(Report producer/analyst: Guosheng Securities, Liu Gaochang, Yang Ran)
First, AI capabilities have been accumulated for a long time, and will focus on 2023
AI has been built up for a long time and has been elevated to a strategic position since 2017.
1) In 2016, AlphaGo set off a wave of AI, followed by Kingsoft Office to accelerate its layout; In 2017, AI was elevated to a strategic position within the company, and the AI middle office was established; In 2018, Kingsoft Office established four strategies of "multi-screen, cloud, AI and content"; In 2020, smart office has taken shape, and nearly 100 AI office capabilities have been developed, covering full-text translation, document proofreading, intelligent writing, PPT beautification, data analysis, etc., landing more than 18 AI applications and winning 5 international AI technology awards.
2) We believe that with R&D strength, user accumulation and office know-how, AI has become an important accelerator for the growth of Kingsoft Office.
Vice President Yao Dong led the establishment, and the AI team has exceeded 100 people.
1) According to Info reports, in May 2017, Yao Dong, a veteran who had been away from Kingsoft for more than ten years, returned to his old club and served as vice president of Kingsoft Office. Yao Dong joined Kingsoft in 1998 and was previously responsible for Kingsoft Wordmaster, Kingsoft Ranger and other products.
This time back to Jinshan, Yao Dong transformed into algorithms and engineering products in the direction of AI, was responsible for algorithm improvement, promoting project landing, and talent echelon construction, and led the establishment of the AI algorithm, engineering and product team of Kingsoft Office.
2) As of November 2022, Kingsoft Office's AI team has about 100 people, divided into infrastructure, platform, products and applications, basic algorithms and other groups.
In the 5 years since its founding, the AI team has focused on different goals at each stage, with a "three-step" strategy. In the first two years, the team emphasized accumulating AI R&D capabilities, including algorithm capabilities, engineering capabilities, data collection, and data analysis capabilities.
In the last two years, more attention was paid to the productization of technology and AI product capabilities.
R&D efforts are high, and about 300 million yuan will be invested in office AI projects, which are planned to be completed in 4 years.
1) In November 2019, the company issued a prospectus, accounting for 16% of the construction of artificial intelligence basic R&D center in the office field in the fundraising project, with a total investment of about 319 million yuan.
2) The construction direction of the construction direction of the artificial intelligence basic research and development center in the office field is to use the basic technology of artificial intelligence in the office field to improve the understanding and processing ability of WPS Office software on natural language, image text information and other content, so as to improve the speed and efficiency of user processing files.
Main efforts: human-computer collaborative assisted writing system research and development project based on massive corpus, AI natural language processing platform project, AI computer vision recognition platform project.
A large number of users and diverse scenarios provide rich soil for AI to grow.
1) As of September 2022, the number of monthly active devices of the company's main products reached 578 million, continuing to grow steadily. On the one hand, this has accumulated rich usage scenarios for the company, and on the other hand, it has also provided a data experience basis for the iterative update of AI technology.
2) At the same time, relying on more than 30 years of R&D experience in the office field, the company has moved from PC to mobile and to cloud, and has also accumulated corresponding industry Knowhow in accommodation and catering industry, manufacturing, construction industry, education industry, retail industry, etc., with strong landing capabilities.
A number of capabilities have been in the first echelon in China, committed to reshaping smart office.
1) As of July 2021, the OCR and machine translation technology independently developed by the company has reached the level of the first echelon in China, and the number of intelligent proofreading words has exceeded 7 billion per month; Intelligently generated content accounted for 33.6% of the overall content resources, and the number of monthly active users of the intelligent beautification function exceeded one million; The AI middle office has also built an AI training platform based on distributed training, including a one-stop platform for training data processing, training scheduling, service monitoring and alarming, and link tracing.
2) According to Vice President Yao Dong, the company has three major AI development strategies: first, focus on intelligent text processing to improve office efficiency; second, knowledge services based on cognitive intelligence; Third, help enterprises with digital transformation.
In 2023, we will focus on the field of AI and focus on empowering the digital transformation of enterprises.
1) On March 14, according to Xinhuanet, scientific and technological innovation is the lifeblood of enterprises, and it is more related to the development of national strategy. In an interview with reporters, Zhang Qingyuan, CEO of Kingsoft Office, said, "Kingsoft Office is a beneficiary and practitioner of self-reliance and self-improvement in science and technology, and has adhered to independent innovation for 35 years since its establishment, solved the 'stuck neck' problem, and created our own national office software brand."
