Microsoft researchers are teaching AI to read spreadsheets

Getting generative AI models to understand spreadsheets can be difficult. To address this problem, Microsoft researchers published a paper on July 12 on Arxiv describing a method that Master of Laws (LLM) in Electronic Formsan encoding framework that enables large language models to “read” spreadsheets.

SpreadsheetLLM could “transform spreadsheet data management and analysis, paving the way for smarter and more efficient user interactions,” the researchers wrote.

One advantage of SpreadsheetLLM in the business world is that formulas in spreadsheets can be used by asking questions to the AI ​​model in natural language, without having to learn how to use them.

Why are spreadsheets a challenge for an LLM?

Spreadsheets are a challenge for LLM students for the following reasons.

• Spreadsheets can be very large, exceeding the number of characters that LLM can process at one time.
• The report said spreadsheets are “two-dimensional in layout and structure” rather than the “linear and sequential input” that LLMs excel at.
• LL.M.s are not typically trained to interpret cell addresses and specific spreadsheet formats.

Microsoft researchers use multi-step technique to parse spreadsheets

SpreadsheetLLM has two main parts:

• Sheet Compressora framework for reducing spreadsheets to a format that LLMs can understand.
• Spreadsheet Chaina method of teaching LLMs how to identify the correct part of a compressed spreadsheet to “look at” when faced with a problem and generate an answer.

SheetCompressor has three modules:

• Structural anchors help LLM identify the rows and columns in your spreadsheet.
• A method to reduce the number of tokens required for LLM to interpret a spreadsheet.
• A technique that increases efficiency by grouping similar cells together.

Using these modules, the team reduced the tokens needed to encode spreadsheets by 96%. This in turn enabled another leading research team to achieve a slight improvement (12.3%) in helping LLMs understand spreadsheets. The researchers tried their spreadsheet recognition approach on these LLMs:

• OpenAI GPT-4 and GPT-3.5.
• Meta’s Llama 2 and Camel 3.
• Microsoft’s Phi-3.
• Mistral-v2 by Mistral AI.

For the spreadsheet chaining feature, they used GPT-4.

What does SpreadsheetLLM mean for Microsoft’s AI efforts?

Microsoft’s obvious advantage in this regard is that it can make its AI Assistant Co-PilotWorks with many Microsoft 365 suite apps to do more in Excel. SpreadsheetLLM represents a continued effort to make generative AI practical—opening up Excel to people who haven’t been trained on more advanced features could be a great area for generative AI to expand.

Look at your business and Microsoft Copilot This will affect which version (if any) is appropriate for your job.

Practical applications and next steps for Microsoft Research

This achievement represents a 12.3% improvement over the results of the previous leading research team and currently has more academic significance than economic significance. Generative AI is notorious for making things upand the illusions that occur in spreadsheets can render large amounts of data useless. As the researchers point out, having an LLM understand the format of a spreadsheet—what a spreadsheet typically looks like and how it works—is different from having an LLM produce understandable, accurate data in those cells.

Furthermore, this approach requires a lot of computing power and multiple reviews by the LLM to arrive at an answer. Alternatively, your office Excel wizard may be able to come up with the answer in a few minutes while consuming almost no energy.

Going forward, the team hopes to find a way to encode details such as the color of a cell’s background and to deepen LLM’s understanding of how words relate to each other within a cell.

TechRepublic has reached out to Microsoft for more information.