Natural language processing advanced by AI

In recent years, the social implementation of AI has progressed rapidly, and AI has become a familiar part of our lives. For example, there is content related to language, such as automatic answers with chatbots in user support and highly accurate Q&A.

When dealing with words, NLP is generally used. NLP is the process of analyzing words with a computer, capturing the characteristics of words by using the appearance frequency and appearance probability of words, and performing task processing related to various words.

Examples of social implementation include classification, extraction, Q&A, and search. For example, in classification, it is conceivable to classify products into categories from product descriptions, classify product reviews from users into positive and negative reviews, and feed them back to operations.

He also talks about the current state and future of NLP with AI:

Currently, most of the AIs that are being implemented in society are AIs that specialize in specific tasks, so-called specialized AIs. This is also called weak AI because it specializes in only one task. On the other hand, strong AI refers to AI that can solve general problems. I'm aiming for a more human-like behavior, but as far as I know, it doesn't exist in the world yet. However, some scholars predict that it will be born in the near future.

GPT-3 is an AI that is a gateway to the future of NLP, although it is not yet versatile enough to be called a strong AI. Opportunities to appear in the media are also increasing, such as being featured in the Nihon Keizai Shimbun.

Blog writing experiment with GPT-3

To test the capabilities of GPT-3, Macnica conducted an experiment in which they wrote a blog using GPT-3, and posted it on "Arima-ku," a blog that is open to the public for recruiting purposes. This blog is a dialogue between a character called "Macnica" and "G," a self-proclaimed futurist who is supposed to be a Macnica employee. This article about "G" was created using GPT-3.

The blog post initially disclosed that it used GPT-3

We asked "G", "I would like you to tell me five important developments in society over the next ten years." As a result, GPT-3 said, ``AI will become a part of our lives,'' ``Bionic limbs will become affordable, improving the quality of life for people with disabilities,'' and ``Solar power generation will become more affordable.'' will become the norm in many countries," "Gadgets will become smarter and more intuitive," and "We will recognize the importance of the environment and start creating the necessary infrastructure for a sustainable future." I did.

To dig deeper into the fifth answer, we asked, ``We will begin to create the infrastructure needed for a sustainable future.'' Three points were returned: ``promotion of waste'' and ``reduction of waste.''

I thought this was exactly the answer given by GPT-3. We believe that it is realistically possible for humans and GPT-3 to co-create articles, not just blogs.

Please note that the questions to GPT-3 were asked in English, and the answers were machine translated and slightly revised before being published in Japanese. GPT-3 can handle Japanese, but perhaps because the amount of Japanese it has learned is small, as far as Macnica has tested it, English seems to be the more natural answer at the moment.

Examples of GPT-3 usage

So how exactly should we use GPT-3? Let's take the example of writing an article for "Arima".

The first thing to do is describe the task. For example, given the input "You are a brilliant futurist. This is an interview with you. What are your thoughts on my question?" We then continue with the question, "What do you think will be the most significant developments in the next 10 years, and why?" The second, deeper question has a similar structure.

At that time, the description content is very important and directly affects the accuracy of the output. The point is how accurately you write to lead to the expected output.

Next, I will explain the source code. Parameters allow you to set whether you want a more creative answer, or a more rigid, accurate answer at the expense of some creativity.

Since this is a blog post, I'm setting myself up to expect more creative responses. The output from GPT-3 will be different every time, the more creative the setting. This time, I tried many times and picked up what I thought was good and published it.

The prompt input is actually achieved by Python code

GPT-3 applicable to various businesses

In this way, GPT-3 can derive answers similar to natural language, but in what fields can it be used other than blogging?

Macnica repeatedly contacted GPT-3 with the assumption that it would be used for chat. This time, I am inputting Japanese directly. Regardless of the truth or falsity of the content, the point is that the conversation is properly structured. If you ask similar questions in English, you may be able to get more creative answers.

GPT-3 also supports Japanese, but currently you can expect better answers in English

Next, we will automatically create a product advertisement for a fictitious ballpoint pen and notebook. Input is the product name and product description in addition to the task description.

The trade name of the ballpoint pen is Smoothwrite. One o was missing, but the o was complemented in the answer. The product description first made it clear what this is. Since the image is not given as input, it is necessary to make it recognize that it is a ballpoint pen by words.

The output advertisement text includes sentences that are not directly used in the product description, such as "written characters look beautiful", "rich color variations", and "cannot be reversed".

In the source code, I specified "Please write a creative advertisement to put the following products on the EC site" as the description of the task.

Creativity settings are conservative compared to chats and blogs. We believe this will prevent the hype.

Change your creativity settings to adjust for richness or precision

GPT-3 safety and risks

GPT-3 is very powerful, so there are some points to consider. There are four points: "safety", "harmfulness", "fairness" and "robustness".

Powerful AI is convenient, but on the other hand, it is risky, so be careful not to mishandle it.

In the case of GPT-3, OpenAI's approval is required for release, and strict measures such as API suspension will be imposed on those released without going through the approval process.

It also provides use case guidelines for developing applications that consider safety and risk.

Low risk ones are always approved. Movie review classification, information extraction, computation, SQL query generation, recipe generation, etc. tend to fall into the low-risk use cases. Medium risk ones are case by case. Creating blogs, creating product advertisements, etc. However, static posting to blog posts, not applications, is free. The high-risk ones are very sensitive use cases that are rarely allowed. Automatic posting to SNS will be the target.

GPT-3 is also a very powerful and versatile natural language processing model. Many challenges can be solved with a single model. It will revolutionize the NLP solution development process by significantly reducing the time to release. On the other hand, it is necessary to develop applications with full consideration of safety and risks.

▼ Business problem-solving AI service that utilizes the data science resources of 25,000 people worldwide
Click here for details.Fully customized AI service CrowdANALYTIX