Matching with Synonyms using KeyLLM OR KeyBERT

[ChettakattuA]

I have been playing with KeyBERT and KeyLLM for a while now. And here is something I would like to achieve.

If have a text "CO2 emissions are high these days" and a list of candidate words, which might contain the word Carbondioxide and not CO2 will KeyBERT or KeyLLM find Carbondioxide as a match?

Text = "CO2 emissions are high these days"
candidate keyword list have the word ["Carbon dioxide"] and not "CO2"

Expected output = ["Carbon dioxide"]

[MaartenGr]

If have a text "CO2 emissions are high these days" and a list of candidate words, which might contain the word Carbondioxide and not CO2 will KeyBERT or KeyLLM find Carbondioxide as a match?

I think it should be possible if you use it as a candidate word. Have you tried it out?

[ChettakattuA]

In this result the acronym and synonyms are not identified by KeyBERT

acronym used = CO2 -> carbon dioxide
synonym used = emission -> release
Plural = emission -> emissions 

The code used

from keybert import KeyBERT 
kw_model = KeyBERT() 
text = "CO2 emissions are high these days"
can = ["carbon dioxide", "emissions","release","emission","co2"]
Keywords = kw_model.extract_keywords(text,candidates=can)

Is there some way to resolve this?

[MaartenGr]

Ah right, that's because the candidates should appear in the original document in order to find them. Instead, you might want to use the seed_keywords parameter which allows you to steer the model towards certain words. Note that you might have to use the global perspective here.

[ChettakattuA]

But do you know why its require the word itself to appear in the text? What I understood from the documentation is it uses embeddings and cosine similarity. Aint it enough to understand similar words or synonyms from the text and candidates?

[MaartenGr]

@ChettakattuA That depends on what you want. Generally, keywords are derived directly from the article that was written for SEO reasons. In KeyBERT candidates are passed to the CountVectorizer as a vocabulary, which means they should appear in the original documents (as they are fitted on the original documents):

KeyBERT/keybert/_model.py

Lines 163 to 182 in f0f96a6

	# Extract potential words using a vectorizer / tokenizer
	if vectorizer:
	count = vectorizer.fit(docs)
	else:
	try:
	count = CountVectorizer(
	ngram_range=keyphrase_ngram_range,
	stop_words=stop_words,
	min_df=min_df,
	vocabulary=candidates,
	).fit(docs)
	except ValueError:
	return []
	# Scikit-Learn Deprecation: get_feature_names is deprecated in 1.0
	# and will be removed in 1.2. Please use get_feature_names_out instead.
	if version.parse(sklearn_version) >= version.parse("1.0.0"):
	words = count.get_feature_names_out()
	else:
	words = count.get_feature_names()