Developing Models for Identifying Inconsistent Declarations in Training Data
Amazon has developed a dataset designed to improve product retrieval systems by identifying counterfactual statements in customer reviews, written in English, German, and Japanese. The dataset, although not explicitly labeled for counterfactual statement identification, can serve as a starting point for researchers to create such resources.
Accessing the Dataset
Amazon Customer Review Data (Multi-lingual)
Amazon offers publicly available datasets of customer reviews in multiple languages via the Amazon Product Review Dataset, accessible on platforms like AWS and public repositories such as the “Amazon Reviews” dataset on the AWS Open Data Registry or Amazon’s own data dumps. These datasets include multiple languages, making them suitable for this purpose.
To find these datasets, search on AWS Open Data or Kaggle.
Processing for Counterfactual Identification
The datasets typically contain raw review texts and ratings without explicit counterfactual labels. To build a counterfactual statement identification dataset, researchers curate or annotate the data based on specific linguistic or contextual criteria, such as sentences expressing "what-if" scenarios or hypothetical alternatives.
Recent research has involved leveraging large language models (LLMs) to filter and annotate such data for counterfactual explanations in recommendation systems based on the Amazon review corpus. However, direct dataset access details for this counterfactual annotation are not public.
Access Requirements on AWS
If you intend to access the raw Amazon review datasets via AWS services like Amazon SageMaker or AWS Glue, you will need to configure appropriate IAM permissions for data access and processing, as described in AWS documentation. These permissions control your ability to load, process, and analyze Amazon’s datasets stored in AWS.
Further Steps
- Check the official Amazon Customer Review Dataset or Kaggle for available review datasets with multiple languages.
- For counterfactual labels, you may need to build the dataset yourself or reach out to authors of recent academic papers to request access if they provide the annotated data.
- Use LLMs or annotation frameworks to extract counterfactual statements from the raw review texts for your specific languages of interest.
While a pre-annotated counterfactual statement dataset from Amazon is not publicly documented as of 2025, the steps outlined above can help you create or access a dataset for counterfactual statement identification in Amazon customer reviews, paving the way for more accurate and reliable product retrieval systems.
Researchers can utilize the Amazon Customer Review Dataset, available on platforms like AWS and public repositories such as the “Amazon Reviews” dataset on the AWS Open Data Registry or Amazon’s own data dumps, to process data for counterfactual identification. This dataset, which includes multiple languages, can serve as a starting point for researchers to create a counterfactual statement identification dataset using large language models (LLMs) or annotation frameworks.
