TAPAS is a neural network model that uses an extended BERT architecture to answer natural language questions by directly querying tabular data.
TAPAS (Table Parsing via Pre-training) is a neural network model developed by Google Research that extends the BERT architecture to jointly encode natural language questions and tabular data structures. This allows the model to directly identify answers within tables, including single cells, multiple cells, or aggregated results, without needing to translate questions into SQL-like queries. The model is designed to work across a wide range of table domains, moving beyond approaches that are specific to certain table styles. The model processes questions by encoding them along with table content row by row, using special embeddings for column index, row index, and a numeric rank index. It outputs a probability score for each table cell being part of the answer and an aggregation operation (e.g., SUM, COUNT, AVERAGE) to apply. TAPAS was pre-trained on millions of Wikipedia tables and has demonstrated competitive accuracy on academic table question-answering datasets like SQA, WikiTableQuestions, and WikiSQL. The code for training and testing, along with pre-trained models, has been open-sourced on GitHub to facilitate further research.
Platforms: Web
View full TAPAS (Table Parsing via Pre-training) profile on Tools-Radar | Browse Data Analysis tools | Alternatives to TAPAS (Table Parsing via Pre-training)
Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com