Hello, Good, I will look at your path over the weekend and get back to you with any more specific comments/suggestions. With respect to the plans, be aware that the SST approach I mentioned is addressed as a sequence labelling problem, not as a classification problem. Instead of learning a classifier for each word as in supervised WSD, sequence labelers learn models to tag sequences of tokens. In practical OpenNLP terms, this means that the implementation of a Super Sense tagger is closer to the NameFinder than to the Doccat.
I think that is a priority to have an all words supervised WSD component working as soon as possible with the usual functionalities (train, tag, evaluate) rather than starting with another paradigm such as SST. In this way we can start making tests and evaluating it on the senseval etc. datasets to know more about the performance of the component. (there are some typos in the javadoc of your patch) Best, Rodrigo On Thu, Jul 2, 2015 at 1:19 PM, Mondher Bouazizi <mondher.bouaz...@gmail.com> wrote: > Dear Rodrigo, > > Kindly find below the details of my progress on the supervised WSD module as > well as the plan for the remaining period of time. > > Current state: > -------------------- > - Implementation of the IMS approach (Done). Please refer to my previous > email for details about the implementation. Please also find the up-to-date > patch attached in [1] > - Documentation of the implemented parts related only to IMS (Done) > - Make a tester to run the approach (Done). It is included in the patch > [1]. > - Collect data used for training: data are collected from Senseval-3 > (Done). > > Next Step: > -------------------- > - Clean the code and add more parameters (e.g., whether the > disambiguation is coarse-grained/fine-grained disambiguation, etc.) > - Implement the all-words-disambiguation > - Create an interface so different users can input any training data. > Particularly, I will work on semcore, and implement the import of training > data from it. > - Implement further approaches including the SST approach (regular > supervised approaches relying on the extraction of textual features from a > regular training data won't be a problem since they will present a simple > variation of the IMS approach). > - Synchronize my work with Anthony's and clean the code. > > Yours sincerely, > > Mondher > > ===================== > [1] https://issues.apache.org/jira/browse/OPENNLP-757 >
