This is a guest post written by BowTiedBear.
I want to thank Hal for the invitation. My guest posts on the substack of BowTied_Raptor took a look at the reality of AI in the broader world LINK. In this post, I want to outline what I have experienced specifically for NLP applications.
To set expectations right: what I give you here is my perspective on use cases that proved to be beneficial in several industries and have been adopted into regular operations. Many more things are currently tested but not deployed at a large scale.
Talk Computer to Me
Let us first talk a bit about what people typically think about when they talk about AI applied to human language: conversational AI aka chat - or voice assistant.
To my memory triggered by the IBM Watson stunt on Jeopardy in 2012:
This made huge waves at the time and was followed by hype between 2015 and 2020 when many major corporations tried to implement some form of voice assistants with huge fanfare.
Let us agree that the results have been rather underwhelming for the most part:
Most companies simply did not have the necessary training data to train good models. What was used instead? Available solutions worked based on pre-defined dialogue trees. Probably some component to map different ways to express a question to keywords the voice assistants could understand, but usually nothing more sophisticated. What I have seen most of the time is voice assistants that do some tricks well but offer too few practical benefits for any consumer to actually use them.
This article here from 2019 makes a similar argument:
“Today, most chatbots rely on simple keyword recognition rather than true AI or machine learning. The result is that conversations may feel stilted and unhelpful to customers.”
And here for a bit of entertainment a few really bad bot cases:
“Also, human conversations are never linear and you can’t expect your customer-chatbot interactions to always run smoothly. Users may want to change the subject or ask a clarifying question. A chatbot is truly conversational only when it can properly process abbreviations, misspellings, jargon, colloquialisms, and subtleties.”
So, all lost here?
Here come pre-trained language models
The introduction of pre-trained language models in 2018 significantly improved the chances to build a good specific language understanding with limited training data. This was nothing short of a breakthrough.
Hal describes this in detail here LINK.
It takes some time to wrap such fundamental improvements into applications used by companies broadly, but I would expect a lot more good bots in the next years.
To nail this point home I would argue that consumers got fairly comfortable with talking to a computer with some of the best implementations out there: The likes of Google Assistant, Amazon Alexa, and Apple Siri. While they have been used more as toys initially, I see a lot more people using them for practical purposes.
I want to point out one more spectacular success of NLP technology: machine translation systems.
The well-tested natural language understanding applications
But now let us talk about the humbler applications that generate very tangible benefits.
First and foremost consequential but forgotten applications of language processing: spam filters. To be fair, mail content is only one part, but an important one.
This points us in an interesting direction:
Generally, language processing that has the goal to categorize documents or extract certain elements of a text in applications that can deal with accuracy < 100% is very widespread in all industries. To be very clear, those solutions do not aim at actual understanding of the text context, but rather performing a text classification of a document into the right category.
Real-world use cases
Let us stick to E-Mails and take it a step further and have a look at this real-world report:
Probably needs to be taken with a grain of salt as it comes from a vendor, but it seems rather realistic to me.
“For email alone, Aflac fielded thousands of requests every week. These requests pointed to seven separate inboxes that required manual classification and assignment of work across Aflac’s contact center. With the volume of email requests continuing to surge due to a rapidly shifting global environment, Aflac needed to improve efficiency and reduce costs while delivering exceptional service.”
Let us take a look at the capabilities described for this solution:
This is clearly less ambitious than a bot that talks to a human as good as most humans would, but a very good example of results that can be achieved with fairly basic natural language processing.
Under-explored natural language application
One personal guess for areas that should benefit soon: knowledge management. This is a complete mess in many organizations. You will find dozens of file systems or SharePoints with no way of searching across all files so most of the time you need somebody who knows where to look.
To be fair, this is often also a cultural problem – knowledge is seen as power and not shared freely. But you can´t convince me that this is the right way to work and we could do much better with current technology.
Outro
To wrap this up:
1. Natural language processing applications that aim at actual language understanding and interaction with a human have been introduced broadly but often with disappointing results – The newest technology however gives you very good odds of success.
2. Much more humble natural language processing applications that aim at text classification or extraction of certain elements are much more established, you can often rely on good off-the-shelf tools.
Thanks to BowTiedBear for writing this post!