Form Latent Semantic Indexing to Language Models and Back

One of the key challenges in information retrieval is the problem of automated indexing. How can computers be used to automatically extract relevant index terms from documents? How should documents be represented to facilitate information access? Primarily, a good document representation should capture the topical and semantical relationships between documents. Thereby, it should support the computation of similarities between documents and queries or other documents. From the early years of information retrieval, it has been realized that automated indexing should get to the semantic level of the meaning of words. An important example is the idea of notional families in the work of H.P. Luhn [5]. Ideally, notional families group together words of similar and related meaning and use these concepts to encode documents. In this paper, we will put our own work [4] in a new context and show how a combination of ideas from latent semantic indexing [2] with the language modeling approach to information retrieval [6] leads to a statistical retrieval model that is very close in spirit to notional families.