To make an Excel table, it is really enough to move your mouth!
Look, enter what you want to do: rank the students’ grades.
Simply press Enter, and the form will be listed in one click!
It’s okay to check again.
It can also be processed across tables.
For example, mark the top ten schools in two different tables.
Oh, it’s a little bit beyond ChatGPT?
After all, if I make a similar request to ChatGPT, it can only help me write the corresponding code, and I have to do it manually.
This is the AI tool ChatExcel that has recently attracted attention on Zhihu, and it was on the hot list as soon as it was released.
It focuses on implementing complex Excel operations (Chat-to-Excel) in the form of dialogue, saying goodbye to memorizing functions and manually setting formulas.
The finished product supports one-click export into a standard Excel table, which saves even the process of copying and pasting.
Moreover, there is no need to register, it can be used on the webpage, and the number of uses is unlimited, which is quite good for lazy people and Excel beginners.
Even people who know Excel skills feel that what they know is going to be useless.
This is a project independently developed by 3 master and doctoral students from Peking University Shenzhen Graduate School/School of Information Engineering.
Demo is currently open for everyone to try.
Seeing this, many people may think, “This is another project developed under the influence of ChatGPT.”
Really not.
The first commit on the GitHub project was submitted in March 22.
And the real opportunity behind it——
In fact, it is a little brother in the team who wants to use technology to help female tickets solve some minor troubles at work.
What the hell is going on? Qubit contacted the main creative team to reveal the secrets to everyone~
How does ChatExcel perform?
Cough cough, we will talk about other women’s votes later, let’s take a look at how ChatExcel performs.
Generally speaking, what it wants to do is to skip learning Excel, skip writing function formulas, and directly turn the needs in its mind into the content in the table.
When the team knew about the first product, it used a picture to explain:
The old rules, let’s get started and experience it.
After entering the main page, the website displays the demo by default, and you can upload your own form by clicking UPLOAD. The format needs to be a standard Excel file.
We asked ChatGPT next door to generate a student report card, and let ChatExcel perform calculations.
Mainly tested Excel operations that require insert functions. for example:
Those with a total score of 255 and above are marked as A, those with a total score of 240-255 are marked as A-, and those with a total score of 210-240 are marked as B.
After typing the content, hit enter, and it will start processing.
What if the teacher wants to know who is the first in a single subject? enter:
Find the person with the highest math score in the table.
Then, it generated a new interface with only the highest-scoring children’s shoes in it.
If you want to perform other operations on the form, click “UNDO” on the page, and it will return to the form before the operation.
We also let it rank the students in the table and list the rankings, and ChatExcel did not make mistakes.
In addition, it can complete tasks continuously.
For example, if you want to add a column of average scores to the table, you can first ask it to add a column:
Then enter the requirements further and let it fill in the calculated results:
Click “DOWNLOAD” to export the final form to an ordinary Excel form.
After experiencing it, I feel that ChatExcel has two main features:
coherence
accuracy
For example, there was also a tool to help people generate Excel formulas before. excelformulabot can realize this function, but it is independent of the table, and you need to bring the cells into the given formula yourself.
If it is a more specific description, there will be situations where the formula cannot be given.
Compared with ChatGPT, the accuracy of ChatExcel calculation is more prominent.
For example, when calculating the average number of the same table, ChatGPT either made a mistake about how many people there are, or it included the wrong number when calculating, and repeated corrections will not save it…
So, how does ChatExcel do it? What is the principle behind it?
Carry out “Party B’s thought” to the end
In a nutshell, the principle of ChatExcel is to directly convert “vernacular instructions” into a programming language similar to VBA, and then execute the program.
The bottom layer is based on the Transformer architecture, and the basic technical route is unsupervised training + fine-tuning for specific scenarios.
However, the NLP model has always been prone to errors in mathematics, and it is difficult to avoid it if it is as powerful as ChatGPT.
To this end, the team focused on symbolic logic in the process of training the model, and also deliberately introduced some new knowledge of logical symbols.
From this we can also see that the probability of its calculation error is not high.
In addition to outstanding mathematical ability, one of the biggest features of ChatExcel is continuous interaction.
This is because every time ChatExcel generates results, it is based on the new requirements put forward by users + the table generated in the previous round. In fact, higher requirements are put forward for the understanding and operation of the model.
Why implement this function?
According to the team, AI tools such as Dall E, ChatBCG, etc., complete tasks in a single and non-continuous way. But in the process of people’s actual use, ideas are advanced step by step.
For example, the relationship between users and ChatExcel is like that of Party A and Party B. ChatExcel is Party B, and we who use it to process tables are equivalent to Party A:
As long as you ask for it, change it until you are satisfied. (Doge)
Similarly, this thinking of Party B is also reflected in the development team itself.
After the demo was released, many people came to watch, try it out and give suggestions. The most frequent feedback is that ChatExcel is a little bit confused about the request and needs to be rephrased.
For these situations, the team first clarified its attitude: it is recommended that we accept them in full.
At the same time, it further explained the reason. The current open demo is not very perfect.
On the one hand, this is to see everyone’s feedback for further optimization; on the other hand, I also hope to collect more corpus to make the model smarter in the future.
