janduplessis886/gp_surgery_reviews_fake_and_real

GP Surgery Reviews Dataset Data Card

Overview

This dataset consists of GP Surgery reviews designed for binary classification tasks. It includes both real and fake reviews, where the label feature marks real reviews as 0 and fake reviews as 1. The fake reviews were generated using DeepSeek LLM (Ollama) and then passed through a processing pipeline to derive additional features.

Dataset Composition

• —Total Records: 9,974

Features:

• —free_text: The original text of the GP Surgery review.
• —sentiment_free_text: The sentiment analysis result derived from the review text.
• —sentiment_score_free_text: A numerical sentiment score computed from the review text.
• —freetextlen: The length of the review text, measured in tokens or characters.
• —rating: The rating associated with the review (if provided).
• —feedback_labels: Additional feedback labels derived from the content.
• —emotionfreetext: Emotion analysis results computed on the review text.
• —label: Binary classification label: 0 for real reviews and 1 for fake reviews.
• —abs_sentiment: The absolute sentiment score for the review.
• —generation_category: A category describing the generation process of the review, which may help differentiate between various types of generated content.

Data Sources and Collection Process

• —Real Reviews: Collected from genuine GP Surgery reviews. vFake Reviews: Generated using DeepSeek LLM (Ollama). Once generated, these fake reviews underwent a post-processing pipeline that computed additional features such as sentiment, length, and emotion. vPost-Processing Pipeline: The pipeline applied various natural language processing (NLP) techniques to derive extra features like sentiment analysis, token count, and emotion classification, adding layers of information beyond the raw review text.

Intended Use

This dataset is primarily intended for:

• —Binary Classification: Training and evaluating models that distinguish between real and machine-generated (fake) reviews.
• —Fake Review Detection: Research into methods for detecting artificially generated content in review systems.
• —NLP Research: Exploring sentiment analysis, emotion detection, and other feature engineering approaches on real-world text data.

Preprocessing and Feature Engineering

• —Text Processing: The free_text field can be tokenized and preprocessed according to the requirements of your chosen NLP model.
• —Sentiment and Emotion Analysis: The sentimentfreetext, sentimentscorefreetext, abssentiment, and emotionfreetext fields were computed using established sentiment and emotion analysis tools.
• —Feature Generation: Additional features such as freetextlen, rating, feedbacklabels, and generationcategory provide more context and are valuable for downstream tasks like classification and regression.

Limitations and Considerations

• —Dataset Size: With 9,974 records, the dataset is moderately sized. Depending on your model’s complexity, you might need to consider data augmentation or additional data sources. vRepresentation of Fake Reviews: The fake reviews were generated by a specific LLM (DeepSeek LLM by Ollama) and might not encompass the full spectrum of potential machine-generated text. Models trained solely on this data may perform differently on fake reviews generated by other systems.
• —Biases Introduced: Be aware of any biases that might have been introduced during the data collection and post-processing stages.