Analyze a 20,000-word paper in 10 seconds, can this large model help me write a graduation thesis | AI measured

Original Source: Number One AI Player

Image source: Generated by Unbounded AI

On July 26, NetEase Youdao officially released the "Ziyue" large model, and launched six major application achievements based on the "Ziyue" large model - "LLM Translation", "AIBox", "Grammar Intensive", " Document Q&A", "Virtual Human Speaking Coach", "AI Composition Guidance".

As the first vertical model in the field of education, how is the performance of "Zi Yue"? Can it meet our expectations for educational vertical models?

After watching NetEase Youdao's "Ziyue" large-scale model conference, "Number One AI Player" evaluated the six application functions of the "Ziyue" large-scale model, hoping to give some references to players.

Translation Productivity Tool - "LLM Translation"

NetEase Youdao has undoubtedly been leading in the domestic translation field for many years.

According to official data, Youdao Translate has more than 100 million monthly active users and is the number one dictionary and translation product in China. As users have higher and higher requirements for translation quality, from "understanding is enough" to "accurate expression", what is the translation quality of Youdao translation equipped with LLM (Large Language Model)? Can it meet the needs of users?

We extracted fragments of reports on the new movie "Oppenheimer" from foreign news websites, and conducted comparative tests on "LLM translation" and the artificial intelligence robot Claude ai.

It can be seen that due to the completely different expression habits, cultural background and grammar between Chinese and English, Claude ai still has some shortcomings in the expression of Chinese sentences.

For example, for the verb collocation of "nuclear weapon", Claude ai uses "development", while "LLM translation" uses the Chinese customary collocation of "research and development".

However, in the follow-up translation and understanding of "Oppenheimer", neither of them accurately translated it as "Oppenheimer's release... coincides with the emergence of a large number of new technologies".

The expression of Claude ai does not conform to the Chinese expression habits, and the "LLM translation" ignores the word "released", which may be due to the lack of specific text background, resulting in inaccurate translation.

We also uploaded this English passage to Google Translate, and the answer it gave was obviously more convoluted:

After trying the English-to-Chinese translation, we decided to test their understanding of the translation of ancient poems.

"Please translate 'flying down three thousand feet, it is suspected that the Milky Way falls nine days' into English."

The result was a bit unexpected. When translating Chinese into English, Claude ai could simultaneously translate into English in the form of confrontation.

On the other hand, "LLM translation" takes the poet as the first point of view and supplements the missing subject object when translating into English, but the translation of "waterfall" into "swift current" (a stream with faster water flow) is not as accurate as "waterfall".

In general, Youdao LLM translation can provide more authentic translation according to the context in the field of English translation, and some expressions are better than Claude ai. Compared with traditional machine translation, the accuracy of understanding, fluency of expression and fluency of semantics have all been significantly improved. In the Chinese-to-English link, LLM translation is not as accurate as Claude ai's performance in the selection of individual words.

Netease Youdao CEO Zhou Feng said at the press conference that this "LLM translation is based on the neural network translation data accumulated by Youdao over the years, and uses a large number of new technologies, including single-language corpus that has not been used before. It can give everyone a better translation experience.”

Sentence polishing assistant - AI box

In addition to accessing the large language model, Youdao Translation also launched a sentence polishing assistant - AI box.

It is worth mentioning that the NetEase Youdao translation client has fully launched the AI box function. We took out the previous English composition and invited the AI box to polish and rewrite it to see how it works.

This article is mainly to write an invitation letter to invite university professors to come to China for investigation. We have chosen the function of "Sentence Polishing-Professionalization".

It can be seen that the AI box can directly rewrite the English sentences in the document in the desired way, and put forward suggestions for rewriting.

Depending on the style of the article written, the user can also choose "more gorgeous, more friendly, more concise" and so on, and the AI box will perform corresponding grammatical error correction at the same time:

The conspicuous "make the vocabulary more in line with SCI paper writing" below is a good gospel for the scientific research party

References are no longer difficult - Documentation Q&A

In academic research, reading literature is the most common and time-consuming link.

