Dataset Card for NoReC_document

Document-level polarity classification of Norwegian full-text reviews across mixed domains.

Dataset Details

This is a dataset for document-level sentiment classification in Norwegian, derived from the Norwegian Review Corpus: NoReC. We here provide two simplified versions of NoReC where the original six-point numerical ratings have been mapped to a reduced set of categorical classes: positive and negative for the binary version, and positive, fair, and negative for the larger ternary version (more details below). This reduces the problem of class imbalance and data scarcity when trying to predict the six-point ratings directly.

• Curated by: The underlying NoReC data was created as part of the SANT project (Sentiment Analysis for Norwegian Text), coordinated by the Language Technology Group (LTG) at the University of Oslo, in collaboration with the Norwegian Broadcasting Corporation (NRK), Schibsted Media Group and Aller Media.
• Funded by: The SANT project is funded by the Research Council of Norway (NFR grant number 270908).
• Shared by: The SANT project (Sentiment Analysis for Norwegian Text) at the Language Technology Group (LTG) at the University of Oslo
• Language(s) (NLP): Norwegian (Bokmål and Nynorsk)
• License: The data is distributed under a Creative Commons Attribution-NonCommercial licence (CC BY-NC 4.0), access the full license text here: https://creativecommons.org/licenses/by-nc/4.0/ The licence is motivated by the need to block the possibility of third parties redistributing the orignal reviews for commercial purposes. Note that machine learned models, extracted lexicons, embeddings, and similar resources that are created on the basis of NoReC are not considered to contain the original data and so can be freely used also for commercial purposes despite the non-commercial condition.
• Paper: A paper by Velldal et al. at LREC 2018 describes the (initial release of the) underlying NoReC data in more detail.

Uses

The data is intended to be used for training and testing models for Norwegian document-level classification of polarity, either binary (positive / negative) or ternary (positive / fair / negative).

Dataset Structure

Each instance in the data has the following three fields:

• 'id': a uniqe document identifier
• 'review': the actual review text
• 'sentiment': a numerical class label indicating polarity, which can have one of three values:
  • 0 = negative
  • 1 = positive
  • 2 = fair (only available in the ternary version)

In terms of the mapping from the six-point numerical ratings in the underlying NoReC dataset, 'negative' (0) here corresponds to ratings of 1–3, 'positive' (0) corresponds to ratings of 5–6, while 'fair' corresponds to ratings of 4. In the binary version of NoReC_document, reviews with a rating of 4 are excluded.

The data comes with pre-defined train/dev/test splits, inherited from NoReC.

Source Data

The document-level labels are derived from the Norwegian Review Corpus (NoReC), which contains over 43K full-text professional reviews collected from major Norwegian news sources and cover a range of different domains, including literature, movies, video games, restaurants, music and theater, in addition to product reviews across a range of categories. The review articles NoReC were originally donated by the media partners in the SANT project; the Norwegian Broadcasting Corporation (NRK), Schibsted Media Group and Aller Media. The data comprises reviews extracted from eight different Norwegian news sources: Dagbladet, VG, Aftenposten, Bergens Tidende, Fædrelandsvennen, Stavanger Aftenblad, DinSide.no and P3.no. In terms of publishing date the reviews of NoReC mainly cover the time span 2003–2019, although it also includes a handful of reviews dating back as far as 1998.

The numerical ratings of the underlying NoReC dataset were provided by the original review authors, i.e. these are the ratings assigned by professional journalists in the published versions of the review articles.

Personal and Sensitive Information

The data does not contain information considered personal or sensitive.

Bias, Risks, and Limitations

Results obtained on this data might not generalize to texts from other domains or genres. Any biases in the sentiments expressed by the original review authors may carry over to models trained on this data.

Citation

@InProceedings{VelOvrBer18,
  author = {Erik Velldal and Lilja {\O}vrelid and Eivind Alexander Bergem and  Cathrine Stadsnes and Samia Touileb and Fredrik J{\o}rgensen},
  title = {{NoReC}: The {N}orwegian {R}eview {C}orpus},
  booktitle = {Proceedings of the 11th edition of the 
               Language Resources and Evaluation Conference},
  year = {2018},
  address = {Miyazaki, Japan},
  pages = {4186--4191}
}

Dataset Card Authors

Vladislav Mikhailov and Erik Velldal

Dataset Card Contact

vladism@ifi.uio.no and erikve@ifi.uio.no