Human reference translations were provided for a subset of the data in
the test set. Each Oromo document was translated into English by two
independent translators, and both translations are presented (with "A"
or "B" appended to the filename).
Genre N_Docs N_Tokens
DF 1 128
NW 70 25676
WL 70 25567
total 141 51371
The translation/ directory under setE/data/ contains source and reference translation files, as follows:
il6/{ltf,psm}/ -- contain 141 ltf/psm pairs
eng/{ltf,psm}/ -- contain 282 ltf/psm pairs: two reference translations,
having "eng_A" and "eng_B" in their respective file
names, for each source file
3.5.3 Annotation
Entity Detection and Linking and Situation Frame annotations were
applied to a subset of the data in the translation set, in order to
identify "entities", "needs" and "issues" to be detected by systems
for scoring purposes:
Genre N_Docs N_Tokens
DF 1 128
NW 47 16356
SN 1265 16580
WL 49 16597
total 1362 49661
Some of the files that received annotation did not yield annotatable
content for one or more annotation types. The next table shows the
number of files containing reference annotations of each type for each
genre:
Number of Files containing:
Genre Ents Needs Issues
------------------------------
DF 1 1 1
NW 47 38 41
SN 1168 136 1212
WL 48 33 46
The annotation/ directory under setE/data/ contains a tab delimited
file "il6_edl.tab" containing the entity linking annotation and a set
of directories containing situation frame annotation as follows:
situation_frame/ -- contains subdirectories for each type:
issues/
mentions/
needs/
Situation Frame annotation is designed to extract basic information
about where needs (such as a need for food) and relevant issues (such
as civil unrest) exist; the information is designed to be of the type
that would be useful for planning a disaster response effort. For
more detailed information about situation frame annotation, see
https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/smerp2017.pdf.
Guidelines for both the EDL and Situation Frame tasks are included in
the docs/ directory of set0.
4.0 Data Formats
The data formats described below are common across all sets.
4.1 PSM - Primary Structural Markup
When original data has structural markup interleaved with the language
content, we apply a filtering process that, in effect, separates the markup
and language content into distinct files. The language content (with
white-space normalization) goes into an RSD file (see below), and the relevant
markup content goes into a corresponding PSM file, which is a simple XML
stream comprising tags with attributes, and no other text content of its own.
(Configuring the filter for a given data source involves determining which
content and markup are "relevant"; the filter eliminates other content and
markup as irrelevant, such as ads, navigation menus, etc.)
Each PSM file has a single "psm" tag as its root element, and contains one
or more "string" tags. Each "string" refers to some span of text in the
corresponding RSD file, using "begin_offset" and "char_length" attributes, and
assigns a label to it, using a "type" attribute. (Note that offsets and
lengths are expressed as Unicode CHARACTER counts, not byte counts.)
The "type" attribute tells what sort of markup tag was used in the original
data to contain the given string (e.g. "p", "quote", etc.); when sentence
segmentation can be done as part of the filtering step, a "string" tag with
type="seg" is used to label the span of each detected sentence.
Some structural tags in original data contain attributes that may be relevant
to language research; for example, in a file that contains a thread from a
discussion forum, it's useful to keep track of the dates and authors of posts
within the thread. For these cases, the "string" element can contain one ore
more "attribute" elements, to preserve the name and value of the given
attribute - e.g.:
As shown in this example, the "attribute" tag is also used, where appropriate,
to assign an ID value (unique within the file) to each string of a given type;
this is also used with the "seg"-type strings to assign IDs to detected
sentences.
PSM files appear in the data/monolingual_text/ and data/translation/
directories of each set.
4.2 LTF - LORELEI Text Format
LTF was originally developed for language packs produced in the REFLEX Program
("LCTL Text Format"). This XML format uses structural tags "SEG" and "TOKEN"
to mark sentence segmentation and word tokenization of the source data. The
full original text of each sentence (SEG) is contained in an "ORIGINAL_TEXT"
tag, and each individual word and punctuation string is contained, in order of
occurrence, in a sequence of "TOKEN" elements, along with various attributes
for each token. Both SEG and TOKEN attributes include character offsets
relative to beginning of the raw source data ("RSD" file format, described
below), with the offset of the first character being 0.
