WordNet Ambiguous Words
In July of 2026 while working on the knowledge graph for the storage of chunks in a databse for my Deterministic AI project because before the GPTs there was work for a deterministic AI and all kinds of libraries to support that goal were created. However, I hit a snag in the knowledge part, how do you know what a word means, for instance Mercury, can be a planet, a chemical element, a god, and several other things.
So I asked Gemini 3.5 on my phone and after a discussion; I decided the library to use was called WordNet. So I then was curious, could local AI’s come to the same or a better conclusion, and after testing 12 of them. I found Gemma 4 26B A4B did the best job of explaining the problem and giving the most resources available for the topic beating in my mind Gemini 3.5 Flash. Below are the prompts given to all local AIs, and the system prompt they used. Also I used a temperature of .6 and used Q4_K_M models for all but LFM 1.6B, if you want to reproduce my results.
Below the system prompt and prompts, are links to the Gemini 3.5 Flash, and Gemini 3.5 Thinking discussions (I tested 3.5 thinking after the local testing), and the 12 markdown files showing the prompts, thinking, web searches, and responses.
I was using Brave-Search in my testing, I know that using a different search engine may have produced different results expecially for Minstral 3B. It is also interesting how many AIs ran down the same track as Gemini 3.5 Flash.
System Prompt:
-If you don’t know something, or the user is searching for something you think does not exist, do a web search.
-if you use brave-search use long tail keywords or sentences for searches.
Prompts:
How do LLMs handle a word that can be used different ways? One of the words I’m thinking of is mercury. It can be a planet, it can be a god, it can be a chemical. All right. Yeah, a chemical or it’s a chemical element. I’m also thinking of words like age. Like what is your age? It talks about how many years old typically that you are or something is or. But there’s also age as in the industrial age where it refers to a period of time. How do LLMs tell the difference between those kinds of words?
So how would I find out what is around a word? In other words, the word Mercury is used in a piece of text. How would I outside of the LLMs matching that text to a semantic area? Even if I had the semantic area, how would I know that that’s planets and that sort of things versus another? And not all words have multiple meanings. Many words have just a single meaning. Is there a dictionary somewhere of words that have multiple possible meanings?
Does WordNet also handle multi word nouns? So, for instance, Paris, France versus Paris, Texas.
So, Paris, France is a bad example of this. But if I have a word that I pull out as a noun or being used in a sentence as a noun, and I send it to WordNet, and that word may be a single word, it may be multiple words that are used as a thing. I’m thinking of like maybe regions like East Coast, Southeast, those kinds of areas. Um, those multi-word would be actually considered a single noun, and it would come back and let me know whether there’s only one definition or there’s multiples. Is that correct?
Okay, so my actual application of this is loading a knowledge graph because in a knowledge graph, I have to make sure that each of my nodes are unique nodes. And I need to know if something like Mercury is this talking about a planet, a god or an element. And I need to know not necessarily in the sentence may not have enough context for an LLM when it’s breaking it out to know that that’s the case. But if I knew what the possible definitions were, I could go back and look at the text that it came from and then I could uh send it off and find out if there’s multiple meanings. If there aren’t, then find the node goes in as is. If there’s multiple meanings, then I know that there’s another research step I need to do to figure out what it is or I need to ask the user to identify what it’s related to.
How does the built-in ontologies work with WordNet?