==================================================== CoNLL-2014 Shared Task: Grammatical Error Correction Description of Preprocessed NUCLE Data Set 7 April 2014 Version 3.2 ==================================================== 1. General ========== This README file describes a preprocessed version of the NUS Corpus of Learner English (NUCLE). For information about NUCLE, please refer to the NUCLE README file. For information about the CoNLL-2014 shared task, please refer to the shared task website. The preprocessed data set, following earlier CoNLL shared tasks, provides syntactic information for the raw texts in NUCLE. For each sentence, the part of speech tags, dependency parse tree, and constituent parse tree are encoded in a column format. In NUCLE, annotations are made at the character level, which means both the start offset and the end offset of an error annotation are character positions in the corresponding paragraph. In this preprocessed version, annotations are made at the token level, which means the start offset and the end offset are indexes of tokens in the corresponding sentence. This README is updated on 4 August 2014. 2. Files ======== Two files are to be generated: conll14st-preprocessed.conll conll14st-preprocessed.conll.ann conll14st-preprocessed.conll contains the preprocessed data in CoNLL-style column format. This file is not included in this distribution due to the size. conll14st-preprocessed.conll.ann contains token-level annotations. 3. Preprocessing systems ======================== The NUCLE corpus is preprocessed with the following steps to generate this preprocessed data set: a). sentence splitting, using nltk punkt [1]. Note: the version used to generate the files is before the fixing of issue 514 b). word tokenization, using nltk word_tokenize [1]. c). POS tags, dependency parse trees, and constituent parse trees, using the Stanford parser [2]. d). projecting character-level annotation to token-level annotation. Results from (a-c) are in conll14st-preprocessed.conll. The projected annotations (d) are included in conll14st-preprocessed.conll.ann. 4. Data format ============== Here is an example sentence in conll14st-preprocessed.conll: NID PID SID TOKENID TOKEN POS DPHEAD DPREL SYNT 829 1 2 0 This DT 1 det (ROOT(S(NP* 829 1 2 1 will NN 7 nsubj *) 829 1 2 2 , , - - * 829 1 2 3 if IN 4 mark (SBAR* 829 1 2 4 not RB 7 dep (FRAG* 829 1 2 5 already RB 4 dep (ADVP*))) 829 1 2 6 , , - - * 829 1 2 7 caused VBD -1 root (VP* 829 1 2 8 problems NNS 7 dobj (NP*) 829 1 2 9 as IN 11 mark (SBAR* 829 1 2 10 there EX 11 expl (S(NP*) 829 1 2 11 are VBP 7 advcl (VP* 829 1 2 12 very RB 13 advmod (NP(NP(ADJP* 829 1 2 13 limited VBN 14 amod *) 829 1 2 14 spaces NNS 11 nsubj *) 829 1 2 15 for IN 14 prep (PP* 829 1 2 16 us PRP 15 pobj (NP*))))))) 829 1 2 17 . . - - *)) The columns represent the following: Column Type Description 0 NID Document id of the sentence, equals to "nid" in NUCLE. 1 PID Paragraph index of the sentence, according to the paragraphing in NUCLE (

). 2 SID Sentence index in paragraph, each sentence has its own index starting from 0. 3 TOKENID Token index in the sentence, starting from 0. 4 TOKEN Word/token. 5 POS Part of speech tag. 6 DPHEAD Index of parent in dependency tree. 7 DPREL Dependency relation with parent. 8 SYNT Constituent tree. The constituent tree can be recovered as follows: (a) Replacing "*" in this column with a string "(pos word)", where pos is the value of column 5, word is the value of column 4. (b) Concatenating all the strings in (a) gives the bracketing structure of the constituent parse tree. ------------------------------------------------------------------------ Here is the corresponding token-level annotation for the above sentence (in conll14st-preprocessed.conll.ann): Vform cause Nn space SVA is The tags represent the following: Tag Description Each section identifies annotations for one sentence. Identifies an error annotation, with the following attributes: nid: Document id of the sentence, equals to the NID column (column 0) in .conll file pid: Paragraph index of the sentence, equals to the PID column in .conll file sid: Sentence index in paragraph, equals to the SID column in .conll file start_token: The token index (TOKENID column in .conll file) which is the start of annotation. end_token: The token index which is the end of the annotation. Error tag (refer to the NUCLE corpus README file for the complete list of error tags). Correction, replacing tokens in the interval [start_token, end_token) with the correction string will result in a corrected sentence. ------------------------------------------------------------------------ How to map a sentence to its annotation? In conll14st-preprocessed.conll, different sentences are separated by empty lines, and sections are also separated by empty lines in conll14st-preprocessed.conll.ann. A sentence maps to one section, with the same nid, pid, and sid. If a sentence has no annotation, there is no section for it. The order of the sections is the same as the order of sentences in the preprocesed file. 5. Updates included in version 2.1 ================================== The error categories Wcip and Rloc have been mapped to Prep, Wci, ArtOrDet, and Rloc-, to facilitate the detection and correction of preposition errors and article/determiner errors. See the NUCLE README file for details on how the mapping was done. 50 overlapping error annotations involving the 5 error tags ArtOrDet, Prep, Nn, Vform, and SVA have been modified such that they do not overlap in the revised error annotations. Some minor mistakes in error annotations have been corrected. This was done for the CoNLL-2013 Shared Task. 6. Updates included in version 3.0 ================================== The overlapping error annotations of the remaining error tags have been modified such that they do not overlap in the revised error annotations. Some minor mistakes in error annotations have been corrected. 7. Updates included in version 3.1 ================================== Duplicate error annotations, each having the same span and correction string but different error type, are removed to keep only one of those. Previously, annotations spanning to the end of the paragraph was detected as cross-sentence and therefore not included in the preprocessed format. This has been fixed such that they are included in the preprocessed format. Some minor mistakes in error annotations have been corrected. 8. Updates included in version 3.2 ================================== The preprocessing script has been fixed such that gold edits that insert empty string are not included in the token-level gold edit and scorer answer files. 9. References ============= [1] Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing with Python. O'Reilly Media Inc. http://nltk.org/ [2] Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pp. 423-430. Stanford parser version 2.0.1. http://nlp.stanford.edu/software/stanford-parser-2012-03-09.tgz