As we all know, ChatGPT has no way to read long papers, and the updated Claude 2 has been strengthened in text analysis and long document processing, but it takes a long time.

Duan Yitao, Chief Scientist of Netease Youdao, said: "Youdao's document question and answer function can increase the user's work efficiency by ten times."

We tried to upload a paper in the field of journalism with nearly 20,000 words. After about 10 seconds, the "Ziyue" model translated the full text of it and made concluding speeches on different pages.

On the document Q&A page, the original text and the translated text will be displayed separately. We can ask questions about the details of the paper, with annotations and screenshots, and youdao speed reading can quickly give answers.

If you don't want to read the whole paper, you can directly let the big model summarize the final conclusion of the paper:

It can also explain the new concepts in the paper and help us better understand the relevant content of the paper:

Compared with Claude ai, the more convenient function of "Document Q&A" is direct screenshot query.

For example, I intercepted the graphs in the paper and asked the big model to explain what it means. While giving the answer, it also labels the source of the information, making it easy for users to do fact-checking.

"Document Q&A" can help users quickly judge whether an article is worthy of intensive reading, but it can only interpret and analyze a single document, and does not have the function of searching similar materials online.

If you want to understand peer-relevant research, the large model is currently unable to answer.

24-hour online foreign teacher - speaking coach

In addition to the above-mentioned several application functions, the "Ziyue" large model also focuses on the area of greater demand in English learning - oral training.

At present, you can experience the function of AI oral English coaching on Netease Youdao's internal test webpage.

We can talk to the speaking coach about any topic of interest, and there are related learning objectives and challenges under each topic page to encourage users to actively express themselves in the conversation.

First of all, we chose "Favorite TV Show" to start today's topic. When we talked about the latest Hollywood movies and the actors' strike, the "Speaking Coach" gave the following answers.

If my pronunciation is not standard, can the speaking coach recognize it?

For example, when I was reading "Oppenheimer", the "speech-to-text" link recognized my sentence as "open", but the speaking coach could still "understand" me and commented that it was a good movie, It seems that Chinglish can't stop him.

When practicing speaking, the "speaking coach" can play different roles in the corresponding language environment, and guide users step by step to actively "speak English".

Compared with similar products, what are the advantages of "Speaking Coach"? Prior to this, the technology company Animato released a real-person dialogue artificial intelligence "Call Annie". This 30-year-old woman named Annie can undertake one-on-one foreign teacher work to help users train their speaking and listening skills. Annie also became a famous Internet AI teacher for a time.

We tried to find that Annie's speech rate is similar to that of VOA (Voice of America), and she is a "talker".

If you start a topic, it will be difficult for you to interrupt her. You can only listen to her share quietly, and you can't tell who is training English for a while.

In addition, she is not proficient in Chinese recognition. If the pronunciation is not accurate during the video, or there is a mixture of Chinese and English, Annie will not be able to recognize it successfully. For example, she didn't understand the title of the movie I shared.

In terms of interactivity, topic discussion, and word accuracy, "Spoken English Coach" adapts to users of different English levels, achieves heuristic learning companionship, and gives users polishing suggestions from the dimensions of pronunciation, grammar, and word usage. It is an English training partner that is more suitable for Chinese babies. These are features that the "Call Annie" app does not currently have.

Youdao R&D team said: "General-purpose conversational robots cannot meet the requirements of oral practice. The product positioning of Call Annie is not for foreign language learning, but more like a virtual assistant. We have optimized the heuristic of the oral language coaching model, and we expect this The model can better guide users to speak English."

Based on the function of the spoken language coach, before the 2023 World Artificial Intelligence Conference, Netease Youdao brought an upgraded version of the spoken language coach-virtual human language training Echo, combined with Youdao's self-developed "Ziyue" large model, and also There is a virtual human generation technology provided by NetEase Fuxi Big Model.

