On average, one high-scoring college entrance examination essay is generated in one second. How does the Fei Paddle Wenxin large model do this?

The national college entrance examination is underway, and Baidu AI technology is also ushering in a "college entrance examination." On June 7, Baidu digital person Du Xiaoxiao transformed into a "digital person candidate" after the college entrance examination Chinese language test and challenged the college entrance examination composition writing, triggering heated discussions across the Internet. AI writing involves many artificial intelligence technologies such as natural language processing, data mining, and knowledge graphs. Baidu used the Feipiao Wenxin large model to support Du Xiaoxiao in writing the college entrance examination essay well and quickly. According to reports, Du Xiaoxiao participated in answering the argumentative essay titled "Excellent Hands, Expert Hands, and Common Hands" in Paper I of the National New College Entrance Examination. He created more than 40 articles based on the topic in 40 seconds, with an average of one article generated per second, and randomly selected them. One article, its score has caught up with about 75% of college entrance examination candidates.

An average of 1 article per second, and the score exceeds 75%. Candidates write quickly and well in Wenxin Large Model

Applying AI to text creation, in technology This is nothing new in the industry. With the development of technologies such as big data and deep learning, AI has greatly improved its text generation capabilities. From financial and sports news flashes to literary creations such as poetry and novels, AI writing covers more and more fields, and the creative forms and contents have become more colorful.

But compared with content forms such as reviews, summaries, and bulletins, it is obviously more difficult to use AI to generate college entrance examination essays. The writing of college entrance examination essays poses greater challenges to AI in three aspects: "question review", "logic" and "creativity". At the question review level, there are many college entrance examination essay topics and forms, coupled with complicated content themes, each ever-changing combination is the ultimate test of AI's understanding ability. At the logical level, college entrance examination essays require more than 800 words. How to ensure the logic and coherence between the various fragments of a long essay is another big challenge for AI. Finally, at the creative level, readability is one of the most important scoring criteria for college entrance examination essays. AI writing must not only be relevant and coherent, but also be good at citing classics, skillfully using rhetoric, and even introducing new ones to avoid empty and boring articles.

Faced with these challenges, Du Xiaoxiao completed the writing quickly and well with the help and guidance of the flying paddle Wenxin large model. Shen Yi, who once served as the leader of the Chinese language marking team for the Beijing College Entrance Examination, scored 48 points (out of 60 points) for Du Xiaoxiao's essay. She said that based on previous years, less than 25% of candidates scored 48 points or above in composition. In other words, Du Xiaoxiao's composition score has surpassed about 75% of college entrance examination candidates. Pan Zhoudan, an internet celebrity who has participated in "The Strongest Brain", also expressed his "appreciation" for Du Xiaoxiao's composition level.

This is due to Wenxin Model’s newly released billion-dollar model ERNIE 3.0 Zeus, which integrates task-related knowledge. This model further learns more than a hundred different forms on the basis of learning massive data and knowledge. The task knowledge enhances the effect of the model and shows stronger zero-sample and small-sample learning capabilities on various NLP tasks. In layman's terms, the Wenxin model is like a well-informed "top student" or "a top student". Its understanding and creative abilities are stronger than those of traditional models. Not only does it have the rapid production capacity of "relying on the horse and waiting for you", but it also has the quality assurance that "it's like a god when you write".

With the support of Wenxin Big Model, the quality of Du Xiaoxiao’s college entrance examination essays has been significantly improved compared to the past. First, the generated composition can closely follow the theme and have a clear intention. Du Xiaoxiao understands the given composition title well, organizes the text around the theme, supports the article's views, and outputs positive insights on the topic; then in 800 In the long text of Chinese characters, Du Xiaoxiao can ensure that the generated content has a complete structure and smooth language, and introduces questions at the beginning and ends with smooth context connection. At the same time, because the Wenxin model has a stronger knowledge reserve, the compositions generated by Du Xiaoxiao are better at quoting scriptures and making appropriate use of rhetorical techniques such as parallelism and metaphor, which greatly improves the readability of the articles.

