A natural language word problem solver written in Python
Python Shell
Latest commit 2a2ae77 Feb 28, 2014 Eric Garside forgot calibration. moeny money



A word problem solver.


  • Python 2.7+
  • SymPy
  • NumPy
  • NLTK

How it works

Zoidberg solves word problems in three distinct steps.


Zoidberg uses a triple interpretation approach to parse word problems. The raw text of the word problem is iterated over three times, each cycle of the iteration designed to pull out or piece together different information about the problem at hand.

Solution inference

What implications does the problem make about itself?

The first analysis of the problem uses the Inference engine to make some quick judgements on what the problem is about. Without going into any specific detail about the problem itself, the Inference engine provides some quick and dirty directionality for the query parser.

For example, a word problem containing the word "another" might be an addition problem; one containing the word "fewer" might be subtraction. The output of the inference engine isn't designed to be right, it's designed to be fast.

This step is intended to model the cognitive process of a rough first impression of the question being asked; it is expected the first impression will often be wrong or incomplete, but even then should be invaluable to the query parser for actually targeting the real question.

Query parsing

What is the question being asked?

The second analysis of the problem uses the Parsing engine to determine the question actually being asked. The parser marries the general impressions of the Inference engine with specifics about the problem itself.

A properly parsed query will be a mathematical formulation of the question being asked, in symbolic notation, which SymPy could then solve for us.

Data extraction

What salient data is proved; what data is missing?

Once the question we need to answer is known, the Extraction engine attempts to fill in whatever data is missing from our expression such that it can be properly solved.

The Extractor exclusively uses the query, and attempts only to fill in missing data from the query. A properly extracted problem will be an expression.


Once an expression has been defined, it can be solved. SymPy is used to handle expression solving. To the best of its abilities, Zoidberg will attempt to respond with a humanized, textual version of the answer.

An ideal solution is "Jane has 6 balloons" instead of "6" or "6 balloons".


Arguably the most important part of the Zoidberg model is the ability to learn new things. A ~/brain.zoidberg.json file will be created in your home directory to store learned behaviors.

The learning process is heavily involved in modifying the default functionality of all it's engines.

Why is it named Zoidberg?

Need a project name? Why not Zoidberg?

Tagging reference

PoS Tagging reference list: ftp://ftp.cis.upenn.edu/pub/treebank/doc/tagguide.ps.gz