Intelligent search AI innovates the health field, reduces the hallucination rate of large models, and provides more professional and accurate health information
The phoneticization of GPT and the emergence of Gemini have allowed the public to invest the key indicators of large models into multimodal imagination earlier. As everyone knows, at the other end of the technology scene, a number of pragmatic large models are also rising.
In the world of verticals, manufacturers have begun to "roll" the illusion rate. Hallucinations are the biggest fatal flaw of large models that have not yet been widely used, and low hallucination rates represent higher scientific and realism. For large vertical models with more demanding professional information, generating the right content is far more important than generating content naturally.
Since the complexity of the dataset is one of the biggest factors leading to the hallucination rate, large model manufacturers need to use and filter the dataset carefully and rigorously in the development process, and alleviate the illusion problem from the quality of the dataset and better model structure.
Traditional health search: high-frequency user demand, inefficient information matching
A few days ago, in the professional evaluation, the hallucination rate of the quark large model on healthy content was reduced to less than 5%, and this figure is currently only reached by GPT among the popular general large models overseas. It can be seen that the large model is expected to become an "innovator" in the field of big health.
After the release of the self-developed large model, Quark first upgraded the service experience of health search, and through the previously accumulated health knowledge graph and the new health model application, it seems to be announcing to the market: the large model is the best fuel for health search.
In fact, the domestic medical and health field has entered the "Internet + era" as early as ten years ago, but there has not yet been a standardized platform, with the increase in the density of health information, the lack of verification of rumors and marketing content is also increasing, for patients and users, screening and judging the cause, it seems that it is "more difficult than before".
In contrast, in recent years, the importance of Chinese people to health has ushered in a significant explosion period. Health anxiety has led to a large number of users' online behaviors and search scenarios. The demand for high-frequency medical consultation and health consultation has also caused a shortage of medical resources in various regions. Online platforms such as search engines and social media are undoubtedly the fastest way for ordinary people to screen.
However, looking at the global market, traditional search engines have not been able to match the needs of healthy searches well, and the search results have been questioned. Based on the health information and diagnosis content of third-party platforms and article pages, etc., are uneven, and once the third-party information misleads users, it is an irreparable loss for the entire health search field.
As a result, in recent years, the "credibility" of social media for healthy search seems to have surpassed that of search engines. However, because social users are more sharing their personal experiences, they do not have the ability to systematically diagnose and popularize science. Therefore, whether it is the public or the entire health industry, what is more needed is the targeted change of search engines.
Health Search in the AI Era: Intelligent Screening, Complete Diagnosis Path
On the basis of the excellent performance of the vertical large model score, Quark has recently carried out a comprehensive upgrade for health search and launched the "Quark Health Assistant", which fully integrates search and AIGC content, which not only helps users to query health information, but also takes the first step in the application of large models to vertical search.
Recently, Lu Yifeng, a senior engineer at Google DeepMind, said in an interview that one of the most promising directions for large models is Retrieval Enhanced Generation (RAG). As early as 2015, Fang Binxing, an academician of the Chinese Academy of Engineering and an expert in network information security, proposed in a paper that traditional search engines cannot meet the changing search needs in ubiquitous cyberspace. At the same time, "how to turn the data value mining service of various industries into a general, publishable, and serving the general public search technology, people are no longer satisfied with the search engine just giving existential results, but need a set of intelligent solutions." ”
For example, the mismatch between health queries and user needs in the traditional search experience is mainly due to two main problems.
One is the limitation of what you can search for. When the product form of the search box is designed, it is to search for the existing information in the database, so it is more suitable for the input and accurate matching of keywords, which will make users instinctively short and keyword when describing the problem, and once they want to make a targeted detailed statement, they will feel at a loss. However, for health queries, short descriptions and keywords are clearly not enough. Even in offline consultations, doctors need to help patients make the right judgment through complete and logical statements.
At this point, the intelligent matching capability brought by AI is only one aspect, and the product form and path also need to be better reconstructed and polished to improve the guidance of health inquiries from the entrance page. Therefore, the upgraded Quark Health Search first changes the presentation form of the content on the result page, and the AIGC content provided by the "Quark Health Assistant" has appeared in front of users. The advantage is that it makes it easier for users to accept information by turning the sea of information into accurate screening.
At the same time, there are also quark health encyclopedia, intelligent screening and other functions to further help users understand the symptoms, medication, employment, maintenance and other suggestions. In particular, the intelligent screening function allows users to further describe the health problems they face by checking the symptom information according to their own situation, and then the system will generate relevant suggestions to assist users in making decisions.
The second crux of traditional health search is that although the current search engine is one of the largest databases for "collecting problems", the chain of problem solving is far from being standardized. As a result, even if the screening results are accurate enough to help users identify symptoms, they cannot tell users what to do for specific symptoms. Especially in the era of medical separation, different stages of the same disease, in the flow of basic care, buying medicine, and going to the hospital for further examination, will produce completely different scenarios and results.
In addition, from the perspective of the data itself, only by accompanying users through deeper query and solution scenarios can search engines better optimize the capabilities of large models and keep the information provided by AI up to date.