However, what surprised Shen Yi was that Du Xiaoxiao made a small mistake in the composition - using the Internet hot word YYDS. The reason why such "point loss" occurred is because Du Xiaoxiao studied some popular network data. Judging from the context of the generated article, Du Xiaoxiao has understood the "essence" of YYDS, and her own usage is appropriate, but it does not conform to the college entrance examination essay writing standards. It also reflects that AI writing can continue to learn and improve.

Wenxin Big Model helps intelligent creation and promotes the arrival of the AIGC era

In the past year, the implementation of Baidu AI technology has accelerated, especially the recently emerging AIGC. AIGC is a new content production method after UGC and PGC. It refers to the automatic production of content using artificial intelligence technology. Digital Human, TTV (Text to Video), etc. are its benchmark applications. Baidu Research Institute predicts that in 2022, AIGC technology will leverage the cross-modal comprehensive technical capabilities of large models to stimulate creativity, improve content diversity, reduce production costs, and achieve large-scale applications.

This time Du Xiaoxiao writes a composition for the college entrance examination, which is based on the AIGC creation based on the large flying paddle model. Wenxin large model has the core characteristics of "knowledge enhancement" and can conduct fusion learning from large-scale knowledge and massive data. It has higher learning efficiency and better effect, and has the characteristics of good versatility and strong generalization.

In addition to intelligent writing, with the support of the Wenxin large model, AI has strong understanding and generation capabilities, and can realize the automatic generation of creative works, including AI painting, AI songwriting, AI editing, etc. . Currently, creative applications of large models such as intelligent painting, intelligent dialogue, intelligent creation, and story generation have been launched on the Wenxin Large Model Creative and Exploration Community - Yanggu Community. Search "Wenxin Large Model" to enter the official website to experience it. In the future, AIGC based on large models will be open to more content production fields.

The rise of pre-trained large models has further enhanced the versatility of artificial intelligence. Large models have the characteristics of good effects, strong generalization, and high degree of standardization of R&D processes, and are becoming the new base for artificial intelligence technology and applications. Baidu has been deeply involved in the research and development of pre-training models since 2019, and has successively released a series of knowledge-enhanced Wenxin models. At the just-concluded WAVE SUMMIT 2022 Deep Learning Developer Summit, Wenxin Large Model ushered in a series of upgrades: 10 new large models were released, including the knowledge-enhanced 100 billion large model that integrates learning task knowledge, and the multi-task unified learning model. Visual large models, cross-modal large models, biological computing large models, industry large models, etc.; propose three key paths to support the implementation of the large model industry: build a large model system that is more suitable for scene needs, and provide full-process support for application implementation. Tools, platforms and methods, building an open ecosystem that inspires innovation, etc.

In addition to its application in intelligent creation, the Wenxin large model has also been used in industries such as industry, energy, education, finance, communications, media, etc., such as zero in the industrial field. Component quality inspection, transmission line inspection in the energy field, contract information extraction in the financial industry, etc., can truly help companies reduce costs, increase efficiency, and stimulate innovation. At the same time, the Wenxin large model is also fully used in Internet products such as intelligent search, information flow, and smart speakers to improve the efficiency and effectiveness of users' acquisition of information, knowledge, and services.

At present, Feipiao has built a model library system with the most comprehensive layout in the industry and the most suitable for industrial applications. As a part of the artificial intelligence "infrastructure", large models further broaden the coverage of scenarios for the implementation of artificial intelligence technology. It deepens the depth of industrial applications. Wenxin Big Model will continue to lower the application threshold, promote the intelligent upgrading of the industry, and make artificial intelligence technology benefit everyone.

The above is the detailed content of On average, one high-scoring college entrance examination essay is generated in one second. How does the Fei Paddle Wenxin large model do this?. For more information, please follow other related articles on the PHP Chinese website!