gnawang / piazzaqa Goto Github PK

The baseline OpenEphyra system contains many features geared toward answering factoid questions. As we restructure the code towards questions typically asked in EECS 280, many of these remnant features from the baseline make Ephyra slow and possible interfere with the answer extract process. Research and remove all non important features.

Focus on...

1. Query Generators
2. Answer Filters

It'd be useful if Ephyra could also respond to queries with related past piazza posts in case the generated answer is not useful or if the student wants to know more. Similarly, adding a reference to particular lecture could also be useful.

Baseline ephyra typically only handles a single question. However, piazza posts typically have more than a single sentence in the query.

There are two ways we can account for this:

1. Leave it as is, and let PiazzaQA return the most relevant answer to the given query
2. Parse the query into its component sentences and run PiazzaQA on each sentence

I opt for the second as it would provide more usefulness for the tradeoff of extra time. However, time would be a huge issue because statements that clearly aren't questions will naturally return without many results from baseline Ephyra.

Figure out the best way to compile and export this program for testing and actual use next Fall.

When IndriKM is used, Ephyra fails to finish its search before my computer freezes and/or crashes. The reason for this is currently unknown, but looking into it could reveal more about whether the issue lies with the IndriKM file itself or the indexed Piazza data.

If Indri proves to have too many issues, or if Solr is just far superior, Solr would be a better place to go for document indexing. The Indri Knowledge Miner was included in the open source package, so it's already tested and approved. We'll have to do extensive research and testing with Solr to see the results. This is a side project for now, but if Indri refuses to work, this issue will become more important.

http://lucene.apache.org/solr/

After getting the desired results from the search, the Ephyra answer extraction pipeline attempts to derive a factoid answer from said results. This significantly limits the spectrum of quesitons we can answer. Research and find a way to modify/remake the pipeline such that Ephyra can questions in a useful manner (information extraction questions)