README.md

metadata

license: apache-2.0
language:
  - en
pretty_name: imdb-single-sentence
size_categories:
  - 10K<n<100K
source_datasets:
  - stanfordnlp/imdb
task_categories:
  - text-classification

We maximized the information within the multi-sentence input reviews of the original IMDb dataset (stanfordnlp/imdb) and transformed it into a single-sentence format. This transformation aims to make benchmark studies more compatible with datasets containing single-sentence inputs (SST-2, HateSpeech, Tweet-Emotion, etc.)

Tasks we performed to obtain this dataset:

• (1) We fine-tuned the RoBERTa model (roberta-base) on the raw IMDb dataset, achieving a classification accuracy of 94.6% for sentiment analysis. We published this fine-tuned sentiment classifier on HuggingFace under EgehanEralp/roberta-base-imdb-ft for public access.
• (2) We split the sentences in each train and test sample of the IMDb dataset's input reviews.
• (3) For each sentence in every review, we queried our fine-tuned RoBERTa sentiment classifier model to obtain sentiment predictions.
• (4) For positive reviews, we selected the sentences with the highest confidence positive label predictions by the model, retaining these sentences within the original multi-sentence reviews and deleting all other sentences.
• (5) For negative reviews, we selected the sentences with the highest confidence negative label predictions by the model, retaining these sentences within the original multi-sentence reviews and deleting all other sentences.

We created a single-sentence IMDb dataset, representing each multi-sentence review with the sentence containing the most sentiment. This strategy allowed us to select and use the sentences that best represented the sentiment of the original reviews.