rvv-karma/mstds-multi-source-target-dependent-sentiment

MSTDS: Multi-Source Target-Dependent Sentiment Dataset

Overview

MSTDS is a consolidated English-language dataset for target-dependent sentiment classification.

Given a text and a specified target, the task is to predict the sentiment expressed toward that target. MSTDS combines 13 existing datasets covering social media, financial text, news articles, and product or restaurant reviews.

The meaning of target depends on the original dataset. It may refer to a named entity, company, financial instrument, product-related aspect, aspect term, or another annotated target. Therefore, MSTDS is broader than a strictly aspect-based sentiment analysis (ABSA) dataset.

MSTDS is a downstream compilation of existing resources. It is not an official release by, or an endorsement from, the original dataset creators.

Intended uses

MSTDS may be useful for:

• Research on target-dependent sentiment classification.
• Comparing sentiment models across financial text, news, social media, and product-review domains.
• Studying transfer learning between different sentiment domains.
• Exploring entity-level and aspect-level sentiment classification.
• Training models that receive both a text and an explicit target.

Because annotation schemes differ between sources, users should inspect the source and source_metadata_json fields before combining all records for a particular training objective.

Dataset files and splits

• train.jsonl: training examples.
• test.jsonl: held-out evaluation examples.

The original validation and test examples are combined into the final test split during preprocessing. This release therefore does not contain a separate validation split.

Both files contain ground-truth sentiment labels. The test set should be reserved for final evaluation rather than repeatedly used for model selection, threshold tuning, or hyperparameter optimization.

Data schema

Each exported JSONL record has the following fields:

Field	Description
text	Input text
target	Target associated with the sentiment label
sentiment	positive, negative, neutral, or conflict
source	Identifier of the originating dataset
split	train or test
source_metadata_json	JSON-encoded source-specific metadata
extra_fields_json	JSON-encoded additional fields, including provenance where available

The conflict label is kept as a separate class where it occurs in the source annotations. It should not automatically be treated as equivalent to neutral.

The metadata columns are encoded as JSON strings to provide a consistent tabular schema while preserving source-specific information.

To decode these fields in Python:

import json

metadata = json.loads(example["source_metadata_json"])
extra_fields = json.loads(example["extra_fields_json"])

Data processing

The preprocessing pipeline used to prepare the consolidated corpus includes:

1. Validating required fields, sentiment labels, and split names.
2. Removing text-target pairs that have conflicting sentiment labels.
3. Collapsing duplicate normalized text-target-sentiment records.
4. Combining validation and test into one held-out test split.
5. Checking normalized text for overlap between train and test.
6. Preserving available source metadata and provenance.

Unicode normalization, case folding, and whitespace normalization are used for duplicate and overlap checks. These operations cannot detect every paraphrase or semantic near-duplicate.

Please consult the manifest and the actual files for the statistics of the specific release being used.

Acknowledgements and upstream citations

We gratefully acknowledge the original dataset authors, annotators, maintainers, and research teams whose work made this compilation possible.

MSTDS combines or derives examples from the following upstream resources. Users of MSTDS should also cite the relevant original publications and follow each resource's attribution and usage requirements.

1. Dong et al. (2014) — Target-Dependent Twitter Sentiment

Paper: Li Dong, Furu Wei, Chuanqi Tan, Duyu Tang, Ming Zhou, and Ke Xu. 2014. Adaptive Recursive Neural Network for Target-dependent Twitter Sentiment Classification. Proceedings of ACL 2014, pages 49–54.

• Paper: https://aclanthology.org/P14-2009/
• Original data resource: https://drive.google.com/file/d/0B8yp1gOBCztyVVVoLTdNZ1JHYVU/view

Included file: dong2014_target_dependent_twitter.jsonl

2. FinEntity

Paper: Yixuan Tang, Yi Yang, Allen Huang, Andy Tam, and Justin Tang. 2023. FinEntity: Entity-level Sentiment Classification for Financial Texts. Proceedings of EMNLP 2023, pages 15465–15471.

