A lemma is the canonical, dictionary form of a word — the headword you would look up in a dictionary. Every inflected variation of a word maps back to a single lemma. The forms "run", "runs", "ran", and "running" all share the lemma run; "better" and "best" share the lemma good; "mice" has the lemma mouse. The process of reducing a word to its lemma is called lemmatisation.
Lemmatisation matters for any analysis where you care about meaning rather than surface form. In word-frequency work, counting "run" + "runs" + "running" + "ran" separately scatters what is really one concept across four low counts; folding them into the lemma "run" gives a single, more meaningful frequency. Search engines lemmatise so that a query for "running shoes" can match a page that says "shoes for runners". The same idea powers topic analysis, sentiment analysis, and document similarity.
An important distinction is lemmatisation versus stemming. Stemming chops affixes with crude rules and can produce non-words — a stemmer might reduce "studies" and "studying" to "studi". Lemmatisation is smarter: it uses vocabulary and part-of-speech information to return a real dictionary word, so "studies" becomes "study". That accuracy comes at a cost, because a lemmatiser must know a word's part of speech to get it right — "saw" is the lemma "see" as a verb but the lemma "saw" (the tool) as a noun. Lemmatisation runs after tokenization and often alongside stop-word removal in a text-analysis pipeline, feeding cleaner word counts and frequency lists. See it applied in our word frequency counter.