Google recently explored a technique called instruction fine-tuning to train models to solve natural language processing problems in a general way. This method is not to train a model to solve a problem, but to teach it how to solve a wide range of problems, thereby increasing efficiency and advancing the most advanced technology.
Google does not use all research in its algorithms
Google’s official statement on research papers is that just because it publishes an algorithm does not mean it is used for Google searches.
Nothing in the research paper says it should be used for search. But what makes this research interesting is that it advances the most advanced technology and improves the current technology.
Understand the value of technology
People who don’t know how search engines work might end up understanding it with pure guesswork.
This is how the search industry ends up with misconceptions such as “LSI keywords” and absurd strategies such as trying to beat competitors by creating content that is ten times better (or just larger) than competitors’ content, with zero consideration The user’s content may need and request.
advertise
Keep reading below
The value of understanding these algorithms and technologies lies in understanding the general outline of what is happening in search engines, so that people will not make mistakes by underestimating the capabilities of search engines.
Problems solved by FLAN
The main problem solved by this technology is to enable the machine to use its vast knowledge to solve real-world tasks.
This method teaches the machine how to solve a specific problem by inputting instructions, and then generalizes these instructions to solve other problems, thereby generalizing the problem solving to invisible problems.
The researchers stated:
“The model has been fine-tuned on different instruction sets and extended to invisible instructions. As more types of tasks are added to the fine-tuned data model, performance will improve.
…We show that by training the model according to these instructions, it will not only be good at solving various instructions seen during training, but also at following general instructions. “
advertise
Keep reading below
The research paper cites a currently popular technique called “zero-sample or few-sample hints”, which can train machines to solve specific language problems and describes the shortcomings of the technique.
Reference zero stroke/less stroke prompt technology:
“This technology formulates tasks based on the text that the language model may see during training, and then the language model generates answers by completing the text.
For example, in order to classify the emotions of movie reviews, the language model may be provided with a sentence such as “Movie reviews “The best RomCom since beautiful women” is _” and asked to use the words “positive” or “positive” Finish the sentence. “Negative”. “
The researchers pointed out that the zero-sample method performs well, but the performance must be measured in terms of tasks that the model has seen before.
The researchers wrote:
“…It requires careful prompt engineering to design tasks to make it look like the data the model sees during training…”
And this shortcoming is solved by FLAN. Because the training instructions are generalized, the model can solve more problems, including solving tasks that have not been trained before.
Can Google use this technology?
Google rarely discusses specific research papers and whether the content described is being used. Google’s official position on research papers is that it publishes many research papers, and they don’t necessarily appear in their search ranking algorithms.
Google is usually not transparent about the content in its algorithms, which is correct.
Even when announcing new technologies, Google tends to give them names that are inconsistent with published research papers. For example, names like Neural Matching and Rank Brain do not correspond to specific research papers.
It is important to review the success of the research, because some research did not achieve their goals and did not perform as well as the current state of the art in terms of technology and algorithms.
Those unqualified research papers are more or less ignored, but it is good to understand them.
The most valuable research papers for the search marketing community are those that are successful and perform significantly better than the current state-of-the-art technology.
advertise
Keep reading below
This is the case with FLAN.
FLAN performs better than other technologies, so FLAN needs attention.
The researchers pointed out:
“We evaluated FLAN on 25 missions and found that it was better than zero shooting tips on all but four missions. We found that our results were better than zero on 20 of the 25 missions. The sample GPT-3 is even better than a few samples of GPT-3 in some tasks.”
Natural language inference
Natural language inference tasks are tasks where the machine must determine whether a given premise is true, false, or uncertain/neutral (not true or false).
FLAN’s natural language inference performance
Reading comprehension
This is a task that answers questions based on the content of the document.
advertise
Keep reading below
FLAN reading comprehension score
Closed-book quality assurance
This is the ability to answer questions with fact data. It tests the ability to match known facts with questions. For example, answer questions such as what color the sky is or who is the first president of the United States.
FLAN closed volume QA performance
Does Google use FLAN?
As mentioned earlier, Google usually does not confirm whether they use a particular algorithm or technology.
advertise
Keep reading below
However, the fact that this particular technology promotes the development of the most advanced technology may mean that it is not unreasonable to speculate that some form of it can be integrated into Google’s algorithms to improve its ability to answer search queries.
The research was published on October 28, 2021.
Have some of them been included in the recent core algorithm update?
Core algorithm updates usually focus on better understanding of queries and web pages and providing better answers.
People can only speculate, because Google rarely shares details, especially in terms of core algorithm updates.
Citation
Introducing FLAN: a more general language model with fine-tuning of instructions
Picture from Shutterstock
