Employees who can work well with Excel spreadsheets are highly valued. But Microsoft researchers have developed a solution that can at least partially replace them. SpreadsheetLLM is a large language model designed to manipulate spreadsheets using natural language commands.
Image source: Rubaitul Azad / unsplash.com
SpreadsheetLLM analyzes and interprets data in spreadsheets using AI, solving most of the problems associated with it – for this, data is serialized, that is, addresses, values and cell formats are included in the stream. The tool contains a SheetCompressor component that compresses spreadsheets for transmission to the AI model. It consists of three modules: the first analyzes the table structure and discards non-table contents; the second transforms the data into a more efficient representation; the third aggregates data.
SpreadsheetLLM has some limitations. It ignores cell colors, which may have meaning, and does not perform semantic compression on cell content expressed in natural language. But this is enough to reduce the consumption of tokens when requesting AI by 96%, which means saving computing resources. As a result, users without proper technical training can send requests to SpreadsheetLLM in natural language and achieve their goals. But the main goal of the project is not to replace a person, but to provide him with assistance in finance, accounting and other areas related to data processing.
The model includes the Chain of Spreadsheet (CoS) framework for analyzing the contents of multiple tables. SpreadsheetLLM can work with structured and unstructured spreadsheet data, an aspect the researchers indicate can reduce incidents of “hallucinations” in AI responses. The project is currently at the research stage and is not yet ready for release as a commercial product.