For the virtual speaking coach, Youdao’s R&D team admitted frankly at the press conference, “After all, oral sparring training is a market worth tens of billions. We look forward to this speaking coach gradually replacing real-person sparring.”

In the future, users can experience a 1V1 oral private teaching experience similar to real people. The virtual coach not only has rich expressions and authentic accents, but also gives personal advice on the pronunciation of practitioners.

AI Composition Guidance

For language learning, NetEase Youdao also launched the "AI Composition Guidance" function to solve the problems that students face in their daily writing, such as lack of materials and unclear themes.

Entering the Chinese writing interface, you can see that this function includes two aspects: "Teach me to write a composition" and "Help me to correct my composition". In the "Teach me to write a composition" section, you only need to enter the topic of the composition to get intelligent analysis, helping students have a clearer understanding of the theme of the composition so that they can write better.

We chose a composition question from the composition of the college entrance examination, and let the "Chinese Writing" function try to write it.

Compared with other language models such as ChatGPT that directly generate an article, "Chinese Writing" guides students how to write and guides them to explore the answers by themselves. It will guide students' composition framework in terms of core ideas, material references, and paragraph assignments.

"What is the hardest thing about writing a good article? Gao Huituan, Product Manager of NetEase Youdao Smart Hardware, pointed out at the press conference, "Finding the creative ideas is the most difficult part of writing. "

He emphasized, "The AI composition guidance function provides more creative methods and ideas in writing, rather than simple materials and model essays."

The function of "help me change" can deeply understand the content of the composition, and rewrite the composition from the four dimensions of expression, structure, emotion, and material richness.

We uploaded the articles generated by Claude ai to "AI Composition", and the articles written by AI were rated as "excellent" by AI. It may be that the article is really impeccable, and the focus of "AI Composition" correction is on good words, sentences and article writing structure.

In the end, this composition correction, which includes correction of typos, suggestions for improvement, good words and sentences, and general comments on the composition, will be synthesized into a complete report. Based on this report, users can improve weak links in a targeted manner.

The above several applications are all teaching and learning interactions completed by humans and AI technology in different teaching scenarios.

At the press conference, Zhou Feng, CEO of Netease Youdao, has also been emphasizing the importance of "scenes". He pointed out: "Whether a good technology has value and whether it can play a huge role depends on the choice of scenarios and applications in many cases." , and the polishing of details.”

When technology giants put their energy into the research and development of large-scale models, various enterprises are involved in the implementation of large-scale models in educational scenarios.

In May of this year, HKUST Xunfei released the world's first AI learning machine equipped with a large model, which provides functions such as composition correction and simulated oral teacher dialogue.

On May 8, Xueersi announced that it plans to develop a large mathematical model, mainly for mathematics enthusiasts around the world, with mathematical problem-solving and lecture algorithms as the core.

Backgammon also updated and launched the backgammon learning machine, launching AI-generated personalized knowledge graphs, AI intelligent recommendation of similar questions and other functions.

After ChatGPT became popular, Zhou Feng said in an interview that Youdao is a leading company in domestic natural language processing capabilities. In fact, after NetEase Youdao released a spoiler video of an AI oral teacher based on the "Ziyue" large model in May, this ChatGPT-like model "Ziyue" was officially launched in July, which seems a bit "long overdue".

"Compared with the chatting and dialogue capabilities of other AI tools, Netease Youdao's large model pays more attention to the implementation of educational applications," Zhou Feng emphasized at the press conference, "Ziyue's large model can not only help students learn better, but also Help teachers teach better, and one-to-one teaching becomes more of a reality."

Education, text rewriting, oral training, Netease Youdao's "Ziyue" model covers almost all these AI hotspots, and we can also find on the relevant evaluation pages that educational functions such as chess education and programming education are also available. in development.

Whether the "Ziyue" large model can become an "all-round teacher" in the education industry in the future remains to be verified by time.