2) At the same time, the favorable policy of "vigorously developing the digital economy" will bring important opportunities for the development of Kingsoft Office. Zhang Qingyuan introduced that the core strategic goal of Kingsoft Office this year is to continue to empower the digital transformation of enterprises, and will focus on the field of AI, especially in AIGC (artificial intelligence generated content) to achieve more technological application breakthroughs, to help customers better achieve digital transformation.
Second, the coordinated development of the three major technologies of text, image and voice reshapes smart office
2.1 Text: Natural language processing accelerates the landing and document intelligence is outstanding
Intelligent document processing has a vast space, and NLP accelerates the landing of the office field.
The iterative upgrading of natural language processing technology has rich landing scenarios in the business field, and plays an important role in office scenarios such as document processing through the integration of knowledge graph and computer vision technology.
Through the in-depth mining of data, document intelligence has excelled in manual information review, translation, and intelligent writing. According to KBV Research, the global intelligent document processing market is expected to reach $4.1 billion by 2027, effectively linking the internal document data problem and accelerating the development of intelligent office.
WPS layout forward, NLP and other existing achievements include machine translation, document proofreading, assisted writing:
1. Machine translation: Multilingual translation capabilities are enhanced, and format processing is done in combination with CV technology.
After the application of AI technology, the translation effect is significantly improved, combined with the layout analysis advantages of CV, machine translation ensures consistent format, alignment, and style. With the help of WPS, rice husk network and other product ecology, the company's previous Kingsoft intelligent translation can realize the translation of text, documents and multiple fields, and can carry out post-editing.
At the same time, the Mongolian version of WPS Office 2021 released by Kingsoft Office is equipped with AI product add-on technology, of which the conversion rate of machine translation exceeds 90%.
2. Document proofreading: Combined with a rich corpus, Dark Horse proofreading V30 ensures high efficiency.
1) In the daily office, typo problems occur frequently, and the document proofreading function can find most of the typos in a very short time, and realize the functions of word segmentation, document classification, identification, extraction and proofreading, listing errors and giving modification suggestions.
2) According to 36Kr, in 2021, Kingsoft Office wholly acquired Dark Horse Feiteng and its Dark Horse proofreading related products, and in September 2022, Kingsoft Office's Dark Horse Proofreading V30 version was officially launched, V30 version analyzes the corpus of trillions of Chinese characters, has a database of about 80 million Chinese knowledge and a database of about 8 million error rules, which is applied to the official document proofreading of governments and enterprises and the quality control of the news and publishing industry, and is embedded in the enterprise business system.
3. Assisted writing: Kumon assistance has been explored, and it has been invested in intelligent writing research and development for 2 and a half years.
1) In June 2018, Kingsoft Office became the chairman unit of China Intelligent Writing Industry Alliance, which was selected as the "2018 Innovation Project for Deep Integration of Artificial Intelligence and Real Economy", and has now launched Kumon intelligent (assisted) writing and WPS intelligent writing functions.
2) Kumon intelligent (assisted) writing helps users to typeset in accordance with GB/T 9704-2012 normative requirements, built-in 19 Kumon templates and 15 legal Kumon types, promoting collaboration efficiency and avoiding wrong versions and running versions in different software and hardware environments.
3) Auxiliary writing products accumulate a number of outline libraries and corpus, the data are from authoritative media and government open networks, support automatic text generation, assisted drafting, sentence intelligent writing and text intelligent proofreading and other functions, in addition, machine independent learning can combine user behavior data and feedback to judge writing preferences.
2.2 Image: OCR has been accumulated for a long time, and complex scenes and beautification functions are gradually superimposed
CV is widely used in multiple scenarios, and the business value of OCR as an underlying general-purpose capability has emerged. Computer vision is the ability of computers to understand digital images and videos and extract target information from multimodal data through deep learning algorithms.
As the underlying general technology of CV, OCR technology is one of the AI technologies with the most application value, and has generated great commercial value in multiple vertical industries such as smart office, smart education, smart finance, smart transportation, smart city, and smart tourism.