LTF files appear in the data/monolingual_text/ and data/translation/
directories of each set.
4.3 EDL (Entity Detection and Linking)
The file "il6_edl.tab" contains all EDL annotations for the IL6 EDL subset.
The table contains eight columns, as follows:
column 1: system_run_id -- "LDC"
column 2: mention_id
column 3: mention_text
column 4: extents
column 5: kb_id -- numeric-ID or "NIL"+numeric, may contain multiple KB links
separated by | ("pipe" symbol)
column 6: entity_type
column 7: mention_type
column 8: confidence
When column 5 is fully numeric, it is a citation to a numbered entity in
LORELEI Entity Detection and Linking Knowledge Base (distributed separately
as LDC2020T10); when it consists of "NIL" plus digits, it refers to an
entity that is not present in the Knowledge Base, but this label is used
consistently for all mentions of the particular entity.
Note that for any annotated Twitter documents, text extents have been
replaced by underscore ("_") characters to comply with the prohibition
against distributing the text of tweets directly. Character offsets
can be used to align the annotations with the tweets once the user has
downloaded them using Twitter's API.
4.4 Situation Frame
Situation frame annotation consists of three parts, each presented as a
separate tab-delimited file: entities, needs, and issues. The details of each
table are described below.
Entities, mentions, need frames, and issue frames all have IDs that follow a
standard schema consisting of a prefix designating the type of ID ('Ent' for
entities, 'Men' for mentions, and 'Frame' for both need and issue frames), an
alphanumeric string identifying the annotation "kit", and a numeric string
uniquely identifying the specific entity, mention, or frame within the
document.
4.4.1 Entities
The grouping of entity mentions into "selectable entities" for situation frame
annotation is provided in the mentions/ subdirectory. The table has 8 columns
with the following headers and descriptions:
column 1: doc_id -- doc ID of source file for the annotation
column 2: entity_id -- unique identifier for each grouped entity
column 3: mention_id -- unique identifier for each entity mention
column 4: entity_type -- one of PER, ORG, GPE, LOC
column 5: mention_status -- 'representative' or 'extra';
representative mentions are the ones which have been chosen by the
annotator as the representative name for that entity. Each entity
has exactly one representative mention.
column 6: start_char -- character offset for the start of the mention
column 7: end_char -- character offset for the end of the mention
column 8: mention_text -- mention string
Again, note that for any annotated Twitter documents, text extents
have been replaced by underscore ("_") characters to comply with the
prohibition against distributing the text of tweets directly.
4.4.2 Needs
Annotation of need frames is provided in the needs/ subdirectory. Each row in
the table represents a need frame in the annotated document. The table has 13
columns with the following headers and descriptions:
column 1: user_id -- user ID of the annotator
column 2: doc_id -- doc ID of source file for the annotation
column 3: frame_id -- unique identifier for each frame
column 4: frame_type -- 'need'
column 5: need_type -- exactly one of 'evac' (evacuation), 'food' (food
supply), 'search' (search/rescue), 'utils' (utilities, energy, or
sanitation), 'infra' (infrastructure), 'med' (medical assistance),
'shelter' (shelter), or 'water' (water supply)
column 6: place_id -- entity ID of the LOC or GPE entity identified as the
place associated with the need frame; only one place value per
need frame, must match one of the entity IDs in the corresponding
ent_output.tsv or be 'none' (indicating no place was named)
column 7: proxy_status -- 'True' or 'False'
column 8: need_status -- 'current', 'future'(future only), or 'past' (past only)
column 9: urgency_status -- 'True' (urgent) or 'False' (not urgent)
column 10: resolution_status -- 'sufficient' or 'insufficient' (insufficient /
unknown sufficiency)
column 11: reported_by -- entity ID of one or more entities reporting
the need; multiple values are comma-separated, must match entity IDs
in the corresponding ent_output.tsv or be 'none'
column 12: resolved_by -- entity ID of one or more entities resolving
the need; multiple values are comma-separated, must match entity IDs
in the corresponding ent_output.tsv or be 'none'
