Most of you have seen “Batman: The Animated Series.” Have you noticed a scene towards the ending of the movie where the concept of scalable text generation is highlighted? Well, that part cannot be missed out. Generating copy for e-commerce category pages is perhaps one of the coolest strategies one can accommodate only if one knows how to do it.
What’s the hack?
Before getting into the hack and related other tactics, it is very important to know what DataToText is. A study of natural language generation. It allows people to make use of structured data to produce copy is known as DataToText. In many academic researches, you might have seen engineers focusing on use cases. These may be generating a copy for e-commerce product pages or providing a recap on some sports activities.
Sometimes, you will find single page applications incorporated for the creation of e-commerce websites. This is an indication of the presence of an API on the client’s side from which he or she can get access to capture the content while constructing or updating a page.
If you go by design, these API endpoints are available and open in most cases with little to zero authentication. However, these endpoints can be used to accumulate features as well as obtain data points that will yield innovative and fitting content.
Once the data is gathered, it can be arranged into some significant sentences to form paragraphs. For completing these paragraphs, you can take the help of a natural language generation library. The uniqueness of your content depends on the length of the paragraphs and where a sentence is placed in a paragraph.
What is GPT-2?
It is an open-source artificial intelligence that performs the function of translating text, summarizes passages, answers questions, and generates output that is barely distinguishable from humans. Over the years, it has revolutionized and developed a more advanced form of text generation that creates a copy.
GPT-2, the second generation of GPT has the capability to write copy that is almost indistinguishable from a human. The first announcement about the launch of GPT-2 was on an account outraging fake news and was said to be detrimental for the mass.
Finally, it was released with lesser factors than they have tested it with. Instead of using 1.5 billion variables, they implied only 774 million. So, if you find a substantial degree of copy on your site, you can test GPT-2 on that.
You can also make use of content spinning and then feed that data into GPT-2. Then, you have to generate as many words as you want to and continue with it until you reach the end of the content.
However, besides GPT-2, there are other text generation models too, like:
- The T5 of Google was tried and tested on a clear version of CommonCrawl.
- CTRL by salesforce with 1.6 billion parameters.
- Microsoft’s TuringNLG caters to 17 billion testing parameters.
How to generate copy for a category?
Now, the time has come to disclose to you the steps that you can follow to generate copy into your category pages on an e-commerce site. You can follow these steps on your own. Else, if you find it difficult, collaborating using Brainvire will ease your job.
Let’s get into the steps straightaway.
Step 1: Review the data model of your site
Pay attention to the data to develop the correct data blocks.
Step 2: Reproduce a good number of sentences that assimilate those data points
Get a series of phrases and break them down to make them fit into the data points.
Step 3: Gather as much relevant text as you can to fit it into your data space
A site with a good copy is enough for pulling text. Else, you can draw text from your competitors. Do not forget to save whatever you are pulling.
Step 4: Find and optimize a GPT-2 model
To build a model with which you can move ahead, you need to put the text file into the GPT-2 model.
Step 5: Use data from the data model to make the sentences fit in rightly
Use content spinning to produce your own sentences. Here, you can use any content spinning tool.
Step 6: Use data-driven statements as cues
Then, feed each of the sentences to the GPT-2 model to get the desired amount of copy that you would prefer to have . You can check out the difference in the length of placement and generated copy.
Step 7: Review and edit
After you have got your copy, make the necessary changes as required.
Step 8: Install
Finally, launch it on your site when you see the copy matches with the tone, brand voice and expectations of the user.
Therefore, try this newly discovered technique of generating data-driven copy for e-commerce category pages with GPT – 2 and see its amazing results.