I have documents with reviews of articles that have the following structure:

  • Introduction: a description of the review, dates and metadata that will be discarded. (avg~180 words, std~30 words)
  • Loop with:
    • Reference: the reference article to review. A single review can talk about multiple articles. (avg~8 words, std~2 words)
      • Subject: topic of the following comments. There can be more than one subject for the following comments. (avg~7 words, std~3 words)
        • Comment: atomic comment about the article. There can be multiple comments per subjects. (avg~45 words, std~30 words)
  • Final: a brief summary of the review and other info that will be discarded. (avg~165 words, std~40 words)

I've to identify orderly the different chunks of references, subjects and comments. I've already designed a pipeline, and have a dataset to test and train my models, I'm using BIO tags for chunking as output, but I can adapt the pipeline to use BMEWO (BILOU).

I'm searching for a similar problem dataset, to compare how my pipeline is performing against the best metrics accomplished in that problem.

I've already searched in:



Until now the closest dataset I've found is the CoNLL for named entity recognition or grammar tagging(noun phrases, prepositional phrases, verb phrases).

But that problems seems very different from mine:

  • Small length per chunk in relation with my problem.
  • The length of every chunk class are similar, In my case the size of the comments are very different than the subjects and references.
  • The outside tag could be everywhere while in my problem only can be at the beginning or the end.
  • These problems seems more focus on short grammatic structure instead of long sequence meaning.

I'm searching with the following keywords: chunk, long-sequence, text. I'm not sure if the word chunk could be misleading.

Where can I keep searching datasets? What keywords can I use to improve my search? There is a similar problem in which I could try to transform mine to test against?