HebrewSentiment - A Sentiment-Analysis Dataset in Hebrew

Summary

HebrewSentiment is a Hebrew dataset for the sentiment analysis task.

Introduction

This dataset was constructed via [To Fill In].

Dataset Statistics

The table below shows the number of examples from each category in each of the splits:

split	total	positive	negative	neutral
train	39,135	8,968	7,669	22,498
test	2,170	503	433	1,234

Dataset Description

Each row in the dataset contains the following fields:

• id: A unique identifier for that training examples
• text: The textual content of the input sentence
• tag_ids: The label of the example (Neutral/Positive/Negative)
• task_name: [To fill in]
• campaign_id: [To fill in]
• annotator_agreement_strength: [To fill in]
• survey_name: [To fill in]
• industry: [To fill in]
• type: [To fill in]

Models and Comparisons

In collaboration with DICTA we trained a model on this dataset and are happy to release it to the public: DictaBERT-Sentiment.

In addition, we compared the performance of the model to the previous existing sentiment dataset - Hebrew-Sentiment-Data from OnlpLab. We fine-tuned dictabert 3 times - once on the OnlpLab dataset, once on this dataset, and once on both datasets together and the results can be seen in the table below:

Training Corpus:	OnlpLab					HebrewSentiment
	Accuracy	Macro F1	F1 Positive	F1 Negative	F1 Neutral	Accuracy	Macro F1	F1 Positive	F1 Negative	F1 Neutral
OnlpLab+HebrewSentiment	87	61.7	93.2	74.6	17.4	83.9	82.7	79.8	81.8	86.4
OnlpLab	88.2	63.3	93.8	72.1	24	41.3	42.2	48.1	56.3	22.2
HebrewSentiment	69.9	51.7	82.2	62.9	10.2	84.4	83.2	81	82.1	86.6

Contributors

[To fill in]

Contributors: [To fill in]

Acknowledgments

We would like to express our gratitude to [To fill in]