pysdcxx 0.1.4

Sørensen–Dice coefficient on string bigram multi-sets (in C++)

This library may be used to compare strings by a similarity measure, defined as Sørensen–Dice coefficient calculated on multi-sets of the strings' "bigrams" (all couples of adjacent characters).

Bigram (multi-)set is relatively easy to generate (O(n) time complexity lower bound in terms of the string length).

Also, compared to just using a (multi-)set of characters, bigrams do retain certain level of word structure. E.g. backwards spelled words are not deemed similar (unlike when single character set is used).

Just note that single-character words are all perfectly similar in this metric, as their bigram sets are empty---and therefore the same. It may be a good idea to augment words like that with an additional character (like a whitespace) to mitigate that problem.

Using implementation of bigram multisets above, a sequence matcher implementation is also available. Given a sequence of text tokens, the matcher then allows "fuzzy" matching of strings (using their bigrams again) to sub-sequences of the text tokens.

See detailed documentation in doc directory. Rendered documentation:

• https://htmlpreview.github.io/?https://github.com/vencik/libsdcxx/blob/master/doc/sequence_matcher.html

List of features

• Bigram multi-set objects implemented as sorted lists of character tuples with count (O(n log n) creation time complexity in terms of the string length)

• Union operation has O(m+n) time complexity (sum of multi-sets' cardinalities at most)

• Intersection doesn’t produce objects; only its size is calculated in O(m+n) time

• Template implementation, allowing for both ASCII/ANSI characters and UNICODE characters

• Implementations using std::multiset and std::unordered_multiset also available

• Performance tests show that, long story short, the "custom" implementation is the best (notably faster unions, intersection size computation in similar or better time)

• Sequence matcher (using the best performing bigrams) with several optimisations

• Python v3 binding is provided (as pysdcxx module, packaged)

• Python multiset based implementation also compared---and is expectedly much slower

Build and installation

You shall need C++ compiler with support for at least C++17. Recent enough gcc (v8.3 or newer) is OK (older versions might do as well, though). You shall also need cmake and make.

Python v3.7 or newer is supported. For the Python package build, you shall need pip and Python distutils and the wheel package. If you wish to run Python UTs (which is highly recommended), you shall also need pytest.

E.g. on Debian-based (or similar, apt using) systems, the following should get you the required prerequisites unless you wish to use pyenv.

# apt-get install g++ cmake make git # apt-get install python3-pip python3-distutils # unless you use pyenv $ pip install wheel pytest # ditto, better do that in pyenv sandbox</programlisting>

On Mac OS X, you’ll need Xcode tools and Homebrew. Then, install the required prerequisites by

$ brew install coreutils cmake make git</programlisting>

If you do wish to use pyenv to create and manage project sandbox (which is advisable), see short intro to that in the subsection below.

Anyhow, with all the prerequisites installed, clone the project:

$ git clone https://github.com/vencik/libsdcxx.git\</programlisting>

Build the project, run UTs and build Python packages:

$ cd libsdcxx $ ./build.sh -ugp</programlisting>

Note that the build.sh script has options; run $ ./build.sh -h to see them.

Most importantly, run with -s or --enable-pt options to perform performance tests. Performance tests compare computation times of object construction, union operations and intersection size computations of the implementations. If you install multiset Python package (using pip), the perf. tests shall also produce results for pure-Python implementation using multiset.Multiset (not necessary).

The perf. tests are written on the Python level, so the measured times also contain some Python code runtime (and not trivial). I’d expect results in native code to be notably better; but the test code is identical, so the measurements should be meaningful for comparison.

If you wish, use pip to install the Python package:

# pip install pysdcxx-*.whl</programlisting>

Note that it’s recommended to use pyenv, especially for development purposes.

Managing project sandbox with pyenv

First install pyenv. You may use either your OS package repo (Homebrew on Mac OS X) or web pyenv installer. Setup pyenv (set environment variables) as instructed.

Then, create pysdcxx project sandbox, thus:

$ pyenv install 3.9.6 # your choice of the Python interpreter version, >= 3.7 $ pyenv virtualenv 3.9.6 pysdcxx</programlisting>

Now, you can always (de)activate the virtual environment (switch to the sandbox) by

# pyenv activate pysdcxx # pyenv deactivate</programlisting>

In the sandbox, install Python packages using pip as usual.

$ pip install wheel pytest</programlisting>

Usage

C++

Using bigrams

#include <libsdcxx/bigrams.hxx>

using bigrams = libsdcxx::bigrams;                            // wbigrams for UNICODE

const auto bgrms_empty = bigrams();                           // empty bigrams set

const auto bgrms1 = bigrams("Hello world!");                  // construct from string
size_t cnt = bgrms1.size();                                   // number of bigrams

std::cout << bgrms1;                                          // serialisation

for (const auto & bigram_cnt: bgrms1) {                       // tuple of (bigram, count)
    const auto & bigram = std::get(bigram_cnt);

    std::cout << "Bigram : " << std::get(bigram) << std::get(bigram) << std::endl;
    std::cout << "Count  : " << std::get(bigram_cnt) << std::endl;
}

