/ppdb

Interface for reading the Paraphrase Database (PPDB)

Primary LanguagePythonMIT LicenseMIT

PPDB

This module contains functions for reading the Paraphrase Database (PPDB), an automatically generated database of paraphrases in different languages.

Long story short, PPDB was generated by selecting words and phrases that were translated in the same way to English. Thus, a known problem with it is that many entries are not real paraphrases, but variations in gender, number, case or other morphological subtleties not present in English.

This package provides both an easy interface to use the PPDB in your code as well as entry points to define filter functions.

Filters

Currently, there are only implemented filters for Portuguese. They discard paraphrase rules that only change gender or number, as well as leading or trailing articles and commas in the rules.

Here is an example of a filter function that you can implement:

def my_filter(lhs, rhs):
    """
    Return True if the pair should be filtered out (it is not a relevant
    paraphrase); otherwise return False.
    """
    stripped_lhs = strip_suffix(lhs)
    stripped_rhs = strip_suffix(rhs)

    if stripped_lhs == stripped_rhs:
        return True

    return False

def strip_suffix(word):
    # language-specific logic

Then you call ppdb.load_dict with it:

import ppdb
ppdb_rules = ppdb.load_ppdb(path, my_filter)

Loading a PPDB file and filtering pairs can be time consuming, especially for the larger ones. For this reason, I recommend using pickle to serialize a TransformationDict after it is created, so the next time it can be loaded much faster. If you pass a path ending in .pickle, ppdb.load_ppdb() will just load it and ignore the filtering logic.

If you want to use the existing Portuguese filters, import ppdb_pt:

from ppdb import ppdb_pt
ppdb_rules = ppdb.load_ppdb(path)

And if you happen to write filter functions for another language, please submit a pull request!

Singleton Usage

Once the dataset is read, ppdb stores the same object it returns as a singleton. You can then call ppdb.get_rhs() to get the RHS of a given LHS.

In order to replace the singleton object inside the module, call ppdb.load_ppdb() with force=True.

TransformationDict

The paraphrase rules are stored in a data structure called TransformationDict, which is a subclass of Python's built-in dict. The TransformationDict returned by load_ppdb maps the left-hand side (LHS) of the rules into the righ-hand sides (RHS).

The values stored by the dictionary are tuples with the set of RHS for that LHS and another TransformationDict with possible continuations of the LHS for other RHS.

Confused? Let's start simple. Suppose there are two paraphrase rules:

A -> X
A B -> Y

A TransformationDict storing it would look like this:

>>> ppdb_rules
{'A': ({('X',)}, {'B': ({('Y',)}, {})})}
>>> rhs, more_rules = ppdb_rules['A']
>>> rhs
{('X',)}  # a set with the only RHS for "A"
>>> more_rules
{'B': ({('Y',)}, {})}  # more nested stuff
>>> rhs, more_rules = ppdb_rules[('A', 'B')]
>>> rhs
{('Y',)}  # a set with the only RHS for "A B"
>>> more_rules
{}

If you only want the RHS for a specific LHS, you can use get_rhs(), like in the singleton usage:

>>> ppdb_rules.get_rhs('A')
{('X',)}