Conceptual schema extraction using POS annotations and weighted edit distance algorithm

Database design process involves analysis of system requirements described in natural language for manual extraction of conceptual schema. This is a tedious process and prone to human error. Earlier approaches of automation of this process had made use of either a finite set of rules, Context Free Grammars (CFG) or semantic understanding. Rule and CFG based approaches are not robust enough to cover all possible scenarios; whereas semantic approaches are not generic and have domain dependencies. We have defined an approach where the sequence part-of-speech (POS) tags of a sentence are annotated to ER components and a set of such annotated sentences serves as our corpus. Use of POS tags instead of the actual terms in a sentence makes the approach more robust, generic and domain-independent. We have also defined our own algorithm, which takes an input sentence and uses an extension of Edit Distance technique, to find out similar matches for a POS sequence of an input sentence with an associated cost, if a perfect match is not found. The accuracy of our current system is 54%. The feedback provided by the user is used to update the underlying model making the approach more interactive and improving the accuracy of future predictions.