crossing

"crossing" originally started as a software project by Dennis Ulmer and Sebastian Spaar during the summer semester 2014 at Heidelberg University, Germany.

In theory, crossing tries to create a transformation matrix from one Vector Space Model in language A to another one in language B using a provided dictionary (for instance, German-English). Then -- taking an unknown vector v in language A (a word not found in the dictionary) -- crossing can transform that vector into language B using the calculated transformation matrix and looking for the most similiar vector in language B.

During the software project, crossing was used to analyze anglicisms found in the German language, and whether that anglicism's meaning has changed compared to the original English word (hence, "CrOssinG" -- CompaRing Of AngliciSmS IN German).

Vector space models were created by using word2vec on an English and German Wikipedia dump, that were converted to plaintext characters beforehand using a slightly altered version of WikiExtractor.py. These tools can be found in the opt/ directory.

Many thanks to http://www.dict.cc that provided us with a German-English dictionary.

Installation

crossing requires the following Python packages:

• NumPy
• SciPy
• scikit-Learn
• nose (a requirement of scikit-learn, sometimes needed for installation)
• BeautifulSoup (for the scripts found in bin/)

Use pip install -r requirements.txt to install crossing and its requirements. Using a virtual environment is recommended for not spamming your system packages with a small software project.

Usage

crossing's usage can easily be learned by using it interactively in a Python interpreter. Make sure to install crossing and its dependencies, open a Python interpreter and import it:

>>> import crossing

There is some example data prepared in the share/ directory:

share
├── de.txt
├── de_dummy.txt
├── de_vectors.txt
├── dict.txt
├── dict_dummy.txt
├── en.txt
├── en_dummy.txt
└── en_vectors.txt

Of these files, de_vectors.txt, en_vectors.txt and dict.txt are of particular interest. They are based on the corpus "Town Musicians of Bremen" found in de.txt/en.txt. Let's create a VectorTransformator object that will serve sevel vector transformation matrices:

>>> vt = crossing.VectorManager.VectorTransformator()

We have to fill our vt object with some language data. vt has three variables that need to be filled: vt.V and vt.W represent two vector spaces, and vt.Dictionary contains the translation of the words found in vt.V to vt.W. For this example, use the data found in the share/ directory and load them into vt using the functions of FileManager.py:

>>> vt.Dictionary = crossing.FileManager.readDictionaryFile("share/dict.txt")
>>> vt.V = crossing.FileManager.readWord2VecFile("share/de_vectors.txt")
>>> vt.W = crossing.FileManager.readWord2VecFile("share/en_vectors.txt")

(Since we are working with word2vec data, FileManager.readWord2Vec() is used. However, you could pass every dictionary in the following format to vt.V/W:)

{"word" = [1.0, 2.0, 3.0, ...], "another" = [0.1, 0.2, 0.3, ...], ...}

Remember that VectorTransformator only wraps several transformation matrices. This way you could create different transformation models and compare their accuracies. Let's create a transformation matrix now -- by default, sklearn.Linear_Model.Lasso with alpha = 0.1 is used (refer to the docstring to see other models):

>>> vt.createTransformationMatrix()

Let's have a look at the word katze (German for cat). Its vector form is, in German and English respectively:

>>> vt.V["katze"]
[0.006136, -0.052587, 0.012688, -0.01403, -0.046991, 0.042845, -0.023529, -0.001199, 0.034139, -0.003296]
>>> vt.W["cat"]
[-0.067114, 0.033746, 0.020565, 0.032246, 0.113999, 0.016741, -0.021005, 0.043264, 0.060346, -0.008794]

We can now see how crossing would transform the vector for katze into the English vector space, using the transformation matrix that was just created:

>>> vt * "katze"
(matrix([[-0.01070324],
        [-0.00699281],
        [ 0.00408598],
        [ 0.00868466],
        [ 0.03515451],
        [-0.00209241],
        [-0.02295664],
        [ 0.01283001],
        [ 0.01598752],
        [-0.00638645]]),)

Summary

Most of the time, when using vector information from word2vec and sklearn.Linear_Models, our algorithm fails miserably to create an adequate transformation matrix. One reason might be that the information provided by word2vec is not useful for creating a vector space model of a language, since word2vec is more of a straightforward approach of representing words by a numerical value.

Using dummy data, like the _dummy files found in share/, creating transformation matrices works fine.