ACE 2004 Multilingual Training Corpus LDC2005T09 February 14, 2005 1.0 Introduction This publication contains the complete set of English, Arabic and Chinese training data for the 2004 Automatic Content Extraction (ACE) technology evaluation. The corpus consists of data of various types annotated for entities and relations and was created by Linguistic Data Consortium with support from the ACE Program, with additional assistance from the DARPA TIDES (Translingual Information Detection, Extraction and Summarization) Program. This data was previously distributed as an e-corpus (LDC2004E17) to participants in the 2004 ACE evaluation. The objective of the ACE program is to develop automatic content extraction technology to support automatic processing of human language in text form. In September 2004, sites were evaluated on system performance in six areas: Entity Detection and Recognition (EDR), Entity Mention Detection (EMD), EDR Co-reference, Relation Detection and Recognition (RDR), Relation Mention Detection (RMD), and RDR given reference entities. All tasks were evaluated in three languages: English, Chinese and Arabic. The current publication consists of the official training data for these evaluation tasks. A seventh evaluation area, Timex Detection and Recognition, is supported by the ACE Time Normalization (TERN) 2004 English Training Data Corpus (LDC2005T907). The TERN corpus source data largely overlaps with the English source data contained in the current release. A complete description of the ACE 2004 Evaluation can be found on the ACE Program website maintained by the National Institute of Standards and Technology (NIST): http://www.nist.gov/speech/tests/ace/ For more information about linguistic resources for the ACE Program, including annotation guidelines, task definitions, free annotation tools and other documentation, please visit LDC's ACE website: http://www.ldc.upenn.edu/Projects/ACE 1.1 Data Profile Data for each language is contained in individual directories. Within each language directory, files are divided into subdirectories by source type: bnews (Arabic, Chinese, English) - Broadcast news data from the TDT-4 Corpus nwire (Arabic, Chinese, English) - Newswire data from the TDT-4 Corpus arabic_treebank (Arabic, English) - data from the Arabic Treebank 1 Corpus; English translations from the MT-2003 translation data set chinese_treebank (Chinese, English) - data from the Chinese Treebank Version 4; English translations from the Chinese Treebank English Parallel Text Corpus fisher_transcripts (English) - data from the Fisher Telephone Speech collection, sponsored by DARPA EARS Each language directory also contains a FileList, containing the following information for each file: FileID annotation status (fpDone - first pass only; qcDone - additional second quality checking pass completed) source type - as described above word count The total volume of data for each language is summarized below: English (All files QC-passed) broadcast news - 220 files (60,291 words) newswire - 128 files (59,840 words) Chinese Treebank translation - 37 files (12,337 words) Arabic Treebank translation - 58 files (12,855 words) Fisher telephone conversation - 8 files (12,630 words) total - 451 files (157,953 words) Chinese (All files QC-passed) broadcast news - 314 files (135,405 characters; 67,702.5 words) newswire - 226 files (120502 characters; 60251 words) Chinese Treebank - 106 files (51,499 characters; 25,749.5 words) total - 646 files (307,406 characters; 153,703 words) Arabic (415 files QC-passed; see Arabic/FileList_Arabic for a list of QC-passed files) broadcast news - 304 files (63,238 words) newswire - 253 files (63,122 words) Arabic Treebank - 132 files (25,010 words) total - 689 files (151,360 words) 1.2 Annotation Content Entity Detection and Tracking (EDT) is the core annotation task, providing the foundation for all remaining tasks. We identify seven types of entities: Person, Organization, Location, Facility, Weapon, Vehicle and Geo-Political Entity (GPEs). Each type is further divided into subtypes (for instance, Organization subtypes include Government, Commercial, Educational, Non-profit, Other). Annotators tag all mentions of each entity within a document, whether named, nominal or pronominal. For every mention, the annotator identifies the maximal extent of the string that represents the entity, and labels the head of each mention. Nested mentions are also captured. Each entity is classified according to its type and subtype. Each entity mention