• Paper: https://aclanthology.org/2023.emnlp-main.956/
• Dataset and code: https://github.com/yixuantt/FinEntity

Included file: finentity.jsonl

3. FiQA Financial Sentiment / ABSA

This collection includes examples prepared from the FiQA financial sentiment resource.

• Dataset resource: https://huggingface.co/datasets/TheFinAI/fiqa-sentiment-classification
• FiQA challenge paper: https://doi.org/10.1145/3184558.3192301

Included file: fiqa_absa.jsonl

Please follow the citation, access, and licensing instructions for the exact upstream version used to prepare this file.

4. MAMS — Multi-Aspect Multi-Sentiment

Paper: Qingnan Jiang, Lei Chen, Ruifeng Xu, Xiang Ao, and Min Yang. 2019. A Challenge Dataset and Effective Models for Aspect-Based Sentiment Analysis. Proceedings of EMNLP-IJCNLP 2019, pages 6280–6285.

• Paper: https://aclanthology.org/D19-1654/
• Dataset and code: https://github.com/siat-nlp/MAMS-for-ABSA

Included file: mams.jsonl

5. NewsMTSC

Paper: Felix Hamborg and Karsten Donnay. 2021. NewsMTSC: A Dataset for (Multi-)Target-dependent Sentiment Classification in Political News Articles. Proceedings of EACL 2021, pages 1663–1675.

• Paper: https://aclanthology.org/2021.eacl-main.142/
• Dataset and code: https://github.com/fhamborg/NewsMTSC
• Hugging Face dataset: https://huggingface.co/datasets/fhamborg/NewsMTSC

Included file: newsmtsc.jsonl

6. PerSenT — Author's Sentiment Prediction

Paper: Mohaddeseh Bastan, Mahnaz Koupaee, Youngseo Son, Richard Sicoli, and Niranjan Balasubramanian. 2020. Author's Sentiment Prediction. Proceedings of COLING 2020, pages 604–615.

• Paper: https://aclanthology.org/2020.coling-main.52/
• Dataset and code: https://github.com/StonyBrookNLP/PerSenT

Included file: persent.jsonl

7. SemEval-2016 Task 5 — Aspect-Based Sentiment Analysis

Paper: Maria Pontiki et al. 2016. SemEval-2016 Task 5: Aspect Based Sentiment Analysis. Proceedings of SemEval 2016, pages 19–30.

• Paper: https://aclanthology.org/S16-1002/
• Official task information: https://alt.qcri.org/semeval2016/task5/index.php?id=data-and-tools
• Dataset mirror: https://huggingface.co/datasets/alexcadillon/SemEval2016Task5

Included file: semeval2016_task5.jsonl

8. SemEval-2017 Task 5 — Financial Sentiment

Paper: Keith Cortis, André Freitas, Tobias Daudert, Manuela Huerlimann, Manel Zarrouk, Siegfried Handschuh, and Brian Davis. 2017. SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Microblogs and News. Proceedings of SemEval 2017, pages 519–535.

• Paper: https://aclanthology.org/S17-2089/
• Official task information: https://alt.qcri.org/semeval2017/task5/

Included file: semeval2017_task5_financial_sentiment.jsonl

9. SemEval-2014 Task 4 — Laptops

Paper: Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. SemEval-2014 Task 4: Aspect Based Sentiment Analysis. Proceedings of SemEval 2014, pages 27–35.

• Paper: https://aclanthology.org/S14-2004/
• Dataset mirror: https://huggingface.co/datasets/tomaarsen/setfit-absa-semeval-laptops

Included file: semeval_2014_task_4_laptops.jsonl

10. SemEval-2014 Task 4 — Restaurants

Paper: Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. SemEval-2014 Task 4: Aspect Based Sentiment Analysis. Proceedings of SemEval 2014, pages 27–35.

