To run the irc version: ./epichef server_name channel_name
To run a command line bot: ./ircharness.py
There are many aspects to this project. At the base level, a sentence is
passed to a parser, the output of which is passed to the response generator.
There is however a lot that goes on behind this process.
First, the input sentence is put through a tagger. This uses the information
schema to tag each word's type. This then goes to a lexer/parser combo.
After the parser returns its output, the responder start generating sentences.
It uses a variety of loose templates and fills them in with related
information from a relevant category or topic.
The responder takes a Sentence Object as input and extracts portions of that to generate responses that try to sound natural. It handles questions and statements from the parser. If it is a question that isn't classified in the 5 W's or H, the responder will know not to answer with another question. There is also a specific set of responses tailored for questions that fall under the 5 W's and H category. Otherwise, if the user makes a statement to the bot, it will choose from all the possibilities it has.
That explains the 5 sections that the responder is segmented into. It was designed to either ask a question back (in two cases), give a comment, make a suggestion, or say goodbye. One of the more unique ones is making a suggestion. This will take a keyword in the sentence and grab a relevant recipe from the internet and suggest it to the user.
A challenge that came with writing the responder was making the responses fit nicely with the input from the user with the limited amount of information we can extract from a sentence. It took our group a long while to come up with responses that fit a large portion of sentences that could be given.
The parser takes a tagged string as input and outputs a Sentence object. Sentence objects and their components (Agent, Action, and Predicate objects) are detailed in the semantics.py file. Unfortunately,
there is a lot of redundancy in semantics.py due to the way the ply parser works. Short of hacking through the ply code, building sentences up from classes representing bits and pieces of sentences (with a
lot of redundant information being passed around) was the only way to extract all the relevant information from the input sentence using the ply parser. The ply parser is the last of four parsers we tried
to use. The first, a top-down parser, failed because English is not that simple. The second, a shift-reduce parser, failed because something went wrong in its parse table, and it ended up getting stuck in
one particular state. The third, a naive parser that took for granted that all sentences were of the form determiner+adjective+noun+adverb+verb, was only used so that we had something to work with when we
started the responder.