Workers meet big models: Is the outside world already working like this?

Image source: Generated by Unbounded AI

In office scenarios, making PPTs is one of the most common jobs. PPT is required for work reports, product releases, event planning, professional lectures, etc.

The traditional PPT production process is boring and trivial, consuming time and energy. Especially to convert the word format report document into PPT, it takes a lot of time to read the document content, sort out the main points, but also carefully typeset, adjust the font picture, etc., and change the word document PPT to re-sort and adjust.

Is there a possibility that doing PPT can be automated?

With the blessing of the big model, Baidu Wenku did it. After accessing Wenxin, Baidu Wenku has fully rebuilt from the “document search tool” to a “one-stop intelligent document platform”, covering multiple functions such as PPT intelligent generation, intelligent document generation, intelligent editing, and intelligent auxiliary reading.

For example, for the following article, upload a word document, and AI can quickly generate a ready-to-use PPT based on the content of the document.

As early as March this year, Baidu Wenku announced the access to “Wen Xin Yiyan” and opened the user internal test. On August 31, Baidu Wenxin was officially opened to the public, as the industry’s first “one-stop intelligent document platform”, Baidu Wenku’s “PPT intelligent generation” and other document creation functions ushered in more than 2 million experiences on the first day of opening.

In order to obtain a “one-stop” intelligent creation experience, we have conducted a comprehensive test of Baidu Wenku’s newly launched AI functions.

PPT Smart Generation

In recent years, the production of PPT has become more and more volumed, and various PPT artifacts are sought after by everyone. But these production tools need to be used according to their requirements, and sometimes even complicate the production of PPT.

In order to simplify and automate the production of PPT, Baidu Wenku Document Assistant has launched two PPT intelligent generation functions: directly generate PPT in one sentence and upload word to generate PPT.

Enter the theme to generate PPT directly

With the advancement of generative AI, there are some AI-generated PPT tools in the industry, most of which have the main function of entering topics to generate PPTs, such as Gamma.

In order to measure the generation effect of Baidu library, we tested and compared gamma and Baidu library with the phrase “generate a PPT about the history of dinosaurs”.

Gamma first generated a content outline, but the final generated PPT was only 7 pages, without a table of contents display page:

And the PPT content is simple, and the typesetting format is relatively monotonous:

* Gamma input theme generates PPT effect example. *

In the same statement, Baidu Wenku first generates a content outline, but it is more detailed and contains two levels of content points:

The final PPT generated by Baidu Library has a total of 24 pages, from the characteristics and evolution of dinosaurs to the overview of dinosaur research, rich and complete content, diverse and beautiful layout, almost direct use, the generation effect and practical value far exceed PPT tools such as gamma.

*Baidu Wenku “Enter the theme to directly generate PPT” effect example. *

Upload word to generate PPT

Compared with entering the theme to generate PPT, it is more difficult to convert a word document into PPT.

On the one hand, parsing a sentence is that the model expands according to the topic; To analyze an article, it is necessary to start from the primary and secondary structure of the article itself, analyze the key points, writing logic, and content style of the whole article, and fully understand and deconstruct the article in order to generate a logical, complete and rich PPT of the article.

On the other hand, the quality of Word articles tends to be uneven. Some articles are very simple and need to be enriched by AI models in the process of generating PPT; Some articles are complex and professional, requiring AI models to refine and summarize, and may also use the knowledge reserves of large models. This places higher demands on the capabilities of large models.

In addition, some pictures are usually required in PPT, and entering the theme to generate PPT only needs to find a suitable picture according to the theme; According to the word to generate PPT, the accompanying pictures should conform to the theme of the article, the general idea of the paragraph, and also adapt to the style of the article, etc., which is more restrictive.

In order to test Baidu Wenku’s ability to convert word documents into PPT, we uploaded an article titled “Frontier Development of Smart Home Technology”:

*Screenshot of the Word document section of “Frontier Development of Smart Home Technology”. *

Baidu Wenku’s document assistant still generates a detailed outline based on the content of the word document:

Click “Generate PPT”, select the template and generate a 36-page PPT:

Overall, this PPT is rich in content, beautifully typeset, and has a sense of technology consistent with the article. It may take at least tens of minutes to manually make such a PPT, but the AI-blessed Baidu library only took about 30 seconds.

