Open source release for Google reranking technology
Google has released a general purpose framework for reranking problems, ReFr (Reranker Framework), as open source. Reranking is a technique that is used when there is a model that can offer several scored hypothesised outputs; rerankers can reorder the ranked outputs based on information not available to the original model.
For example, a speech recognition system may be able to identify various phrases from audio input in its trained model, but better results can be achieved by applying a reranker that reorders the results using information about the context the audio was generated in.
ReFr is a framework designed to allow exploration of the possibilities of reranking on different datasets. To that end it has a C++ based implementation with a runtime configuration system to make experimenting easier. For situations where more processing power is needed, it can also make use of Apache Hadoop's parallel processing, to train and use reranking models on a large-scale. The BSD-licensed project is available on Google Code and further documentation is available on the project's GitHub repository.