An Architecture for Scientific Document Retrieval Using Textual and Math Entailment Modules

We present an architecture for scientific document retrieval. An existing system for textual and math-ware retrieval Math Indexer and Searcher MIaS is designed for extensions by modules for textual and math-aware entailment. The goal is to increase quality of retrieval (precision and recall) by handling natural languge variations of expressing semantically the same in texts and/or formulae. Entailment modules are designed to use several, ordered layers of processing on lexical, syntactic and semantic levels using natural language processing tools adapted for handling tree structures like mathematical formulae. If these tools are not able to decide on the entailment, generic knowledge databases are used deploying distributional semantics methods and tools. It is shown that sole use of distributional semantics for semantic textual entailment decisions on sentence level is surprisingly good. Finally, further research plans to deploy results in the digital mathematical libraries are outlined.