Professor Ehud Reiter’s latest blog post, Pragmatic Correctness is a Challenge for NLG, clearly articulates a problem that we’ve noticed in some GPT3-based NLG systems. Specifically: sentences that might be factually accurate when presented by themselves will often sound strange–might even lead a reader to inaccurate conclusions–when strung together in a sequence.
In the case of automated property listing descriptions, such as provided by Nila June Instant Property Descriptions, it is critical that sentences be aware of their larger context. That is why, for example, Nila June does not give real estate agents three separate sentences to cover a short commute, great view, and cedar deck, but instead might produce something like this:
The short commute to Denver gives you more time to enjoy the amazing views of the Rocky Mountains as you grill on your cedar deck.
In addition to being semantically (factually) correct, the sentence is also pragmatically (contextually) informative, and written in a conversational tone.
[Side note: Not sure how long this offer will last, but Nila June is currently offering free automated property listing descriptions as a welcome to real estate agents who want to try the service. Use code welcome530 at checkout.]
Remember: NLG stands for NATURAL language generation. It’s worthless without the “N.” Merely generating vast quantities of written material can amount to a form of pollution. When poorly written text is generated in great quantities and then, by default, is used as the basis for machine learning NLP algorithms, the language regresses. As practitioners of natural language generation, our goal is to produce great writing instantly. We do no one any favors by producing bad writing for its own sake.
In an environment where many traditional writers point to bad NLG and say, “Told ya,” and many technologists say, “Fine then, we’ll keep doing it without you,” sticking to the high road might not be easy to do, but the map is certainly not mysterious, and the result can be an enduring business asset.