Explicit semantic analysis (ESA) is a vectoral representation of text used in natural language processing and information retrieval. It represents individual words or entire documents as vectors in a high-dimensional space, where each dimension corresponds to a concept explicitly defined by humans, such as a Wikipedia article. The technique was designed by Evgeniy Gabrilovich and Shaul Markovitch to improve text categorization and to compute semantic relatedness between texts using cosine similarity. The name contrasts with latent semantic analysis (LSA), because ESA's use of a knowledge base allows human-readable labels to be assigned to the concepts forming the vector space.
ESA operates by leveraging a document corpus as a knowledge base. In its basic form, a word is represented as a column vector in the term frequency-inverse document frequency (tf-idf) matrix of the corpus, and a document is represented as the centroid of the vectors of its constituent words. While English Wikipedia is the typical corpus, other collections such as the Open Directory Project have also been used.
Model
The basic ESA variant begins with a collection of texts, such as all Wikipedia articles, numbering N documents. Each document is converted into a "bag of words" - a term frequency histogram - and stored in an inverted index. For any given word, the inverted index returns the set of documents containing that word, with each document scored based on how often the word appears, weighted by the total number of words in the document. Mathematically, this output is an N-dimensional vector of word-document scores, where documents not containing the word receive a score of zero.
To compute the relatedness of two words, their vectors u and v are compared using cosine similarity: sim(u, v) = (u · v) / (||u|| ||v||). This yields a numeric estimate of semantic relatedness. The scheme extends from single words to multi-word texts by summing the vectors of all words in the text, producing a document-level representation.
Analysis
ESA was originally posited under the assumption that the knowledge base contains topically orthogonal concepts. However, Maik Anderka and Benno Stein later demonstrated that ESA also improves information retrieval performance when based on the Reuters corpus of newswire articles, which does not satisfy the orthogonality property. In their experiments, newswire stories served as "concepts". This observation was explained through links between ESA and the generalized vector space model. Gabrilovich and Markovitch responded by noting that Anderka and Stein's result was achieved using a single application of ESA for text similarity and a small, homogenous test collection of only 50 news documents.
Applications
Word relatedness
ESA is considered by its authors a measure of semantic relatedness, as opposed to semantic similarity. On benchmark datasets for word relatedness, ESA outperforms other algorithms, including WordNet-based semantic similarity measures and skip-gram neural network language models such as Word2vec. The approach has been applied in text categorization tasks, where the concept-based representation improves classification accuracy.
Document relatedness
ESA is used in commercial software packages for computing relatedness of documents. Domain-specific restrictions on the ESA model are sometimes applied to provide more robust document matching in specialized fields. The technique's ability to produce interpretable concept vectors makes it valuable for applications requiring transparent representations of textual content.
Extensions
Cross-language explicit semantic analysis (CL-ESA) is a multilingual generalization of ESA. CL-ESA exploits a document-aligned multilingual reference collection, again typically Wikipedia, to represent a document as a language-independent concept vector. The relatedness of two documents in different languages is assessed by cosine similarity between their corresponding vector representations. This extension enables cross-lingual information retrieval and comparison of texts across language boundaries.
ESA has also been connected to broader developments in Machine learning and Artificial intelligence, particularly in the context of Natural language processing tasks. Its explicit, interpretable concept space distinguishes it from later Deep learning approaches such as Neural network embeddings, though both aim to capture semantic relationships in text. The method remains relevant as a foundational technique for semantic representation, complementing more recent models like Large language models and Transformer (architecture) architectures.