column 13: description -- string of text entered by the annotator as
memory aid during annotation, no requirements for content or language,
may be 'none'
4.4.3 Issues
Annotation of issue frames is provided in the issues/ subdirectory. Each row
in the table represents an issue frame in the annotated document. The table has
9 columns with the following headers and descriptions:
column 1: user_id -- user ID of the annotator
column 2: doc_id -- doc ID of source file for the annotation
column 3: frame_id -- unique identifier for each frame
column 4: frame_type -- 'issue'
column 5: issue_type -- exactly one of 'regimechange' (regime change),
'crimeviolence' (civil unrest or widespread crime), or 'terrorism'
(terrorism or other extreme violence)
column 6: place_id -- entity ID of the LOC or GPE entity identified as
the place associated with the issue frame; only one place value per
issue frame, must match one of the entity IDs in the corresponding
ent_output.tsv or be 'none'
column 7: proxy_status -- 'True' or 'False'
column 8: issue_status -- 'current' or 'not_current'
column 9: description -- string of text entered by the annotator as
memory aid during annotation, no requirements for content or
language, may be 'none'
5.0 Software tools included in this release
All software tools are provided in the tools/ directory of Set 0.
5.1 "ltf2txt" (source code written in Perl)
A data file in ltf.xml format (as described above) can be conditioned
to recreate exactly the the "raw source data" text stream (the rsd.txt
file) from which the LTF was created. The tools described here can be
used to apply that conditioning, either to a directory or to a zip
archive file containing ltf.xml data. In either case, the scripts
validate each output rsd.txt stream by comparing its MD5 checksum
against the reference MD5 checksum of the original rsd.txt file from
which the LTF was created. (This reference checksum is stored as an
attribute of the "DOC" element in the ltf.xml structure; there is also
an attribute that stores the character count of the original rsd.txt
file.)
Each script contains user documentation as part of the script content;
you can run "perldoc" to view the documentation as a typical unix man
page, or you can simply view the script content directly by whatever
means to read the documentation. Also, running either script without
any command-line arguments will cause it to display a one-line
synopsis of its usage, and then exit.
ltf2rsd.perl -- convert ltf.xml files to rsd.txt (raw-source-data)
ltfzip2rsd.perl -- extract and convert ltf.xml files from zip archives
5.2 "twitter-processing" (source code written in Ruby)
Due to the Twitter Terms of Use, the text content of individual tweets
cannot be redistributed by the LDC. As a result, users must download
the tweet contents directly from Twitter and condition/normalize the
text in a manner equivalent to what was done by the LDC, in order to
reproduce the Oromo raw text that was used by LDC for annotation. The
twitter-processing software provided in the tools/ directory enables
users to perform this normalization and ensure that the user's version
of the tweet matches the version used by LDC, by verifying that the
md5sum of the user-downloaded and processed tweet matches the md5sum
provided in the twitter_info.tab file. Users must have a developer
account with Twitter in order to download tweets, and the tool does
not replace or circumvent the Twitter API for downloading tweets.
The twitter_info.tab file provides the twitter download id for each
tweet, along with the LORELEI file name assigned to that tweet and the
md5sum of the processed text from the tweet.
The file "README.md" in the tools/twitter-processing/ directory
provides details on how to install and use the source code in this
directory in order to condition text data that the user downloads
directly from Twitter and produce both the normalized raw text and the
segmented, tokenized LTF.xml output.
5.3 Encoding
The common framework for text processing in LORELEI includes a
“normalization” step, which allows for rectifying variations in
orthography and/or punctuation that may occur with some frequency in
this or that particular language. For overall simplicity and
consistency in processing across all languages, this normalization
step is always invoked; in languages such as Oromo that require no
special normalization, this step leaves the data unchanged.