Specifically, PPT expands a lot on the basis of word documents. Taking “smart lighting” as an example, the content in a Word document is only a few short lines of text:

*All about “Smart Lighting” in the Word document. *

In the PPT generated by Baidu Wenku, the “smart lighting” section first introduces the intelligent lighting control system and intelligent lighting appliances, then explains the scale and development trend of the intelligent lighting market, then points out the advantages and shortcomings of the intelligent lighting system, and finally looks forward to the development trend and challenges of intelligent lighting technology.

We found that the PPT generated by Baidu Library contains a lot of information other than word documents. This requires Baidu Wenku to use the “Wen Xin Yiyan” large model to deeply analyze the content of Word documents and generate knowledge-based content. In addition, the accompanying pictures in the PPT are also in line with the theme of the word article - smart home, which also requires the use of large models to understand the strong ability to understand.

Generate PPT with charts

Data chart is a common form of content in PPT, which can quantitatively display relevant results, intuitive and clear. In order to test whether Baidu Library can generate PPT with data charts, we enter the requirement in the dialog box of the document assistant: “Generate a financial analysis PPT of A smart home company”.

The document assistant is still Mr. into a PPT outline, it is worth noting that the outline not only has the content of financial data and analysis, but also the basic introduction of the company and the trend outlook of the industry. This shows that the document assistant understands what are the common uses of financial report analysis PPT, and knows what professional data is required for financial report analysis, such as profits, assets, cash flow, total revenue, and so on.

In the generated PPT, Document Assistant generates different types of data charts for different financial data, including bar charts, line charts, data tables, etc., and each chart has text interpretation.

For example, the Total Asset Details page contains a histogram of total assets, a data table of year-over-year growth rates of total assets, and an analysis of changes in total assets. Among them, the year-on-year growth rate is calculated based on total asset data. When manually making financial report analysis PPT, data such as year-on-year growth rate need to be calculated separately and then added to the PPT, and Baidu Wenku’s document assistant directly generates all the data with the help of AI large models.

If we need to modify the PPT generated by the Document Assistant, we can also directly let the Document Assistant help with the operation, such as modifying the accent color of the PPT:

In this way, it only takes a few minutes to make a PPT from the demand to the draft, and the office efficiency is not improved by a little.

Move the mouth to generate PPT

Finally, we found that all of the above functions can be used on the Baidu Wenku app, and there is an additional function: the mouth can generate PPT, that is, on the mobile Baidu Wenku app, we can directly voice input the demand, and the document assistant can complete the task of generating PPT.

Intelligent Document Generation

With word documents, AI can generate PPT, and word documents can also be directly generated by AI.

We know that one of the tasks that large language models are best at is text generation, and AI-assisted text creation is also one of the most common application directions of current large models, especially in office scenarios.

As a one-stop intelligent document platform, Baidu Wenku has launched a number of text-oriented functions such as “Generate Outline” and “Brainstorm”, and you can directly use these AI functions when creating new documents in Baidu Wenku.

AI helps you write

To test our ability to create text from scratch, we used Baidu Wenku to create an “editorial recruitment copy”.

As shown in the figure below, the intelligently generated results meet the requirements of the recruitment copywriting format, including company profile, job description, job requirements, benefits, application methods, company address, and introduce the job description and job requirements according to the specific position of “editor”. Just adjust some of the information according to the specific situation and you can actually use it.

Then we tested Baidu Wenku’s English writing ability, using Chinese input requirements: “Write an English essay titled “Autumn””. The articles generated by Baidu Library are written from the autumn scene to the autumn people’s behavior activities, with fluent writing and rich vocabulary.

Write an outline

Unlike office texts such as recruitment copywriting and emails, writing knowledge-based introduction articles often requires preliminary preparation such as collecting information and writing an outline. As a platform that includes a large number of knowledge-based documents, Baidu Wenku can quickly list the outline framework of articles according to the topic provided by the user in the function of writing an outline in AI.

For example, we tested Baidu Wenku’s “write an outline” function with the theme of “tea”, and the generated outline framework included the history, classification, production, tasting, culture, and future of tea, and each part was subdivided into several subsections.

Brainstorming

In addition to sketching and writing, in real working life, the most crucial step in text creation is to find ideas. Based on the generation ability of Wenxin’s big model and the rich document reserves of Baidu Wenku, the function of “brainstorming” can quickly find multiple angles for users.

