--- language: - 'no' - nb - nn license: cc-by-nc-4.0 size_categories: - 10K - **Repository:** https://github.com/ltgoslo/norec_tsa - **Paper:** [A Fine-Grained Sentiment Dataset for Norwegian](https://aclanthology.org/2020.lrec-1.618/) ### Dataset Summary The dataset contains tokenized Norwegian sentences where each token is tagged for sentiment expressed towards that token. The dataset is derived from the manually annotated [NoReC_fine](https://github.com/ltgoslo/norec_fine) with rich annotations for each sentiment expression in the texts. The texts are a subset of the Norewegian Review Corpus [NoReC](https://github.com/ltgoslo/norec). ### Supported Tasks and Leaderboards [NorBench](https://github.com/ltgoslo/norbench) provides TSA evaluation scripts using this dataset, and a leaderboard comparing large language models for downstream NLP tasks in Norwegian. ### Languages Norwegian: Predominantly Bokmål written variant. | variant | split | sents | docs | |:-----|:--------|--------:|-------:| | nb | dev | 1531 | 44 | | nb | test | 1272 | 47 | | nb | train | 8556 | 323 | | nn | train | 78 | 4 | ## Fine-tuned models [`ltg/norbert3-large_TSA`](https://huggingface.co/ltg/norbert3-large_TSA) ## Dataset Structure The dataset comes in two flavours: - `default` configuration yields labels with binary Positive / Negative sentiment description - `intensity` configuration yields labels with additional sentiment intensity, 1: Slight, 2: Standard, and 3: Strong. The config is required for accessing the version with intensity. `tsa_data = load_dataset("ltg/norec_tsa", "intensity")` The dataset comes with predefined train, dev (vallidation) and test splits. ### Data Instances Config "default" example instance: ``` {'idx': '701363-08-02', 'tokens': ['Vi', 'liker', 'det', '.'], 'tsa_tags': ['O', 'O', 'B-targ-Positive', 'O']} ``` Config "intensity" example instance: ``` {'idx': '701363-08-02', 'tokens': ['Vi', 'liker', 'det', '.'], 'tsa_tags': ['O', 'O', 'B-targ-Positive-2', 'O']} ``` ### Data Fields - idx(str): Unique document-and sentence identifier from [NoReC_fine](https://github.com/ltgoslo/norec_fine). The 6-digit document identifier can also be used to look up the text and its metadata in [NoReC](https://github.com/ltgoslo/norec). - tokens: (List[str]): List of the tokens in the sentence - tsa_tags: (List[str]): List of the tags for each token in BIO format. There is no integer representation of these in the dataset. ### Data Splits ``` DatasetDict({ test: Dataset({ features: ['idx', 'tokens', 'tsa_tags'], num_rows: 1272 }) train: Dataset({ features: ['idx', 'tokens', 'tsa_tags'], num_rows: 8634 }) validation: Dataset({ features: ['idx', 'tokens', 'tsa_tags'], num_rows: 1531 }) }) ``` ## Dataset Creation ### Curation Rationale The sentiment expressions and targets are annotated in NoReC_fine according to its [annotation guidelines](https://github.com/ltgoslo/norec_fine/blob/master/annotation_guidelines/guidelines.md) Since a sentiment target may be the target of several sentiment expressions, these are resolved to a final sentiment polarity (and intensity) using the conversion script in [NoReC_tsa](https://github.com/ltgoslo/norec_tsa). There is no "mixed" sentiment category. When a target is the receiver of both positive and negative sentiment, the strongest wins. If a tie, the last sentiment wins. ### Source Data A subset of the Norwegian Review Corpus with its sources and preprocessing described [here](https://github.com/ltgoslo/norec). ### Discussion of Biases The professional review texts in NoReC that NoReC_tsa is a subset from, are from a set number of Norwegian Publishing channels and from a set timespan that can be explored in the [NoReC metadata](https://raw.githubusercontent.com/ltgoslo/norec/master/data/metadata.json). Both language usage and sentiments expressed could have been more diverse with a more diverse set of source texts. ### Licensing Information The data, being derived from [NoReC](https://github.com/ltgoslo/norec), 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. ### Citation Information ```bibtex @InProceedings{OvrMaeBar20, author = {Lilja Øvrelid and Petter Mæhlum and Jeremy Barnes and Erik Velldal}, title = {A Fine-grained Sentiment Dataset for {N}orwegian}, booktitle = {{Proceedings of the 12th Edition of the Language Resources and Evaluation Conference}}, year = 2020, address = {Marseille, France, 2020} } ```