For example, someone suggested:
Since there is already an “undo” (undo) operation, when can it display the process of processing the form.
Seconded, too much table data makes it difficult to check that it’s processed correctly:
Improve the transparency of ChatExcel’s execution tasks to facilitate error checking.
Regarding these suggestions, the team told us that they are already working on adjustments.
For example, many people reported language comprehension problems. They plan to recommend some standard operation sentences based on the user’s input in the dialog box. You can choose directly, and you don’t need to study the wording yourself. It’s a bit like AI customer service now.
In addition, these functions are also on the way to development:
Form upload format is more compatible
Show the specific process of ChatExcel processing tables, more transparent…
As well as a more specific introduction to the principles of ChatExcel, it will also be published in the form of a blog on the official website.
Even the code will be open sourced.
Inspired by women’s tickets, make amateur projects into official topics
Speaking of the birth of ChatExcel, it is also quite a result.
It was inspired by the female votes of the team members, the instructor immediately approved it, and turned the amateur project into a formal topic…
This starts with Kid and CY pulling together a small team.
They are all from the research group of Mr. Yuan Li, School of Information Engineering, Shenzhen Graduate School, Peking University. They are already PhD candidates, and one member is studying for a master’s degree.
As basic theoretical researchers of machine learning, they deal more with papers on a daily basis.
Why the sudden shift from theoretical research to application?
This has to start with Kid’s female ticket.
He said that his girlfriend is a middle school information teacher, and she often needs to help older teachers deal with Excel spreadsheets in their daily work.
In order to help his girlfriend share more work (bushi), and of course seeing that many people are troubled by the complex operation of Excel, Kid came up with the idea of giving it a try.
I think the use of Excel formulas still requires programming thinking, which is very unfriendly to many people, especially those who are not good at computer operations. It is indeed more troublesome to gnaw from scratch with a reference book.
At that time, OpenAI’s Copilot was on fire. Seeing that AI had great potential in improving productivity, he took CY together and planned to study this matter carefully.
But things didn’t go as smoothly as expected.
On the one hand, it started off as a side project that everyone worked on when they found time.
On the other hand, the team has also made major adjustments in the technical route.
At first, they thought that ChatExcel should go towards the discriminant model, so they followed this line of thought, but they didn’t expect:
The pre-processing workload is too large, and a lot of symbolic logic needs to be introduced, and the data requirements are very high.
This is basically fatal for small teams, because the amount of data means that the cost of manpower, material resources and time will soar.
There was no other way, Kid and CY sat down and reviewed the matter again:
We thought that Mr. Li Mu had posted a video about “the battle between BERT and GPT”. At first, we all felt that BERT performed better in various indicators. As for GPT, I am a little confused about what it wants to do.
But if you think about it carefully, from the perspective of the requirements of the task itself, generative is what we really want.
Coupled with the fact that OpenAI’s series of generative work started to catch fire last year (such as Copilot, Dall·E), the two finally decided to overthrow the original thinking and start from the generative model again.
Once the technical route is clarified, the subsequent advancement will be smoother.
In addition, this work has also been recognized and supported by the mentor Yuan Li, and the team has more resources and funds.
(ps But the team also explained to us that it can’t afford too much cost at present, the running speed will be affected, and there are still problems with some command line recognition)
Anyway, after 1 year, ChatExcel finally released the demo.
But this is only the first step of the plan.
The team said that in the future, more iterative upgrades will be made to the product, and the code will be open sourced to reveal the technology behind it.
For the positioning of ChatExcel, they insist on “small and beautiful”.
Kid said that large models are indeed the current trend, such as OpenAI, Baidu, etc. are constantly advancing results in the field of general large models. But they believe that in some sub-task areas, the ability of the model needs to be improved, which may be a trend in the future.
As a result, they will now focus more on improving the capabilities of the model in specific scenarios.
As for the long-term goal, the team hopes to create a “Chat ecology” and interactive AIGC.
The ultimate “Chat Ecology” (fantasy version) will probably look like this:
Back to ChatExcel itself, the team said that it does have commercial considerations, but it will be at the To B level.
(I heard that many companies have contacted them as soon as the demo was released)
On the To C side, we will adhere to the public welfare and open form, which can be used by everyone.
Looking back at the entire development process, many feelings of Kid and CY are still fresh in my memory.
For example, they were surprised and delighted when they saw ChatGPT make a form that became popular.
“Surprised” because I started to worry that there will be many similar products coming out? Your own project has no room to survive?
Kid said he couldn’t sleep at night after seeing all these new projects pouring out.
“Happy” is because I feel that my ideas are in line with the development of the trend, which is also a kind of affirmation.
As for the project itself, their attitude is very calm, accepting all problems, but they are also very confident in their results.
We still think our solution is “elegant”.
I feel that people who do theory turn to do application, and they can always think of some strange ideas hhh.
(At the request of the interviewee, Kid and CY are pseudonyms in the article)
Portal: https://chatexcel.com/
Reference link:
[1] https://www.zhihu.com/question/586673687
[2] https://zhuanlan.zhihu.com/p/607997627