is further tagged according to its class . specific, generic, attributive, negatively quantified or underspecified. During the LNK annotation task, annotators review the entire document to group mentions of the same entity together; they also label cases of metonymy, where the name of one entity is used to refer to another entity (or entities) related to it. Relation Detection and Characterization (RDC) involves the identification of relations between entities. RDC targets physical relations including Located, Near and Part-Whole; social/personal relations including Business, Family and Other; a range of employment or membership relations; relations between artifacts and agents (including ownership); affiliation-type relations like ethnicity; relationships between persons and GPEs like citizenship; and finally discourse relations. For every relation, annotators identify two primary arguments (namely, the two ACE entities that are linked) as well as the relation's temporal attributes. Relations that are supported by explicit textual evidence are distinguished from those that depend on contextual inference on the part of the reader. The full annotation guidelines that were followed in creating this data release are included in the /doc directory. Previous and more recent versions of the ACE annotation task specifications can be found on LDC's ACE website: http://www.ldc.upenn.edu/Projects/ACE 1.3 File Format Description Each of the English, Chinese and Arabic directories contains the following types of files. APF files (.apf.xml) : The official ACE Program Format. Please refer to README--APF-FORMAT for a brief description of this data format and its mapping to the other data formats. ALF files (.alf.xml) : The ACE LDC Format is an intermediate format similar to APF designed to store all annotation content represented in the AG files. AG files (-pp.ag.xml): The LDC Annotation Graph Format (postprocessed). LDC's internal annotation files format for ACE. These files can be viewed with LDC's free ACE annotation tools. Source text files (.sgm): Source text files. All files, including the Chinese and Arabic data, are in UTF-8 (Unicode). 1.4 Character Counting The annotation files use a character-based, not byte-based, offset counting method: the first Unicode character has the offset of zero; the second character has the offset of one, and so on. The APF and ALF file formats do not count SGML tags in their offsets, while the AG format does count them. The source SGML files use the UNIX-style end-of-line character (LF), and each end-of-line character is counted as one character. 1.5 DTDs There are three DTDs included in the dtd/ directory. APF DTD file (apf.v4.0.1.dtd): The DTD for the APF format. ALF DTD file (alf.v4.0.1.dtd): The DTD for the ALF format. AG DTD file (ag-1.1.dtd): The DTD for the AG format. 1.6 Source File Text Formats All source files have the following tags: DOC DOCID/DOCNO (published TDT4 files used the DOCNO tag; ATB and CTB files use the DOCID tag) TEXT 1.6.1 Chinese Treebank Files/Arabic Treebank Files The Chinese treebank files were taken from the published Chinese Treebank Version 4 corpus. The Arabic treebank files were taken from the Arabic Treebank 1 Corpus. These files have the same headers as the original files (e.g., DOCID, HEADLINE, etc.), but the format of the TEXT part of the files have been modified in the following way: o P tags and S tags have been removed --- sentence boundaries are indicated with blank lines. When there is a new P tag after a sentence, an additional blank line is used. 1.6.2 Fisher Telephone Conversation Transcripts Speaker IDs are indicated with a single capital letter followed by a colon (i.e., A: and B:). Start and end times have been removed, and empty segments (i.e., segments without transcripts) have been removed. A: what do you think of the topic B: so what ... 1.7 COPYRIGHT Information Portions (c) 1994-1998, 2000 Xinhua News Agency (c) 1997 Department of Information Services, Hong Kong Special Administrative Region (c) 1996-1998 & 2000-2001 Sinorama Magazine (c) 2000 Agence France-Presse, New York Times, Associated Press Newswire, Zaobao, An-Nahar, Al-Hayat, Nile TV, Voice of America, Public Radio International, Cable News Network, American Broadcasting Corporation, National Broadcasting Company, China National Radio, China Television System, China Central TV, China Broadcasting System. The World is a co-production of Public Radio International and the British Broadcasting Corporation and is produced at WGBH Boston.