TF-IDF

BM25 and TF-IDF are classic representatives of lexical retrieval, based on literal keyword matching and term frequency statistics.

BM25是一种基于词频统计（TF-IDF改进版）的经典信息检索算法，属于稀疏检索方法

Traditional search systems mostly rely on statistical algorithms like TF-IDF or BM25 that rank results by term frequency and uniqueness in documents

TF-IDF represents each email as a sparse vector weighted by term frequency and inverse document frequency; logistic regression is a linear classifier that outputs class probabilities by learning a weight for each word feature; when label volume is limited, this combination has few parameters, is not prone to overfitting, trains and infers quickly, and has interpretable weights.

Keyword matching splits a query into terms and relies on inverted indexes and statistical signals such as TF-IDF or BM25 to judge relevance; semantic matching encodes the entire context into a dense vector and measures semantic closeness with cosine similarity.

BM25 builds on TF-IDF by introducing document length normalization and term frequency saturation, making it more robust for job descriptions of varying length.

The keyword retrieval simulator uses lexical matching (based on statistical models such as BM25 or TF-IDF) rather than semantic similarity of embedding vectors, so it cannot predict the actual results of a vector store.

DSupport's RAG implementation splits articles by paragraph, uses TF-IDF for word-weight vectorization, optionally integrates VoyageEmbedder for semantic retrieval, and uses a relevance cutoff of minScore = 0.25 in SearchKbTool.

The authors used an LLM judge to identify suspicious reports and obtained an AUROC of 0.48 (no better than random), whereas a simple TF-IDF word-frequency classifier achieved an AUROC of 0.65.

A common approach to relevance ranking is a TF-IDF/BM25-like weighted model that also considers field weights; MiniSearch allows configuring weights for different fields and supports applying different score coefficients to prefix-match and fuzzy-match results.

TF-IDF/BM25 retrieval fails when there is vocabulary drift (the same concept expressed differently), which is a known limitation of the current approach and needs embedding models or query rewriting to compensate.

The system's retrieval is entirely based on TF-IDF/BM25 done in-process, without using a vector database or embedding server, sacrificing semantic generalization ability in exchange for interpretability and deployment simplicity.

The classifier uses both TF-IDF word features and sentence vectors generated by all-MiniLM-L6-v2, concatenating the two before training a balanced logistic regression classifier.

TF-IDF uses the title plus up to 1,500 characters of the description, while the sentence vectors use the title plus up to 1,000 characters of the description.

The study introduced human players and TF-IDF as control group baselines.

TF-IDF is a keyword weighting method that performs only lexical-level matching without understanding semantics; as a non-neural-network baseline, its performance ceiling is lower than that of embedding models or large language models.

For tasks with short text and clear category boundaries (such as spam filtering), the TF-IDF approach is often sufficient; BERT fine-tuning yields more significant gains when fine-grained semantic understanding is required.

The traditional TF-IDF + logistic regression approach does not rely on pre-training; it converts text into term-frequency statistical vectors and trains a linear classifier, offering fast speed and low resource consumption, but weaker semantic understanding than BERT.

Classical machine learning methods (such as logistic regression, SVM, and Naive Bayes), combined with pre-computed text features such as TF-IDF or fixed word vectors, have extremely small parameter counts, with CPU inference latency as low as the millisecond level.

In TF-IDF, IDF measures the rarity of a word across the entire corpus; the more documents a word appears in, the lower its IDF value.