Leveraging technology that generates text is coming to the main theaters and Forbes is the most recent one: “The Biggest Opportunity In Generative AI Is Language, Not Images”
Different names are in use: generative AI, as in the article; synthetic text, following the popular term “synthetic data”; NLG (Natural Language Generation) is the most traditional term maybe not so trendy just for that reason.
Synthetic Text, as we will call it, started to follow the path of synthetic image recently. Synthetic image and video have been a huge success in sectors like self-driven cars.
For text, the initial successes have come from tabular data. In structured or tabular text, what’s generated is names (James O’Reilly, Bethesda Pharmaceuticals Inc.) or phrases (Junior Accountant, out of order) properly combined in tables or relational structures.
The next step in synthetic text seems to be unstructured data, where actual full sentences are produced, rather than phrases or names in tables.
Report generation, based on numeric tables, is an intermediate step between generating tabular data and actually generating full sentences from scratch. It’s very popular for sectors like e-commerce, finance or pharma.
At Bitext, we are focused on generating unstructured text for customer service applications and solving problems like:
- How do I generate hundreds/thousands of variations of a customer request (like “cancel my account”) so I can train a virtual assistant?
- Can I use text generation to produce comprehensive evaluation datasets?
- How do you express a given request (“can I cancel my account now?”) in colloquial register (“can u pls cancel account”) because my target is young adults?
You can take a look at a sample data in our GitHub Repository