NUCLE Release 3.3 24 Jan 2019 This README file describes the NUS Corpus of Learner English (NUCLE). It was collected in a collaboration project between the National University of Singapore (NUS) Natural Language Processing (NLP) Group led by Prof. Hwee Tou Ng and the NUS Centre for English Language Communication (CELC) led by Prof. Siew Mei Wu. The work was carried out as part of the PhD thesis research of Daniel Dahlmeier at the NUS NLP Group. The corpus is distributed under the standard NUS licensing agreement available when downloading the corpus. Any questions regarding NUCLE should be directed to Hwee Tou Ng at: nght@comp.nus.edu.sg If you are using the NUCLE corpus in your work, please include a citation of the following paper: Daniel Dahlmeier, Hwee Tou Ng, and Siew Mei Wu (2013). Building a Large Annotated Corpus of Learner English: The NUS Corpus of Learner English. Proceedings of the Eighth Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2013). (pp. 22 -- 31). Atlanta, Georgia, USA. 1. About ======== NUCLE is a corpus of learner English. It consists of about 1,400 essays written by university students at the National University of Singapore on a wide range of topics, such as environmental pollution, healthcare, etc. It contains over one million words which are completely annotated with error categories and corrections. All annotations have been performed by professional English instructors at the NUS CELC. 2. Data Format ============== The corpus is distributed in a simple SGML format. All annotations come in a "stand-off" format. The start position and end position of an annotation are given by paragraph and character offsets. Paragraphs are enclosed in

...

tags. Paragraphs and characters are counted starting from zero. Each annotation includes the following fields: the error category, the correction, and optionally a comment. If the correction replaces the original text at the given location, it should fix the grammatical error. Example:

Engineering design process can be defined as a process ...

Firstly, engineering design ...

... ArtOrDet The engineering design process ... ... Below is a complete list of the error categories in NUCLE release 2.0: ERROR TAG ERROR CATEGORY --------------------------- Vt Verb tense Vm Verb modal V0 Missing verb Vform Verb form SVA Subject-verb-agreement ArtOrDet Article or Determiner Nn Noun number Npos Noun possesive Pform Pronoun form Pref Pronoun reference Wcip Wrong collocation/idiom/preposition Wa Acronyms Wform Word form Wtone Tone Srun Runons, comma splice Smod Dangling modifier Spar Parallelism Sfrag Fragment Ssub Subordinate clause WOinc Incorrect sentence form WOadv Adverb/adjective position Trans Link word/phrases Mec Punctuation, capitalization, spelling, typos Rloc Local redundancy Cit Citation Others Other errors Um Unclear meaning (cannot be corrected) 3. Updates included in version 2.1 ================================== The major change made in version 2.1 is to map the error categories Wcip and Rloc to Prep, Wci, ArtOrDet, and Rloc-. In the original NUCLE corpus, there is not an explicit preposition error category. Instead, preposition errors are part of the Wcip (Wrong collocation/idiom/preposition) and Rloc (local redundancy) error categories. In addition, redundant article or determiner errors are part of the Rloc error category. To facilitate the detection and correction of preposition errors and article/determiner errors, we perform mapping of error categories in the original NUCLE corpus. The mapping relies on POS tags, constituent parse trees, and error annotations at the token level. (a) Conditions to change from the error category Wcip or Rloc to Prep: This applies to replacing a preposition by another preposition, or deleting a preposition. The string to be replaced is one word w with POS tag IN or TO, the parent of w is a PP in the constituent parse tree, and the replacement is either a preposition or the empty string. (b) Conditions to change from the error category Wcip to Prep: This applies to inserting a preposition. The replacement is a preposition (one word only) and the immediately following word is tagged as VBG or is the first word of a noun phrase (NP). (c) Conditions to change from the error category Rloc to ArtOrDet: The single word has POS tag DT and the replacement is the empty string. The remaining unaffected "Wcip" errors are assigned the new error category "Wci" and the remaining unaffected "Rloc" errors are assigned the new error category "Rloc-". List of 36 Prepositions: about along among around as at beside besides between by down during except for from in inside into of off on onto outside over through to toward towards under underneath until up upon with within without Below is a complete list of 28 error categories in NUCLE release 2.1: ERROR TAG ERROR CATEGORY --------------------------- Vt Verb tense Vm Verb modal V0 Missing verb Vform Verb form SVA Subject-verb-agreement ArtOrDet Article or Determiner Nn Noun number Npos Noun possesive Pform Pronoun form Pref Pronoun reference Prep Preposition Wci Wrong collocation/idiom Wa Acronyms Wform Word form Wtone Tone Srun Runons, comma splice Smod Dangling modifier Spar Parallelism Sfrag Fragment Ssub Subordinate clause WOinc Incorrect sentence form WOadv Adverb/adjective position Trans Link word/phrases Mec Punctuation, capitalization, spelling, typos Rloc- Local redundancy Cit Citation Others Other errors Um Unclear meaning (cannot be corrected) 3. Updates included in version 2.2 ================================== - Fixed the bug on expanding an error annotation involving part of a token to the full token. - Other miscellaneous corrections were made. 4. Updates included in version 2.3 ================================== - Fixed the bug involving tokenization of punctuation symbols in the correction string. - Fixed the tokenization example in the README file of the M^2 scorer to reflect the real tokenization to be used and removed irrelevant codes from the scorer package. 5. Updates included in version 3.0 ================================== - Resolved overlapping annotations in the NUCLE corpus to make them non-overlapping. - Corrected some minor mistakes in error annotations. 6. Updates included in version 3.1 ================================== - Removed duplicate annotations in the NUCLE corpus with the same span and correction string but different error type so as to keep only one of those annotations. This fix only affects 0.1% of all annotations. - Fixed end-of-paragraph annotations so that the end offset of such annotations is the last character position in the paragraph. This fix only affects 0.7% of all annotations. - Corrected some minor mistakes in error annotations. - Inclusion of the CoNLL-2013 test data, with all the known problems described above fixed. Participating teams in the CoNLL-2014 shared task can make use of the CoNLL-2013 test data in training and developing their systems if they wish to do so. - Fixed a minor bug in the M2 scorer that caused duplicate insertion edits to receive high scores. 7. Updates included in version 3.2 ================================== - Fixed the preprocessing script such that a gold edit that inserts an empty string is not included in the token-level gold edit and scorer answer files. - Removed one edit that inserted an empty string from the CoNLL-2014 test data. Also removed such instances from the NUCLE training data. - Fixed a bug in the M2 scorer arising from scoring against gold edits from multiple annotators. Specifically, the bug sometimes caused incorrect scores to be reported when scoring against the gold edits of subsequent annotators (other than the first annotator). - Fixed a bug in the M2 scorer that caused erroneous scores to be reported when dealing with insertion edits followed by deletion edits (or vice versa). 8. Updates included in version 3.3 ================================== - Added a subdirectory "bea2019" to contain the ERRANT typed NUCLE M2 file for the BEA 2019 shared task on grammatical error correction.