Dialogue with Liang Jianzhang: Vertical large models are more complex than AI poetry, and the tourism industry still needs human intervention

Image credit: Generated by Unbounded AI tools

Source: Sohu Technology

Author: Pan Yuyu

On July 17, Ctrip Group released the first large-scale vertical model of the tourism industry “Ctrip Asking”. Liang Jianzhang, chairman of the board of directors of Ctrip Group, introduced that under the background of intelligent society, the tourism industry is “a spiritual demand industry that is difficult to automate”.

He believes that AI is economically complementary to human beings. In the short term, many service jobs cannot be replaced by AI. In the long run, innovative and emotional jobs cannot replace humans. "In general, human beings will have more time to engage in entertainment, tourism, and innovative activities in the future. Tourism will become an industry with a higher proportion of GDP in the future.

Liang Jianzhang called the AI transformation triggered by the large model “the moment of evolution”. He believes that from a commercial point of view, how many disruptive applications and companies AI technology can produce is still open to debate, “but AI has already achieved human self-knowledge. Had a real impact, and this is the moment of human cognitive iteration.”

In the past two months, companies in many fields have released large vertical models. In the collaborative office scenario, Kingsoft Office released the AI office product “WPS AI” based on large models; in addition, NetEase integrated large models with education and games, and released a virtual human language coach.

Liang Jianzhang pointed out that AIGC’s assistance to OTA can cover pre-sales, after-sales, and en route. In the media dialogue after the meeting, Liang Jianzhang also added to the media including Sohu Technology, “The pre-sales itinerary planning is what customers spend the most time on, and it may also be the most complicated. Therefore, the AI large model can maximize efficiency. part.”

According to reports, as a large vertical model, Ctrip.com has screened 20 billion unstructured tourism data, combined with Ctrip’s existing structural real-time data and robots and search algorithms trained by Ctrip’s history, and conducted self-developed vertical model training.

Liang Jianzhang told Sohu Technology that Ctrip’s investment in AI is long-term. “Our data plus AI team does not know hundreds of people.” What is more special is that Ctrip has recently closely integrated these teams with large-scale models, and also used large-scale models to integrate real-time data currently owned by Ctrip, such as distance and route geometry. Data is integrated.

Liang Jianzhang pointed out that the opportunity for large vertical models is that the answers generated by AIGC are still not so reliable. Therefore, Ctrip has invested manpower to generate and verify the general travel reply content, so that the reliability of the recommendation is better than that of the basic large model.

He believes that algorithms cannot solve all problems, because tourism is a novel experience, so human intervention is indispensable.

Liang Jianzhang introduced that in order to solve the problem of general AIGC, Ctrip manually verified the common theme recommendations of hotels, scenic spots and itineraries on the basis of intelligent algorithms and formed the “Ctrip Word of Mouth List”, “Ctrip Hot List” and “Ctrip Hot List” of travel hotspot trends. Special Offers".

In the media dialogue, Liang Jianzhang also said that pre-departure itinerary planning is the most complicated link, so the AI large model can greatly improve the efficiency of this part. Liang Jianzhang pointed out that the biggest problem with tourism-related content is reliability, “it is indeed more complicated than AI writing poems and articles and novels.”

When asked about the capability of Ctrip’s vertical large model among the three elements of the large model, Liang Jianzhang believes that the evaluation indicators of the vertical large model are different from those of the general large model. “The language model is only one part of it. Others include data combination, manual verification, etc., which are more important.” When the model falls to the application layer, the efficiency and accuracy should be tested by whether it can satisfy customers’ questions and answers and reliability.

The following is an edited transcript of the conversation:

Media: What kind of problems and challenges does the tourism vertical model face in the industry? How to break through? In particular, the tourism industry chain is relatively long and the scenarios are relatively diverse, so what link is our large model mainly used in?

Liang Jianzhang: In terms of scenes, our roughest classification is itinerary planning, which is pre-sales. This is also the part where customers spend the most time, and it may be the most complicated, and it is also the part where the AI model can maximize efficiency.

We are the most qualified to do it, but it is also a long-term and continuous improvement process. It is indeed more complicated than AI writing poems, articles, and novels. The biggest problem in tourism is reliability. The long-term work is to improve the accuracy to 80%, 90%, or even 99%.

Media: Due to the changes in this era of AI intelligence, Ctrip is investing heavily in AI products and strategies. What energy have we invested in these? In the organizational structure, what new changes are there to support the development of this strategy?