• Paper: https://aclanthology.org/S14-2004/
• Dataset mirror: https://huggingface.co/datasets/tomaarsen/setfit-absa-semeval-restaurants

Included file: semeval_2014_task_4_restaurants.jsonl

11. SENTiVENT

Dataset: Gilles Jacobs. 2020. SENTiVENT Bench: Economic Events and Implicit Sentiment in Financial News. Version 1.1, OSF.

• Dataset: https://doi.org/10.17605/OSF.IO/8JEC2
• Hugging Face resource: https://huggingface.co/datasets/GillesJacobs/sentivent
• Event extraction methodology: https://doi.org/10.1007/s10579-021-09562-4
• Sentiment methodology: https://doi.org/10.3390/electronics10202554

Included file: sentivent.jsonl

The upstream licence covers specified SENTiVENT materials but excludes third-party source news material. See the source licence before redistribution.

12. SuperTweetEval — TweetSentiment Subset

Paper: Dimosthenis Antypas, Asahi Ushio, Francesco Barbieri, Leonardo Neves, Kiamehr Rezaee, Luis Espinosa-Anke, Jiaxin Pei, and Jose Camacho-Collados. 2023. SuperTweetEval: A Challenging, Unified and Heterogeneous Benchmark for Social Media NLP Research. Findings of EMNLP 2023, pages 12590–12607.

• Paper: https://aclanthology.org/2023.findings-emnlp.838/
• Dataset: https://huggingface.co/datasets/cardiffnlp/super_tweeteval

Included file: supertweeteval_tweetsentiment.jsonl

Follow the benchmark's instructions and any terms applicable to the underlying source task.

13. Sentiment of Tweets by Brand or Product

The source used for this file is the public db495/phase_4_project repository. Its README describes a tweet dataset sourced through CrowdFlower and points to the tweets_dataset.csv resource.

• Repository and source notes: https://github.com/db495/phase_4_project
• CSV used in preparation: https://raw.githubusercontent.com/db495/phase_4_project/main/Data/tweets_dataset.csv

Included file: tweets_by_brand.jsonl

The repository is cited here as the source used during preparation. Verify the original provider's licensing and attribution requirements before redistribution.

Licensing and redistribution

MSTDS does not claim a single licence for all constituent datasets.

Licences, access conditions, copyright, redistribution permissions, and citation requirements vary across sources. The public availability of a paper, repository, or dataset mirror does not automatically grant permission to redistribute every underlying text.

Before publishing or using this compilation commercially:

• Review the current licence and terms for each constituent dataset.
• Follow its required scholarly citation and attribution instructions.
• Check restrictions concerning tweets and third-party news or article text.
• Confirm that redistribution of the actual text is permitted for each source.
• Preserve required notices and indicate modifications where required.

FinEntity identifies ODC-BY as its licence. SENTiVENT documents the scope of its licence and exclusions for third-party source content. Other sources may have different or additional requirements.

This dataset card does not override any upstream licence or terms of use. The original authors and maintainers have not necessarily reviewed or endorsed MSTDS.

Limitations

• Target semantics and annotation schemes differ across sources.
• Some sources focus on aspects, while others focus on companies, financial entities, document-level targets, or other annotated targets.
• Sentiment conversion and split construction may differ from the original tasks.
• The meaning and prevalence of the conflict class can vary by source.
• Duplicate and leakage checks cannot guarantee the absence of every paraphrase, near-duplicate, or semantic overlap.
• MSTDS should not be treated as a brand/product-only sentiment dataset without filtering according to target semantics and source.

Citation

When using MSTDS, cite this repository and the exact version used, together with the relevant original publications for the constituent datasets.

MSTDS is a downstream compilation. It should not be cited as the original creator of the constituent annotations.

Acknowledgements

We thank the original dataset authors, annotators, maintainers, and research teams for making their research resources available. All source-specific attribution and usage requirements continue to apply to the corresponding data in MSTDS.

Resource

Colab Notebook