Therefore, in addition to AIGC content and intelligent screening, Quark has also launched the function page of "Quark Health Assistant", so that AI can learn to "talk" and "ask" like people, which is also to fully empower users and reduce the threshold for information query.
The targeted analysis of whether you need medical treatment and treatment suggestions given by Quark Health Assistant can better improve the accuracy of medical treatment. This is because, although the health search is based on the vertical scenario of the current symptoms, the real diagnosis and treatment requires a comprehensive observation of the patient. Generative AI can extract and proactively ask for key information in conversations, providing deeper, practical guidance.
This also echoes the industry's expectations for the application capabilities of large models. Previously, Shen Yang, a professor and doctoral supervisor at the School of Journalism of Tsinghua University, said that an important measure for large models to reduce the error rate is to collaborate with search engines. In overseas markets, the in-depth collaboration between Microsoft Bing and GPT, and Google's efforts and breakthroughs in Bard and Gemini have also verified this view. Even in the eyes of tech giants, the application within search engines is the foundation and focus of the imagination of large models.
However, in the vast technological innovation, overseas head search engines have spent more energy to temper the general solution ability of products, so that the search in the AI era is closer to an all-round, multi-modal AGI. In the domestic market, which has always been good at putting future technologies into vertical applications, manufacturers are more concerned about how to synergize emerging technologies and user experience in the way of productized exploration due to the higher traffic and frequency of users' standardized needs.
As a result, health search may become one of the earliest fields and scenarios in China to benefit from the application capabilities of large models, and at the same time, it can also be used as a pioneer in the era of intelligent search, adding more content from the first party and more controllable to the search engine, so as to break the user's prejudice against traditional health search, and shorten the distance between ordinary people and correct, professional and authoritative knowledge in the interactive experience with a lower threshold for expression and a freer interaction.
The future of large-scale products: everything is about user experience
In the AI transformation stage in full swing, how to integrate the capabilities of large models into their own product forms, Quark Health Search gives a powerful example of innovation.
Quark is well aware that innovation and inheritance of product interactions are indispensable: large-model-based search cannot just exist in the form of generative conversations, and professional diagnostic recommendations must be combined with existing user query paths.
Although Google has not yet launched an intelligent search page for health inquiries for C-end users like Quark, in October this year, Google Cloud also released a search product that helps doctors accurately extract clinical information from medical records with the help of Vertex AI.
As such, attempts to provide AI solutions for the health industry are in full swing. Globally, there are also more and more cases of AI-assisted diagnosis. Since the advent of Internet healthcare, the proportion of patients who conduct online dialogue consultations through chatbots and online doctors has been increasing, and these past processes have laid a cognitive foundation for today's AI health search market. Compared with the previous technical period, although large models cannot completely replace doctors' diagnosis, their diagnosis rate of diseases and underlying diseases has been quite good. Especially after quark and other large-scale model self-developed manufacturers have taken health ability as a technical target: quark large-scale model has passed the clinical practitioner qualification examination with a high score of 485.
Perhaps in the short term, AI will not be able to "become" doctors, but at present, AI can solve a lot of auxiliary and predictive work for doctors. According to Quark's technology and resource layout, it is foreseeable that there will be two trends in the field of health search next year:
First, the industry map will be more vertical and segmented. For example, according to the type of disease, population, and age stage, different vertical query scenarios are generated, and the types of health needs covered by structured intelligent Q&A will also be more flexible. At the same time, in terms of medical and health solutions developed by health search, the search engine will be used as a digital platform to better connect users with hospitals, servers, etc., because the user's online query can also quickly provide doctors with diagnostic directions, and the test application of Vertex AI in North American healthcare and some health organizations has provided feasibility for this.
As a result, the second trend has been triggered: out of consideration for the practicability and correctness of health diagnosis, cutting-edge judgments and professional data from the industry will be rapidly integrated into the ecosystem. In addition to the advantages of technology and user data, the quark model can show high professional ability in the field of health, as well as the cooperation of more than 200 authoritative medical experts, more than 60 national public tertiary hospitals, and more than 40 medical institutions. At this stage, the relationship between humans and AI is not a confrontation, but a synergistic relationship. Limited human resources are the source of "unlimited" wisdom, and an excellent large model cannot be separated from professional knowledge accumulation and correction. In this process, an open and active industry ecosystem is particularly important.
Whether it's Google, Bing or the domestic search engine Quark. From college entrance examination volunteering to health search, Quark has always been one of the few platforms that is good at "scene breakthroughs" around the industry map. In the era of large models, how to balance the sensibility of technological change and the rationality of practical scenarios has also reopened the concept of user experience.
Nowadays, the "stubborn disease" of traditional health search is being eliminated by technology. What is more worth looking forward to is that more and more medical practitioners can embrace and help more people obtain and maintain health conveniently and cost-effectively in the form of AI.
The views in this article only represent the author's own, and the 36Kr platform only provides information storage space services.
The image in this article is from: Enterprise Licensing