Korean word sense disambiguation evaluation dataset

We introduce the Korean word sense disambiguation evaluation dataset built by the Bareun team.
An ambiguous word is an eojeol (a space-delimited word unit) whose surface form is the same in different sentences but whose meaning differs. It is a broader concept than "homonym," and can be seen as including the conjugated forms of predicates and the inflected (declined) forms of substantives. Resolving this kind of ambiguity is also called word sense disambiguation (WSD), a long-standing problem in natural language processing.

• For example, "용인" and "트는" differ in meaning as follows.

  Copy

  > 전설속의 용인 이무기의:
     [('용', 'NNG'), ('이', 'VCP'), ('ㄴ', 'ETM')]
  > 용인 제2 캠퍼스를 확대 운영:
     [('용인', 'NNG')]
  
  > 제가 좋아하는 곡을 트는 것이 원칙이지만:
     [('곡', 'NNG'), ('을', 'JKO'), ('틀', 'VV'), ('는', 'ETM')]
  > 동이 트는 아침 시각에:
     [('동', 'NNG'), ('이', 'JKS'), ('트', 'VV'), ('는', 'ETM')]
  

Building the disambiguation evaluation data

A team from the Department of Korean Language and Literature at Seoul National University selected sentences after correcting and reviewing the tagging of ambiguous words. The words in the evaluation data were selected so that the numbers of words with the same form but different meanings are similar. For words with too few surface forms, sentences were written separately.

• 35,396 sentences in total
• 8,285 surface forms (가, 가가, 가거나 ...)

Contributing

Please join us in building the disambiguation evaluation data. Together we can create a smart performance benchmark for natural language processing. The data is not perfect at the moment and may contain errors. If you find an error, please open a pull request.

• When you open a pull request, a Bareun maintainer will review it and merge it.
• If you request an error fix, you will be added to the contributor list.

If you need bigger changes, create an issue and leave a comment in the data repository to discuss.