At present, Kingsoft Office has explored the application of text recognition technology in multiple complex scenarios, and developed computer vision technology to realize multiple functions such as document correction, intelligent cutout, font recognition, table restoration, layout restoration, and PDF editing.
Computer vision has a rich heritage, and existing achievements include OCR, typography restoration, and image recognition.
1. OCR: extraction-screening-sorting-aggregation, supporting multi-environment deployment, and realizing structured understanding of documents. With the increasing frequency of OCR technology in mobile terminals, Kingsoft Office OCR technology has been iteratively upgraded, in addition to understanding multiple text information such as Chinese and English, simplified and traditional characters, but also can understand the semantic and structured information behind the text.
1) In the CSIG Image and Graphics Technology Challenge, Kingsoft extracted and filtered the OCR text box and text through algorithms in "Chinese and English Shopping Receipt Information Understanding", and understood Chinese and English information to organize and collect, and won the single track championship and the double championship of the finals.
2) In terms of deployment environment, Kingsoft OCR supports high-precision model deployment on the server side and supports small model deployment on mobile phones and PCs, and in 2019, the OCR model is less than 10M in the mobile inference model, and the accuracy is only 2% lower than that of the server.
2. Format restoration system: correction-identification-analysis-reconstruction, restoration of a variety of complex scenes.
Layout restoration refers to parsing complex picture documents into editable documents to help users re-edit non-editable documents in complex scenarios; Kingsoft integrates document scanning, document conversion, extraction and editing into a system to meet the reusability under complex requirements. The technology is currently leading in the industry, and the company integrates 30 deep learning models, 100+ algorithm modules, and 50+ 10,000 lines of code to achieve complex scene layout restoration such as bending correction, polluted light, with seals & signatures, and no table lines.
3. Image enhancement and intelligent typesetting: The company can use algorithms to achieve accurate restoration of text style (highly recognize text color, font, bold, italic, underline and other formats), and file format beautification and intelligent typesetting.
At present, Kingsoft Office has applied CV technology to many product functions such as intelligent cutout, ID photo production, document quality improvement, filters, watermark smearing and so on.
At the same time, Kingsoft Office invested in the online design platform of Maker Sticker (holding 12.79%), which can realize the function of intelligent cutout and graphics to generate video with the help of AI intelligence, significantly improving the work efficiency of designers and clerical personnel.
2.3 Voice: Voice conversion enriches office scenarios and creates a multi-level office experience
The speech conversion function breaks the boundaries of language and enables efficient and convenient office.
Kingsoft Office's text-to-speech conversion function in office scenarios also has rich landing scenarios, among which, voice interaction technology can realize 36 language conversions to achieve rapid text-to-speech conversion and reading; On the other hand, the company also implements language shorthand, which can be applied to both mobile and PC; In addition, the review function also adds voice annotation, which uses voice to realize data search and content input, which significantly improves the convenience of office review.
2.4 AI Middle Office: Launched the KSAI-Lite open source framework, which is universal, high-performance, lightweight and professional
AI middle office capability output, launched the KASI-Lite deep learning inference framework.
The company built an internal platform in 2017 and achieved external technology output in 2021 through long-term continuous technology investment. On July 22, 2021, Kingsoft Office released the KSAI-lite open source framework, which is free and open source, which not only adapts to mainstream software and hardware platforms at home and abroad and domestic information and innovation environments, but also optimizes performance, power consumption, and memory, and provides technical support for scenarios such as OCR, machine translation, and intelligent proofreading.
The KSAI-Lite framework integrates a variety of AI functions to suit multi-device offline computing scenarios. With the help of TensorFlow and TensorFlow Lite's algorithm optimization capabilities at the framework layer, the open source framework provides AI offline computing on stand-alone machines, mobile phones, PCs and other devices, adapting to scenarios such as private data processing on the client side, fast algorithm execution and real-time.
Object edge detection: The CNN document detection network designed by Kingsoft WPS allows Android users to quickly detect results, automatically determine edges and adjust filters;
Automatic recognition of image types: Kingsoft WPS uses TensorFlowLite to implement an OCR model that can automatically identify image types, and provides corresponding filters and OCR output formats.
