2
$\begingroup$

In this question I asked about the role of knowledge graphs in the future, and in this answer I found that If curation and annotation are not sufficient, the knowledge base maybe cannot apply in AI.

ChatGPT does not utilize a knowledge graph to understand or generate common sense, then I wonder how knowledge graphs can be utilized in the future. Will they be replaced by LLMs?

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ It has a reason ChatGPT is not named as KBGPT/Q&AGPT/ExpertGPT, etc, OTOH though lots of curated plausible background knowledge belief network is there it doesn't mean any rational agent can utilize them to the fullest extent either deductively or inductively due to lack of insight of relevance. $\endgroup$
    – cinch
    Commented Feb 11, 2023 at 21:37
  • $\begingroup$ @mohottnad ChatGPT is not named KBGPT/Q&AGPT/ExpertGPT but has some basic common sense, you know Siri does not and Watson was very limited. $\endgroup$
    – Lerner Zhang
    Commented Feb 12, 2023 at 3:47

2 Answers 2

Reset to default
2
$\begingroup$

A couple of days ago, Jordi Torras from Inbenta posted that chatGPT fails at classifying a particular integer as prime, while their chatbot nails it. But the goal of a chatbot is no way factoring integers, is it?

Some weeks ago, Stephen Wolfram suggested some combination of chatGPT and their WolframAlpha, a curated engine for computational intelligence.

A wealth of domains could benefit from integrating preexisting knowledge into the conversational skill of transformers.

As a simple example, take "explain how 30 is 2x3x5", where the verified information plugged as a prompt may be obtained from a curated system and the natural language exposition could be finally written by a conversational system.

I don't foresee knowledge absorbed by LLM, but some form of combination between both techiques. Consider the times tables, the chemical elements, or lots of well known and established knowledge pieces. Is there any advantage in texting all that structured information to afterwards gradient descent train on it? Not to mention algorithms, from Viterbi to Quick Sort to the Fast Fourier Transform. Those look like specialized intelligence modules to be interfaced by Large Language Models, rather than (re)learned from scratch.

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ The idea of combining logical or rule-based AI with statistical AI is not new. Even before these GPTs, there were already people claiming that we need to combine both approaches. I agree that ultimately an AI will need to use, and understand when and how to use a tool like a knowledge graph or Wolfram Alpha in order to be reliable. Humans have some kind of database of facts in their heads, which we use when we need. Unfortunately, many people think they can just differentiate end-to-end and get a magical model that works like a human, without even knowing how a human brain works. $\endgroup$
    – nbro
    Commented Feb 11, 2023 at 10:54
  • $\begingroup$ To make this answer even more useful, I would recommend that you provide the links to the blog posts that you mention. $\endgroup$
    – nbro
    Commented Feb 11, 2023 at 10:55
0
$\begingroup$

It is correct that curation and annotation are crucial to knowledge graph. At the same time, such annotation has been accumulated intensely in few areas like medical and manufacturing (some publicly, and some internally within the organization) - partly accelerated by the need of data interoperability and standardization within the industry.

So while it may not be very ready yet for generic use cases, some form of knowledge graphs/ontologies are already utilized for a long time in domains mentioned earlier.

Besides that, the current active research on knowledge graph generation/inference will potentially increase the breath and depth of the graph in a more scalable way.

Improve this answer
$\endgroup$
3
  • 1
    $\begingroup$ Have you ever compared ChatGPT-like LLMs with knowledge graphs in such domains? $\endgroup$
    – Lerner Zhang
    Commented Feb 10, 2023 at 10:29
  • $\begingroup$ If you also want to produce language models that are reliable, you also need to "curate" them (either with RL from human feedback or whatever). I even claim that you need more "curation" to make models like GPTs be more reliable and thus useful than knowledge graphs because they work in ways that are less understandable than knowledge graphs, which may also be built using ML techniques, just to clarify. In the end, a knowledge graph is some kind of graph database, which contains hopefully true facts, but it's not exhaustive. $\endgroup$
    – nbro
    Commented Feb 11, 2023 at 10:38
  • 2
    $\begingroup$ OTOH, models like GPT are stochastic parrots that seem to be able to produce a lot of stuff, but that may spit out misinformation, and you wouldn't know if you are unaware of the topic, unless you check yourself in a reliable source (which immediately implies that ChatGPT is not reliable, which is clearly true), but at that point why are you using ChatGPT if you can't rely on it? Maybe some form of creative art? Or maybe to fool people? That most useful use of these models is to fool people. It may be good for politicians. $\endgroup$
    – nbro
    Commented Feb 11, 2023 at 10:47

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .