"Which technologies does this video talk about?" sounds like a named-entity task, and it is, with a twist: the entities are product names that are short, ambiguous, constantly new and often misheard. This post covers the three problems you meet, and a practical approach to each.
Problem 1: what counts as a technology?
If you do not decide, the model will. Without a definition you get everything from "JavaScript" to "the cloud" to the presenter's company. Write the definition into the prompt:
- Include: named products, frameworks, libraries, languages, services, CLIs and protocols
- Exclude: generic words ("API", "database", "framework"), company names that are not products, channel names, people
Then maintain a block list of the noise you actually observe. In our runs, recurring false positives included generic words such as "API" and "Google", and the names of channels.
Problem 2: the same tool, many spellings
Speakers say "Postgres", "PostgreSQL", "Postgres DB". Captions write "Next JS", "next.js" or "Nexjs". Two things help:
-
Normalise: lower-case, remove spaces and punctuation, strip common suffixes (
js,db) to create a matching key - Keep an alias list for the display name and for deduplication
Watch for duplicates that differ by one letter. We saw a tool appear twice because one occurrence was misspelt; a fuzzy match on names within the same video catches it.
Translation can also distort names. If your pipeline summarises in another language, instruct the model to keep product names exactly as spoken. A product name turned into a translated phrase is no longer searchable.
Problem 3: mentioned does not mean recommended
A technology can be:
- Used in the demo
- Recommended as an alternative
- Compared against another
- Criticised or dismissed
Classify the mention type, or at least keep the quote so a reader can judge. This is the difference between a useful tool list and a misleading one.
Ambiguous names need context
"Neon", "Serena" and "Phantom" are all valid product names, and all share their name with something unrelated. Give the model the surrounding sentence, and when you look up the official site, pass the category as well. A good result for an ambiguous name needs context at both extraction and resolution.
Test it on real videos
Pick ten videos in your field, run the pipeline and read the list yourself. You will quickly see what kind of noise appears and where the prompt needs a rule. Update the block list and the prompt, run again.
Our implementation
We built these rules into VidScope, which returns the tools mentioned in a YouTube video with a timestamp and an official link, and avoids returning a link when the match is unclear.
FAQ
Is this a named-entity recognition problem?
Yes, with the twist that names are short, ambiguous and often misheard in captions.
How do I stop generic words being returned?
Define what counts as a tool in the prompt and keep a block list of observed noise.
How do I handle different spellings of the same tool?
Normalise names to a matching key and keep an alias list.
Top comments (0)