Subspace Representations of Unstructured Text

Since 1970 vector-space models have been used for information retrieval from unstructured text. The initial simple vector-space models suffered the same problems encountered today in searching the internet. These difficulties were significantly relieved by Latent Semantic Indexing (LSI), introduced in 1990 and improved through 1995. Starting with the simple vector-space model’s sparse term-by-document matrix, LSI used a truncated singular-value decomposition to obtain a low-rank approximation, reinforcing similarities between documents. This approach stalled, owing primarily to a lack of interpretation for the low-rank approximation and consequently a lack of controls for accomplishing specific tasks in information retrieval. The text mining team in Boeing Phantom Works has taken a broad systematic approach to vector-space models for unstructured text. The two pillars of our emerging technology Trust are the handling of subspaces to capture latent semantics and the assignment of natural-language labels to elements in the subspace. On these, Trust is developing into a consistent natural system for addressing a variety of information-retrieval tasks. Our focus on subspaces enables us to compute with flexible and efficient algorithms. The natural-language interpretation in the subspace allows us to cast the tasks conceptually within linear algebra and yet to retain the association to natural language. Thus the display, controls and parameters of the system can all be revealed to the user in a familiar natural language setting. Our broad approach to subspace representations is described here.