OpinionRelationCorpus

Introduction

• The test set in paper "Large-scale Opinion Relation Extraction with Distantly Supervised Neural " [2017 EACL].

• Reviews are from Amazon datasets (4 domains: Phone, Movie, Food and Pet).

• The annotation tool is brat, and the data follows brat's format.

Format

Both the raw text (sentence split and word tokenized) and annotations are provied for each review:

• review_%d.txt: a raw review text. Each row is a sentence.

• review_%d.ann: an annotation file for review_%d.txt:

  • Each row is an annotation, which could be an opinion target, an opinion expression or a relation.

  • An opinion target row (opinion expression is similar) contains three fields (separated by tab): T%d``\t``OpinionTarget left right``\t``RawString.

    • T%d: id of the target.
    • OpinionTarget left right: offsets in the plain text.
    • RawString: raw string of the opinion target.

  • A relation row contains two fields:

    • R%d: id of the relation.
    • Arg1:T%d: id of the opinion expression.
    • Arg2:T%d: id of the opinion target.