TF-IDF (Term Frequency-Inverse Document Frequency) is a statistical method used in NLP to quantify the importance of a word in a document relative to a collection of documents (a corpus). It helps identify words that are significant in a specific document but less common across the corpus, making them useful for tasks like search, text classification, or keyword extraction. The method combines two metrics: term frequency (TF), which measures how often a word appears in a document, and inverse document frequency (IDF), which penalizes words that appear too frequently across many documents.
The TF component is calculated as the number of times a term occurs in a document divided by the total number of terms in that document. For example, if the word “algorithm” appears 5 times in a 100-word document, the TF is 5/100 = 0.05. The IDF is calculated as the logarithm of the total number of documents divided by the number of documents containing the term. If “algorithm” appears in 10 out of 1,000 documents, the IDF is log(1000/10) ≈ 2. The final TF-IDF score is the product of TF and IDF (e.g., 0.05 * 2 = 0.1). This ensures common words like “the” or “and” (high TF but low IDF) get low scores, while rare, meaningful terms receive higher scores.
Developers often use TF-IDF in search engines to rank documents by relevance to a query. For instance, a search for “machine learning” would prioritize documents where this phrase has a high TF-IDF score. It’s also used in text classification (e.g., spam detection) to convert text into numerical features. However, TF-IDF has limitations: it doesn’t capture semantic relationships between words (unlike word embeddings) and treats all occurrences of a term equally, ignoring context. Despite this, its simplicity and effectiveness make it a foundational tool for many NLP pipelines.