Skip to content

Latest commit

 

History

History
41 lines (26 loc) · 2.58 KB

README.md

File metadata and controls

41 lines (26 loc) · 2.58 KB
Raw

Sanaphor - semantic anaphora resolution

System

  1. IntermediateCoreferenceSystem.java generates a mentions file, containing all mentions extracted by Stanford Coref + their attributes. The file has the following format:

     DOC_ID  PARAGRAPH_ID    SENT_ID MENTION_ID      START_INDEX     END_INDEX       MENTION NER_ENTITY      HEAD_WORD_LEMMA HEAD_POS_TAG    COREF_ID        GOLD_COREF_ID   NER_ENTITY_TAG  POS_TAG_SEQ
 bc/cctv/00/cctv_0000    0       0       10099   9       10      people  null    people  NNS     10099   -1      O       NNS
 bc/cctv/00/cctv_0000    0       0       23      23      25      Hong Kong       Hong Kong       Kong    NNP     23      23      GPE     NNP NNP

  2. At the same time the system generates two more files in the CoNLL format:

  3. Gold coreferences file, which contains gold coreference resolutions;

  4. Stanford coreferences file, which contains coreferences produced by the Stanford Coref System.

These files can be used by the official CoNLL scorer script to evaluate coreference resultion systems. 3. Next, we link entities from the mentions file using state-of-the-art Entity Linking system, which produces the following linked mentions file:

    DOC_ID  PARAGRAPH_ID    SENT_ID MENTION_ID      START_INDEX     END_INDEX       MENTION NER_ENTITY      HEAD_WORD_LEMMA HEAD_POS_TAG    COREF_ID        GOLD_COREF_ID   NER_ENTITY_TAG  POS_TAG_SEQ     null
    bc/cctv/00/cctv_0000    0       0       10099   9       10      people  null    people  NNS     10099   -1      O       NNS     null
    bc/cctv/00/cctv_0000    0       0       23      23      25      Hong Kong       Hong Kong       Kong    NNP     23      23      GPE     NNP NNP http://dbpedia.org/resource/Hong_Kong
  1. Finally, we run sanaphor.py on the linked mention mentions, this package implements methods described in the paper. It produces modified file in the CoNLL format, that can be evaluated using CoNLL scorer as usual.

CoNLL Scorer

Download from: https://github.com/conll/reference-coreference-scorers

Running:

~/reference-coreference-scorers/v8.01/scorer.pl all conll-test.predicted.new.txt conll-test-gold.txt

Extra

Data

Ontonotes 5.0 - https://catalog.ldc.upenn.edu/LDC2013T19

Generate CoNLL files from Skeleton files:

~/conll-2012/v3/scripts/skeleton2conll.sh -D ~/ontonotes-release-5.0/data/files/data ~/conll-2012/

Stanford coreNLP shell

/Library/Internet\ Plug-Ins/JavaAppletPlugin.plugin/Contents/Home/bin/java -cp "*" -Xmx2g edu.stanford.nlp.pipeline.StanfordCoreNLP -annotators tokenize,ssplit,pos,lemma,ner,parse,dcoref