For example, taking “Shampoo Product Marketing Strategy” as an example, the “Brainstorming” function quickly gives multiple ideas such as “brand story”, “target market analysis”, “product features”, “price strategy” and so on.

Of course, these functions can also be used directly in the Document Assistant, for example, directly enter the requirement in the dialog box: “Help me write a product promotion plan”, the result of the Document Assistant is shown in the following figure:

The whole copy includes seven parts: target market analysis, promotion goals and objectives, promotion strategy formulation, promotion content, promotion execution plan, promotion budget and resource requirements, promotion effect evaluation and summary, covering all aspects of event planning.

It is worth noting that the text generated by Baidu Library is of high quality, complete and detailed, which stems from Baidu Library’s ultra-large-scale high-quality document resources. Over the years, the total number of content included in Baidu Library has exceeded 1.2 billion, which gives Baidu Wenku Document Assistant a unique advantage in intelligent text editing.

In actual work, copywriting such as event planning and work reports is an extremely common daily work. It can take days to do these tasks manually, but Baidu Wenku’s document assistant can complete these tasks quickly and well. It seems that as long as we describe the writing needs in as much detail as possible, we can use AI to assist in many work tasks, and our work efficiency will increase by orders of magnitude.

Smart Editing

Compared to creating text from scratch, large models are not very good at editing text. This is because the generation of large models is relatively random, the generated text varies in length, and text editing requires accurate, meticulous adjustments to the text, and often has a word limit.

Currently, it is difficult to balance the completeness of expression with strict word limit for large models. The knowledge learned during the training process of the model affects the number of words it outputs, and the diversity requirements of the output content of the large model itself may lead to unstable output. Therefore, applying large models to text editing can be challenging.

We found that Baidu Wenku has launched a number of intelligent editing functions, overcoming some technical difficulties. When editing a document in Baidu Library, the “AI Intelligent Editing” button will automatically pop up when selecting the paragraph in the Chinese file, and click on it to bring up a function menu for AI editing text, which can be polished, revised, summarized, abbreviated, expanded, and changed in tone of the text.

We tried AI polishing a piece of text, and the result is shown in the figure below, a piece of text is enriched into two paragraphs, and the text description is more delicate:

In order to test the AI text revision function, we slightly modified the original text to make it contain language errors and sentences that did not flow smoothly, and then selected the “Vocabulary & Grammar Revision” function, and the result of Baidu Wenku AI revision is shown in the figure below:

We also selected an autopilot-related article in Baidu Library to test the abbreviation and amplification functions, aiming to evaluate how effective intelligent editing is for highly professional articles.

As shown in the figure below, after selecting the abbreviation function, Baidu Library abbreviated the two paragraphs into one paragraph, and clearly explained the important concepts and causal relationships in the original text.

In terms of expansion, we found that the expanded content added professional expressions such as “self-driving cars obtain information about the surrounding environment through lidar, cameras, ultrasonic sensors and other devices”, which is extended by Baidu Wenku according to the development status of autonomous driving, which shows that Baidu Wenku has mastered some knowledge and can intelligently edit highly professional content.

Baidu Library can complete a variety of text editing tasks with the help of AI, which shows that it has mastered the grammar, semantics, and language style of text. In actual text writing work, such auxiliary editing tools will save us a lot of time and effort.

In addition, we found that the Documentation Assistant can generate data charts based on text content: select the paragraph containing the data, and the Documentation Assistant on the right will automatically pop up the option to “Generate Chart”. This function can not only generate data charts, but also analyze according to the content of the article and the data situation.

For example, we tested this feature with a paragraph from a “Case Study of Corporate Financial Statement Analysis” that touched on sales margins. As shown in the following figure, Document Assistant generates a histogram of sales margins with cause analysis, solutions, insights, recommendations, and more.

Smart Assisted Reading

In the office scenario, the long summary ability of large models also has many practical uses, such as consulting reference materials, refining meeting minutes, speed reading contract terms, and so on.

As a platform with more than 100 million monthly active users, on Baidu Wenku, we used to search for information using search keywords, and after finding information, we needed to roughly check the bibliography and content to find the information we needed.

