Automated document labeling uses machine learning to scan, classify, and tag documents without manual effort. Labels are later used for document search. It works differently from keyword search, which uses exact keyword matches to find documents, thus requiring to know what is written in document. Machine learning based labeling, on other hand, allows document containing words “cats and dogs” to be found by contextually similar words such as “cat”, “dog”, “animal”, and “pet”, as well as “cats”, “dogs”, “animals”, and “pets”.

Automated labels help administrative and research roles working with large amounts of documents to navigate the document repositories. It allows such roles to find related documents without having read them before, or by searching with inaccuracies. It also allows document viewers to see 30 most relevant labels for any specific document.

During document upload, most types of office documents (text, word, excel, csv, pdf, etc.) gets converted to text representation. Text representations then gets converted to a list of words, and after that each word gets tagged using machine learning model with a relevance score. All of the tags gets summed with final relevance score and sorted from most relevant to least relevant tag. Top 10% of tags gets assigned to document as labels that can be used for advanced search.

Sparse labeling algorithms struggle to capture deep conceptual meaning, limiting tags to contextually similar words from trained data-sets. Sparse labeling also fails to account for spelling errors both in documents and search queries. Also, in case of large documents, only maximum of 1024 labels can be calculated and stored with a document.