// (Const.) iterators are supported via cbegin, cend and begin, end method calls

const auto bgrms2 = bigrams("Hell or woes.");

size_t isect_size = bigrams::intersect_size(bgrms1, bgrms2);  // intersection cardinality
double sdc = bigrams.sorensen_dice_coef(bgrms1, bgrms2);      // similarity, in [0,1]

auto uni0n = bgrms1 + bgrms2;                                 // 2 bigrams union
auto uni0n = bigrams::unite(bgrms1, bgrms2 /* , ... */);      // variadic union

uni0n += bigrams("more stuff");                               // objects are mutable

Using sequence_matcher

#include <libsdcxx/sequence_matcher.hxx>
#include <libsdcxx/bigrams.hxx>

using sequence_matcher = libsdcxx::sequence_matcher;    // wsequence_matcher for UNICODE
using bigrams = libsdcxx::bigrams;                      // wbigrams for UNICODE

auto matcher = sequence_matcher();
matcher.reserve(10);    // reserve space for bigrams matrix for text of 10 tokens
                        // (reservation is not strictly necessary, but advisable)

const auto bgrms_hello = bigrams("Hello");
const auto bgrms_world = bigrams("world");

matcher.emplace_back("Prologue");   // create token bigrams in-place
matcher.emplace_back(" .", true);   // it's a good idea to pad single-char strings...
matcher.emplace_back("  ");         // to 2 characters (so that they produce a bigram)
matcher.push_back(bigrams_hello);   // push existing bigrams back
matcher.emplace_back("  ", true);   // true here stands for "strip" token; matched...
matcher.push_back(bigrams_world);   // sub-sequences are restricted not to begin/end...
matcher.emplace_back(" !", true);   // with such tokens
matcher.emplace_back("  ", true);
matcher.emplace_back("Epilogue");
matcher.emplace_back(" .", true);

const auto bgrms_helo_wordl = bigrams::unite(           // note thatbigrams of the whole
    bigrams("Helo"), bigrams("  "), bigrams("wordl"));  // sentence would differ

auto match = matcher.begin(bgrms_helo_wordl, 0.7);      // match with threshold 0.7
for (; match != matcher.end(); ++match) {
    std::cout << match << std::endl;    // simple string form of match info, try it

    std::cout
        << "Match bigrams: "  << *match        << std::endl  // sub-sequence bigrams
        << "Match at index: " << match.begin() << std::endl  // beginning token index
        << "Match end: "      << match.end()   << std::endl  // index just past the end
        << "Match size: "     << match.size()  << std::endl  // number of tokens
        << "Match score: "    << match.sorensen_dice_coef() << std::endl;
}

// ... you may of course continue matching other sequences...

Pyton v3

Using Bigrams

from pysdcxx import Bigrams   # Python Bigrams are implemented by wbigrams, support UNICODE

bgrms_empty = Bigrams()                 # empty bigrams set

bgrms1 = Bigrams("Hello world!")        # construct from string
cnt = len(bgrms1)                       # number of bigrams

print(str(bgrms1), f"{bgrms1}")         # string serialisation

for bigram, cnt in bgrms1:              # Bigrams are tuple[str, int] generators
    assert len(bigram) == 2

bgrms2 = Bigrams("Hell or woes.")

isect_size = Bigrams.intersect_size(bgrms1, bgrms2)     # intersection cardinality
sdc = Bigrams.sorensen_dice_coef(bgrms1, bgrms2)        # simiarity, in [0,1]

union = bgrms1 + bgrms2                                 # 2 bigrams union

union += Bigrams("more stuff")                          # objects are mutable

Using SequenceMatcher

from pysdcxx import SentenceMatcher, Bigrams

matcher = SequenceMatcher()             # empty matcher
matcher = SequenceMatcher(reserve=4)    # empty matcher, reserved space for 4 tokens
                                        # (not necessary, but speeds up token addition)

matcher.append("Hello")                 # append token (Bigrams are constructed)
matcher.append(Bigrams("  "))           # append token bigrams directly
matcher.append("world", strip=False)    # "strip" token means that no match may begin...
matcher.append(" !", strip=True)        # nor end by that token

# Alternatively, you may pass `Iterable` of tokens directly to the constructor.
# If the `Iterable` length can be taken, reservation of space is done; otherwise,
# you may still use the `reserve` constructor parameter if you know how many
# tokens there shall be...  Again, if you don't, the constructor will handle it anyway
# (construction may just take a bit longer).
strip = True
matcher = SequenceMatcher([
    "This", ("  ",strip), "uses", ("  ",strip),
    "Sørensen", (" -",strip), "Dice", ("  ",strip),
    "coefficient", (" .",strip),
])

for match in matcher.match(["Sørenson", "and", "Dice"], 0.65):  # matching
    print(f"Match begin : {match.begin}")       # 4 (index of the 1st match token)
    print(f"Match end   : {match.end}")         # 7 (1 past the last match token)
    print(f"Match score : {match.score}")       # >0.65, <1.0 as it's not a perfect match

# You may continue matching other sequences
# Note that this is only a quick summary; see `SequenceMatcher` docstrings for more...

License

The software is available open-source under the terms of 3-clause BSD license.

Author

Václav Krpec <[e-mail]>