Scanned OCR: Using TensorFlow to deploy the model, you can implement rotation correction and text line detection and other operations on documents, saving a lot of document editing time;
Natural scene OCR: Use TensorFlowLite to run natural scene OCR on mobile phones, so that it can accurately locate text from complex scenes in a short time and obtain ideal recognition results;
Layout analysis of image to document: Kingsoft WPS combines TensorFlow and scikit-learn framework to carry out graphic layout analysis algorithms, which greatly reduces the research and development cost of the algorithm.
Simultaneous release of KSAI OCR open source model for lightweight deployment.
At the 2021 Kingsoft Office Technology Open Day, Kingsoft Office also released the KSAI OCR open source model on the same day.
The OCR model and library file size does not exceed 9MB, can be lightweightly deployed, the model has shown good performance in text detection, text classification, and text recognition, and the essence of OCR is to convert photos into machine-encoded text.
At present, OCR technology has gradually become popular in the market and has become an important supplement to the way of document information entry.
The KSAI-Lite framework has been launched into the mainstream open source community and is expected to provide solid support for the company's mainstream products and the industry. According to the official WeChat public account of Kingsoft Office, Yao Dong, vice president of Kingsoft Office and head of AI middle office, said that as of 2021, the KSAI-lite framework has been launched on the mainstream open source community GitHub.
In the future, KSAI-lite will continue to make efforts in richer platform adaptation capabilities, more personalized development methods, and more stable business support capabilities, providing solid AI support for mainstream products and industries under Kingsoft Office.
Third, entering a new era of cognitive intelligence, AIGC attracts overseas giants to continue to enter
From perception to cognition, "computing power + algorithm + data", knowledge has become the fourth pole of AI development.
1) The perception system can identify digital information and the physical world, and the cognitive system goes further on this basis to realize the induction, reasoning, deduction, decision-making, feedback, and attribution of perception results. From the perspective of office scenarios, only "cognitive intelligence" can realize the mining of invoice format, semantics, and fragmentary information, and intelligently understand invoice types, reimbursement risks, and compliance issues.
2) The future of artificial intelligence from simple information reading to information understanding, in addition to the help of algorithms, computing power and data, it is necessary to integrate prior knowledge into the algorithm model, in addition to the pre-training of general education, it is also necessary for industry experts to make fine adjustments to achieve more professional training results.
Microsoft plans to integrate OpenAI tools into its entire line of products.
2023/01/23, Microsoft announced through its official blog to expand its partnership with OpenAI, which will make a multi-year, multibillion-dollar investment in OpenAI to help its technological breakthrough in the field of AI.
2023/01/17, Microsoft CEO Nadella said at the World Economic Forum in Davos that the next stage of Microsoft will focus on accelerating various tools to the market and commercializing OpenAI's tools, and the company plans to integrate artificial intelligence tools including ChatGPT, DALL-E into Microsoft's full line of products, including Bing search engine, Office Family Bucket, Azure cloud services, etc.
AI office or one of the first fields of generative artificial intelligence to land.
On March 7, Microsoft announced the extension of ChatGPT's technology to its PowerPlatform platform, which will allow users to develop their own applications with little to no code.
Separately, Microsoft will host an online event called "The Future of Work with AI" on March 16, where CEO Nadella may demonstrate how ChatGPT-like AI can work in Office productivity suites such as Teams, Word and Outlook.
We believe that with solid R&D strength, massive user and scenario accumulation, and deep office know-how, AI is expected to become an important accelerator for the growth of Kingsoft office performance.
Risk Warning
Cloud services are not advancing as expected. At present, cloud computing is in a period of rapid penetration in mainland China, and if the acceptance of cloud services by downstream customers is suspended, it may affect the company's business promotion.
State-owned IT spending was lower than expected. Information innovation is highly correlated with policies, budgets and other factors, and if the IT expenditure of the party and government and industry is not as expected, it may have a fluctuating impact on the company's business in the short term.
Personnel growth exceeded expectations. The important assets of software enterprises are employees, and their salary expenses, bonuses and benefits, subsidies, etc. account for an important part of the company's costs, and if the personnel growth rate increases rapidly, it may affect the company's profit release in the short term.
Macroeconomic risks. There are many macroeconomic influencing factors, which may have an impact on the company's business promotion.
——————————————————
The report belongs to the original author, we do not make any investment advice! If there is any infringement, please delete it by private message, thank you!
The report is selected from [Foresight Think Tank] Wenku - Foresight Think Tank