GPT-3 vs Humans

I just read this insightful, albeit controversial for many, article on OpenAI's GPT-3 and Large Language Models and what they mean for language understanding and AI progress, and it resonated with me:

https://garymarcus.substack.com/p/noam-chomsky-and-gpt-3?s=r

The basic tenet is supported not only by Noam Chomsky (the "father of modern Linguistics"), but wholly embraced by myself, too, as a Computational Linguist, Language Engineer and Conversational Experience Designer: it is that Large Language Models only mimic human language use without any sense of meaning or of the relation of the words to the world. And by mimic we mean "monkey"-type of mimicking, not "baby"-type of gradual brain moulding and learning.

NLP vs NLU

As I've long been saying, there is a big difference between NLP (Natural Language Processing), which models, predicts, spots and generates word sequences learned over terabytes of uncurated language data scraped from the internet (hence potentially biased, fake or nonsensical) and NLU (Natural Language Understanding), which abstracts beyond the words and even word sequences and tries to assign some kind of meaning to the word sequences (user intent and goal, semantic category, ontology mapping).

There is a parallel in the Voice AI and Speech Recognition world: we juxtapose "word accuracy" (whether the ASR got all the words right) and "concept accuracy" (whether the ASR got the right "words", i.e. extracted the right user intent or slot value). A Speech IVR, Voicebot or Chatbot may only get right half the words that you spoke or typed (50% word accuracy), but may still get at what you actually want to say or do (100% concept accuracy). Very often the opposite can be true too; the ASR may recognise all the words typed in a chat, as they are already in its vocabulary (100% word accuracy), but may still not have a clue what you're on about (0% concept accuracy), because this concept, goal, intent, slot, slot value has not yet been included in its domain model.

Statistics + Linguistics = Explainable NLU

This is why language analysis and modelling, whether it's for content extraction, text summarisation or conversational experience design, should be a combination of robust data modelling and Computational Linguistics. Neither by itself can get it right:

• ? Conversation Design that is not based on a sufficient amount of real-world data (80k utterances is the standard in the Contact Centre world) is bound to give rise to Voicebots and Chatbots that will invariably fail in the face of actual language use and user behaviour, irrespective of how natural, friendly or even endearing the system messages are.
• ? A Conversation Designer who hasn't been trained in Linguistics is also bound to get the system messages wrong, by confusing, misleading or even putting the user off with messages that are too vague, ambiguous or chatty.
• ? Data Science that does not supplement string manipulation (NLP) with semantic ontologies and interaction patterns (NLU), will, too, result in natural language interfaces that have limited functionality, and are, hence, unhelpful, if not nonsensical.
• ? Engineers who design conversations and write system prompts always get the messaging wrong (verbose, full of jargon, too formal, confusing), which trips the user and the conversation.

This is also my approach to Conversational Experience Design that I have been using for the past 13+ years at my company, DialogCONNECTION, and now at GlobalLogic: "Explainable VUI & CUI Design", Voice and Conversational User Interface Design that is anchored in both statistical modelling of real-world data and Computational Linguistics and NLU (Morphology, Syntax, Semantics, Pragmatics). Along the same lines, this translates to "Explainable NLU".

Enter Google Imagen

More recently, the?GOOGLE?text-to-image diffusion model?IMAGEN?wowed everyone by generating incredibly realistic - usually beautiful - pictures. However, once again, it was trained on large, mostly uncurated, web-scraped datasets, meaning it unsurprisingly also replicates social biases and stereotypes, as the researchers themselves admit:

- "an overall bias towards generating images of people with lighter skin tones"

- "portraying different professions to align with Western gender stereotypes"

- "social and cultural biases when generating images of activities, events, and objects".?

Good job, the code is not in the public domain yet, as, in the wrong hands, Imagen could generate fake, defamatory or even harmful content. The future should be trodden with caution!

TL:DR

In short, don't get overexcited over GPT3 and such LMs or even text-to-image algorithms such as Imagen; they may have impressive (creative, beautiful, funny, even scary) generative power, but next to no explanatory power. That is, despite the name, current large Language Models only represent a tiny fraction of what human language use is about and what human language conveys. The right terminology is "word sequence modelling", which does not equal actual language understanding (or even ... human language modelling)! We should instead strive for explainability: Explainable NLP, NLU, VUI and Conversation Design.

Many thanks to Maria Spyropoulou for spotting the article!

#gpt3 #lms #datamodels #datascience #imagen #imageprocessing #language #nlp #nlu #linguistics #computationallinguistics #asr #speechrecognition #voiceai #voicebots #chatbots #speechivrs #conversationalai #conversationdesign #cx #cxstrategy #cxtransformation #cxdesign #xAI #AI #AIethics #openai #GOOGLE #responsibleai?#AIbias?#ML?#futuretrends

Dr Maria Aretoulaki is Principal Consultant CX Design (Voice & Conversational AI) at GlobalLogic UK&I and Director at DialogCONNECTION. She has been designing Voice User Interfaces (VUIs and Speech IVRs) for the past 25+ years, mainly for Contact Centre automation and Customer Self-Service. In 2019, she coined the term "Explainable VUI Design" to promote the principle of data-based Conversation Design and Computational Linguistics-based Language Engineering.