Liang Jianzhang: In terms of AI, we did not start investing today. We are investing in machine learning or big data for a long time. Our data plus AI team does not know hundreds of people. Recently, what we have been special about is that these teams are closely integrated with the big model.

We not only use large models, but also use large models to combine our real-time data and algorithms. Today’s AI cannot combine geometric things such as distances and lines. It can also be said that it is an old team or a new team that integrates these.

This work will not be completed today, and it will take a long time to iterate. Ctrip has a certain foundation in this field. Another thing that is special about Ctrip is that our list is based on algorithms and human labor. Human labor does have certain advantages in the global scope. Our landing team is very large, and they need to help complete our content recommendation work.

Algorithms cannot solve all problems, because travel is a novel experience, and sometimes you cannot completely replace human intervention, so I think we need human intervention indispensable.

Media: Do you think the AI strategy and products we released this time will stimulate more demand from users?

Liang Jianzhang: If your recommendations are more accurate, you can find hotspots and price depressions to meet individual needs. It will not be so laborious. Very accurate recommendations will definitely better match demand and supply. Whoever will do this, we are the most suitable to do this. Now with the help of AI, the matching work can be done better.

Media: How much capital can be invested in the field of vertical large-scale models? What is the plan for subsequent investment?

Liang Jianzhang: I don’t know how much technical details can be disclosed, because it is true that their plans are constantly changing. But Ctrip will spare no effort to invest in this aspect.

Sun Jie: We should be very firm in investing in something that is valuable to our customers or merchants in the long run. Our technical team is constantly optimizing the amount of their investment, but we do not set limits, and it is helpful to our customers, and we will greatly increase investment.

Media: What is the training process of this model? I see the scale of our parameters, and the corpus has not been specifically announced. How does Ctrip choose the data in its own site?

Liang Jianzhang: Our vertical large model is different from the general large model. The most important indicators of the general large model are parameters, GPU and so on. But for the vertical large model, its language model is only one component. What it uses is a small part of the general large model, and other data combination, manual verification, etc., are more important. I think the parameters and data are not able to reflect the problem.

There are so many models now, and we accomplish different tasks based on different models. We have several internal schools, and we use different models to solve different problems. In the future, other large vertical models may be similar to ours, and they will not take the number of parameters they use as the most important indicator to see whether they can satisfy customers’ questions and answers, so as to test efficiency, accuracy and reliability.

Media: There is a hot topic recently. Everyone is worried that the current economy as a whole has a downward trend. In this context, the proportion of the tourism industry will increase in the long run. What do you think of the tourism industry in the short term? The development trend in recent years?

Liang Jianzhang: I think tourism will gradually become a pillar industry in the entire economy. Some countries, such as Thailand, have long regarded tourism as an important pillar industry of the economy. I think tourism will also be the same in the future, especially for provinces without high technology. It is very important for balanced regional development.

From the perspective of consumption, tourism is still in great demand from 10 years old to 70 years old. China’s total population is not growing, and this age group is still growing. It may not increase in this age group after 20 years. This is another topic. . In the short term, in the foreseeable ten or twenty years, this age group will continue to grow. I think tourism is an increasingly important consumption, especially as we know that material consumption will be saturated to a certain extent, but if the demand for tourism is truly free to travel, the per capita income is not 10,000 US dollars, but the per capita income may be 10 10,000 to 200,000 US dollars will go to Travel Freedom, so there is a long time for growth.

Media: In terms of market value, we have returned to the approximate level of 2019, but in terms of the growth potential of the industry scale, in fact, we have been underestimated for a long time. Are we not fighting for market value? How to reflect the market value of Ctrip?

Liang Jianzhang: Ctrip is a platform for comprehensive tourism. Its current market value and performance are far better than those in 2019. During the past few years of the epidemic, we have not been idle and have done a lot of technical work. The work during the current epidemic is reflected . In terms of market value, as far as the entire industry is concerned, as long as we do our job well, we shouldn’t worry too much.

Sun Jie: When I joined Ctrip more than ten years ago, I asked Liang Jianzhang the same question. At that time, our market value was 500 to 1 billion. I asked him what he thought? He said that as long as we operate well, there is no need to worry about market value. This is also the eternal focus that we encourage every BU. Don’t worry about the market value. As long as we can serve our customers well, keep innovating, and create valuable products, everything else will come naturally.