6.0 Documentation included in this release
Each set has its own docs directory, but the types of files found
there are consistent across the sets, as described below.
IL6_IncidentDescription.pdf and IL6_IncidentDescription_Appendix.pdf:
provide description and additional links and information about the
incidents that were the focus of the evaluation data set. Found in
set0/docs/ only.
SimpleNamedEntityGuidelines_IL6_V1.2.pdf,
EntityLinkingGuidelines_V1.2.1.pdf and
SituationFrameGuidelines_V3.0.pdf:
guidelines for entity annotation, entity linking, and situation frame
annotation. Found in set0/docs/ only.
twitter_info.tab:
contains tab-separated columns: doc uid, tweet id, normalized md5 of
the tweet text, and tweet author id for all tweets in the
release. Found in all sets (except set S, which contains no Twitter
data).
source_codes.tab:
contains tab-separated columns: genre, source code, source name, and
base url for each source in the release. Found in all sets.
urls.tab:
contains tab-separated columns: doc uid and url. Note that the url
column is empty for documents from older releases for which the url is
not available; they are included here so that the uids column can
serve as a document list for the package. Found in all sets.
annotated_filelist_EDL.txt, annotated_filelist_MT.txt,
annotated_filelist_SF.tab:
list of all files annotated for the EDL task, all files with human
reference translations, and all files annotated for the Situation
Frame task. Found in setE only.
domain_filelist.tab:
lists all documents for which human reference translations were
produced and provides a domain judgement: eval_incident (document
contains information about the incidents that were the focus of the
evaluation), indomain (document is relevant to the overall LORELEI
domain of humanitarian assistance and disaster relief and related
situations, but not specifically the incident of focus), or nondomain
(document is of unspecified topic, not related to the LORELEI domain
or incidents). Found in setE/docs/ only.
filelist.txt:
lists the doc id for all documents in set E. Found in setE/docs/ only.
7.0 Acknowledgement
The authors would like to acknowlege the following contributors to
this corpus: Song Chen, Dana Delgado, Neville Ryant, Brian Gainor,
Neil Kuster, University of Maryland Applied Research Laboratory for
Intelligence and Security (ARLIS), formerly UMD Center for Advanced
Study of Language (CASL), and our team of Oromo annotators.
This material is based upon work supported by the Defense Advanced Research
Projects Agency (DARPA) under Contract No. HR0011-15-C-0123. Any opinions,
findings and conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the views of DARPA.
8.0 Copyright
Portions © 2015 Addis Standard Magazine, © 2015-2016 Agroindustrial
Association of Ukraine, © 2015-2016 Al Jazeera Media Network, © 2016
AllAfrica, © 2015-2017 Ayyaantu, © 2016 BBC, © 2013-2017 Bilisummaa, ©
2016 Cable News Network. Turner Broadcasting System, Inc., © 2015-2017
Daandiihaqaa.net, © 2016 dmg media, © 2016-2017 Ethiopian Satellite
Television Service (ESAT), © 2016 Geeskaafrika, © 2015 Guardian News &
Media Limited or its affiliated companies, © 2016 IPI International
Peace Institute, © 2016-2017 Oromedia, © 2015-2017 Oromiya National
Regional State Office of the President, © 2016 Oromo Democratic Front,
© 1998, 2012, 2015-2017 Oromo Liberation Front, © 2016 Reuters, ©
2016-2017 Robe Media Network, © 2015-2017 Shaggar, © 2014-2017 The
Awash Post, © 2016 The Washington Post, © 2016 Tigraionline.com, ©
2017 Watch Tower Bible and Tract Society of Pennsylvania, © 2015-2017
Xalayaa, © 2017, 2019 Trustees of the University of Pennsylvania
9.0 Contacts
If you have questions about this data release, please contact the
following personnel at LDC.
Stephanie Strassel - LORELEI PI
Jennifer Tracey - LORELEI Project Manager
Jonathan Wright - LORELEI Technical Lead