license: cc-by-4.0
TeSent_Benchmark-Dataset
TeSent_Benchmark-Dataset is a large-scale, open-source benchmark for Telugu sentiment analysis, with a focus on explainability through human-annotated rationales. This dataset is designed for sentence-level sentiment classification and captures the nuanced subjectivity of real-world text via multi-layer annotation.
Overview
- Language: Telugu
- Size: 22,505 sentences
- Sentiment Classes: Positive, Negative, Neutral
- Annotators: Each sentence is independently annotated by 3 annotators
Annotation Protocol
For each sentence, three annotators independently provide:
- Primary Sentiment Label: The main sentiment (Positive, Negative, or Neutral).
- Rationale: Words or phrases from the sentence that justify their primary label.
- Required for positive/negative labels
- Optional for neutral labels (since neutral content may lack explicit cues)
- Secondary Sentiment Label (Optional): If the annotator believes an additional sentiment reasonably coexists, they can specify a secondary label (no rationale required for this).
This protocol enables the dataset to represent ambiguity and subjectivity inherent in natural language sentiment.
Data Format
The dataset is provided as a table with the following columns:
| Column Name | Description |
|---|---|
| Content | The Telugu sentence. |
| Annotations | Three annotators’ sentiment labels for the sentence, separated by | (vertical bar). Each annotation can be a primary label alone (e.g., Positive), or a primary and secondary label joined by a hyphen (e.g., Neutral-Positive). |
| Rationale | The rationale(s) selected by each annotator for their primary label, separated by |. If an annotator chooses multiple rationale words/phrases, they are separated by commas. Rationales are required for positive/negative, optional for neutral. Empty if no rationale. |
| Label | The final label used for modeling and benchmarking (usually the majority or consensus primary label among annotators). |
Example row:
| Content | Annotations | Rationale | Label |
|---|---|---|---|
| పూర్తిగా కావస్తున్న విద్యుత్ కేంద్రాల నిర్మాణం | Neutral-Positive|Positive|Positive | |పూర్తి,కావస్తున్న|పూర్తి,కావస్తున్న | Positive |
Splits
The repository includes both the full dataset and standardized splits for model development:
- Tesent.csv: Contains the complete dataset of 22,505 annotated Telugu sentences.
- train.csv: Stratified training split (80% of the data), used for model training and, together with the validation set, for hyperparameter tuning.
- val.csv: Stratified validation split (10% of the data), intended for hyperparameter tuning and model selection, in conjunction with the train set.
- test.csv: Stratified test split (10% of the data), reserved exclusively for final evaluation and reporting performance metrics.
All splits are created using stratified sampling to preserve the distribution of sentiment classes across partitions. This ensures that each subset is representative of the overall dataset and suitable for robust benchmarking.
Typical Usage:
- Train set: Used for training models and, together with the validation set, for hyperparameter tuning.
- Validation set: Used alongside the train set for hyperparameter tuning and model selection.
- Test set: Used exclusively for final evaluation and reporting performance metrics.
Key Features
- Fine-Grained Explainability: Explicit rationales for every positive/negative annotation.
- Ambiguity Capture: Optional secondary labels reflect subjective or mixed-sentiment content.
- Multi-Annotator Design: All three annotator opinions are provided, supporting robust analysis of inter-annotator agreement and label ambiguity.
- Flexible Usage: Supports explainable sentiment classification, rationale extraction.
- Standard Splits: Predefined train/val/test splits for reproducible experiments and fair comparisons.
Acknowledgements
We thank the annotators and contributors to the TeSent project for their valuable efforts.