The PDF in question attempts to address the carbon footprint of NLP models and provides some estimates. However, it is the first time I have been able to get some idea around the cost structures related to NLP (or any other ML algorithm). No where in the mainstream media, they address the cost to “run” the system.
I had attempted to create a chatbot in Telegram by hooking it up to a NLP library. However, it was pricey, though it is relatively straightforward to hook up the API end point and then get on with it. Integration costs with the chatbot are relatively less; NLP training (which is imperfect) causes more grief in the long-term. The bot could be trained only in English, which excludes a vast majority, and there was no way to control for every scenario. While it may appear futuristic to have the bot auto-answer the queries, the current NLP models appear reading from the script – they don’t understand the intent. GPT-3 might circumvent these problems, but if you observe at the costs for GPT-2- they are most expensive. The ROI will not be worth it.