Now, Baidu Wenku can generate short summaries of the documents it includes with the help of the Wen Xin Yiyan model, allowing users to quickly understand the content of the documents, achieve intelligent reading assistance, and save office time. This makes it more convenient and fast for hundreds of millions of users to check information in Baidu Wenku, becoming the “natives” of AI learning office.

For example, we asked the document assistant to summarize a long article titled “The Development and Application of Artificial Intelligence” in Baidu Library:

You can also answer related questions based on the content of the documentation. For example, according to the article, the answer is: “When was artificial intelligence proposed?” The Documentation Assistant can give the correct answer and will indicate what the referenced article is based on.

Features such as summarizing document content, answering relevant questions, and more also apply to PDF documents. For example, when reading a 10,000-word long article about the basics and applications of multi-agent reinforcement learning, directly select “Help me summarize the general meaning of the document” in the document assistant on the right, and the AI will quickly give a summary of the content of the document, so that we can roughly understand the content of a 10,000-word long article in just a few seconds.

New office mode in seconds

This year, generative AI has revolutionized production tools. From the initial generation effect of the large model is amazing, to the beginning to explore the application direction, and now there are some more mature applications, the large model has moved from technology to landing. The Baidu library blessed by Wen Xin is a good example.

In the past, word documents, PPT, and search tools performed their respective roles in office scenarios, and office often needed to be cross-platform. Although the content is very relevant, writing a word document and making a PowerPoint are two separate tasks, and each work takes more time to complete, for example, writing a document requires searching for materials, building an outline, writing an article, editing and other steps.

Now, with just one sentence, Baidu Library can generate complete and detailed document content, upload documents to directly generate PPT, and the whole process may only take a few minutes. From this point of view, Baidu Wenku has solved the long-standing pain point of “office cross-platform”, and “one-stop” office has become a reality.

So, what specific benefits can Baidu Library bring to daily work?

In terms of work efficiency, it takes about 30 seconds for Baidu Library to generate PPT, about 15 seconds to create a document, and 10 seconds to summarize a 10,000-word long article. For any one of these tasks, the manual completion time is at least tens of minutes, sometimes even days. We can use the saved time to complete more innovative work, and relatively stylized work such as making PPT is done by AI, so that office efficiency is improved by orders of magnitude.

From the perspective of generation quality, the content generated by Baidu Library is of very high quality, clear logic and rich content, and often only needs people to adjust the generated content according to the actual situation and can be used directly. This is also an important reason why Baidu Wenku can be practically applied as a “one-stop intelligent document creation platform”. In just one month of full launch, the cumulative number of users of Baidu Wenku’s AI new features has exceeded 10 million, with a cumulative amount of generated content of more than 20 million and a total of more than 2 million PPTs.

From the perspective of usage scenarios, in the past, our office usually relied on the PC side, and the time and space conditions were limited. Now, using the Baidu Wenku app, you can complete tasks such as writing documents and making PPT with only very simple operations on the mobile terminal, breaking the time and space limitations of the office scene.

In fact, since the advent of large models, there has been an effort within the industry to improve the performance of large models so that they can be put into practical use. Baidu Wenku can become the industry’s first one-stop intelligent document creation platform due to three key factors.

First of all, based on very fine data, the Wenxin Yiyan model has trained strong comprehension, generation and logic capabilities, which will enable the Baidu library it supports to accurately understand user needs, generate clear and rich content logic, rich and reasonable.

Secondly, Baidu Library itself has 1.2 billion high-quality documents, which is the content advantage of Baidu Library reconstruction. These documents are also one of the important training data of the Wenxin Yiyan model, and Baidu Wenku and Wenxin Yiyan complement each other.

Third, Baidu Wenku’s own R&D team has been working hard to explore algorithm development and application landing for many years. This provides technical support for Baidu Wenku’s reconstruction into a “one-stop intelligent document creation platform”.

In just a few months after the words of Wen Xin came out, Baidu Wenku has overcome a number of technical difficulties, and will continue to iteratively upgrade in the future.

“The gold standard we have set for ourselves is the most usable and convenient,” said Wang Ying, vice president and head of interactive entertainment and pendant platforms at Baidu.

We look forward to seeing more intelligent creation features launched by Baidu Wenku, and we look forward to seeing the big model bring further improvements to productivity.