*** UPDATE 8-SEP-22 I have published a blog on another boring use of a language model, Using language models to improve rule/template NLG

I recently had a chat with one of Arria’s development groups (ie, people build who build real-world NLG solutions) about some possible uses of language models, which ranged from mundane to very ambitious. The team was excited about some of the mundane uses (which were varied, but I cannot give details publicly) and politely expressed concerns about the ambitious uses as being too unreliable and too difficult to control. I have had similar feedback in the past.

Language models of course can be used in many different ways in an NLG context. The media focuses on “gee whiz” stuff like generating narratives just by prompting a large language model (eg, “A robot wrote this entire article. Are you scared yet, human?” in the Guardian), and most academics seem to focus on the very ambitious goal of training end-to-end models to generate texts (eg, Sports NLG: Commercial vs Academic Perspective).

However, when I talk to people who are actually building real-world NLG systems (at Arria and elsewhere), they are far more interested in using language models in very limited and controlled ways to do specific focused tasks. An example I can give from Arria (because of a patent application) is that Arria is looking at using language models as part of the quality assurance process, to help testers find questionable narrative outputs. I dont think I’ve ever seen a media article on advanced NLP technology for software testing, and most academics show little interest in the topic (withsome notable exceptions, such as Ribeiro et al 2020). But quality assurance is a huge challenge in real-world NLG, and better tools can absolutely make a difference!

I’m reminded a bit of Named Entity Recognition (NER) in the wider NLP world. No one claims that NER is a sign of “artificial general intelligence” and I’ve never seen a media article extolling the wonders of advanced NER technology. But its a very useful capability in all sorts of real-world applications, and its great to see that papers on NER do appear in major academic NLP venues. So are there things like NER in NLG, ie tools and capabilities which are in some sense “boring” but are very useful in real-world applications?

Unfortunately I cant say much more about specific opportunities here, because of Arria commercial confidentiality. But I did want to make the general point to people who are working on using language models (and other advanced techniques) in NLG that they may want to consider a range of tasks and use cases, some of which are pretty mundane and indeed “boring”.