Named Entities and Word Representations
This code was created for an experiment to try recognizing German nouns that can refer to persons by their proximity to a class of named entities in a word representation. This is giving a small tutorial how to reproduce this experiment. You find the theoretical insights in the uploaded pdf file. I am assuming a basic understanding of Python and the command line. This experiment was done on Ubuntu 14.04, and I do not know how reproducable it is on other OS. You do need at least 2GB of RAM to run this experiment, even though it is much more convenient with a a lot more.
Required Software
You will have to install some software, unless you already have it.
Clone this Repository
Since you are using github I'm assuming you have git installed. Open your terminal and go to the folder where you want to install the repository.
git clone this_sites_url
Stanford's Named Entity Recognition
This is the weak point. Stanford's NER takes an aweful lot of memory (about 1000 times the input files size). Download from the following site the Named Entity Recognizer and also the German classifiers from further below.
http://nlp.stanford.edu/software/CRF-NER.shtml
http://nlp.stanford.edu/software/stanford-ner-2012-05-22-german.tgz
Follow the instructions there, which is basically unpacking it to some folder where you find it later on.
Google's word2vec
This is a nice piece of software, but read the whole paragraph before you start installing. You find everything you need here
https://code.google.com/p/word2vec/
To follow their instructions you would need Apache's Subversion (svn). If you do not have it you might just try to download the files by hand. You do not have to run the demo scripts, but it is nice.
gensim (Python API for word2vec)
I hope you have pip, then you can just type
$ pip install --upgrade gensim
Otherwise I suggest you install pip or you check out this page
http://radimrehurek.com/gensim/install.html
Downloading the Data
The experiment worked with the Gutenberg Corpus. To download all German books (about 500MB) type
$ wget -t inf -o log -w 2 -m -H "http://www.gutenberg.org/robot/harvest?filetypes[]=txt&langs[]=de"
This is actually not too much data. If you feel like writing another preprocessing script, you can also download the newest German Wikipedia dump from
http://dumps.wikimedia.org/dewiki/
Running the Experiment
Cleaning the Data
With this repository you downloaded some scripts. Run the script to strip the headers and footers of Project Gutenberg and concatenate all zip files into one text file. The output will be saved in a file gutenberg_corpus.txt in the same folder where the Gutenberg corpus is located (if you did not rename it its called www.gutenberg.lib.md.us/)
$ ./gutenberg.py www.gutenberg.lib.md.us/
Creating the Input for word2vec
The repository also contains a script that is running the Stanford NER with the German classifier. You might want to open it and give it more memory. Note that you will probably have to split the Gutenberg Corpus into chunks of about 1/1000 the size of your memory to run this script and concatenate the outcome together. This is a nice practice in Python for you. On each chunk you can run
$ ./path/to/stanfordNER/ner-ger.sh www.gutenberg.lib.md.us/gutenberg_corpus_chunk1.txt > output_file1
Once you are done concatenating everything back together, you can use the following script to convert this text to be read by word2vec (converting umlauts, stripping of unknown characters, lowercasing)
$ ./NER-to-w2v.py input_file > output_file
Running word2vec with Parameters
Now you can create your own binary word representation with word2vec (for an explanation of the variables see their homepage). Here are two different parameter settings to play with:
$ time ./word2vec -train /media/arkadi/arkadis_ext/NLP_data/Rammis_gutenberg/gutenberg_corpus_ner_real_clean.txt -output guten.bin -skipgram 10 -size 300 -window 5 -negative 0 -hs 1 -sample 1e-4 -threads 2 -binary 1 -save-vocab defull-vocab.txt > guten.bin
Vocab size: 262369
Words in train file: 71619017
Alpha: 0.000005 Progress: 100.00% Words/thread/sec: 52.51k
real 70m43.637s
user 113m49.045s
sys 0m26.959s
$ time ./word2vec -train /media/arkadi/arkadis_ext/NLP_data/Rammis_gutenberg/gutenberg_corpus_ner_real_clean.txt -output guten_neg10.bin -skipgram 10 -size 300 -window 10 -negative 10 -hs 1 -sample 1e-3 -threads 2 -binary 1 -save-vocab defull-vocab.txt > guten_neg10.bin
Vocab size: 262369
Words in train file: 71619017
Alpha: 0.000005 Progress: 100.00% Words/thread/sec: 20.70k
real 168m32.287s
user 288m13.476s
sys 0m44.350s
Playing around!
You are finally done with creating your binary word representations?! Start python, or ipython and play!
from gensim.models import word2vec
model = word2vec.Word2Vec.load_word2vec_format('path/to/your/binary_word_repr', binary=True)
model.most_similar(positive=['koenig', 'frau'], negative=['mann'])
model.most_similar(positive=['nertagiper'], negative=[])
model.doesnt_match(['nertagiper', 'nertagiloc', 'nertagimisc'])
model.similarity('nertagiper', 'student')
model.similarity('nertagiloc', 'haus')
model.similarity('nertagimisc', 'nation')
To play around you might as well alter the text file test_set by adding new lines with a word and a tag. Then run
python similarities.py --model /path/to/your/binary_word_repr --test /path/to/your/test_file
What does this actually mean? It means you can compare nouns for their similarity to the person, location and miscellanious tags to to which the might refer.
A first evaluation with the given test set
| word | person | location | misc | match |
|---|---|---|---|---|
| polizist | 0.10749246 | 0.01489580 | 0.03354150 | TP |
| student | 0.20737271 | 0.02934145 | 0.04367834 | TP |
| lehrer | 0.12964790 | -0.12548146 | -0.01557306 | TP |
| lehrerin | 0.08681190 | -0.08561248 | -0.04076674 | TP |
| gott | 0.15922870 | -0.01843374 | -0.08631332 | FP |
| baum | -0.06563630 | -0.05539430 | -0.05834422 | TN |
| haus | 0.06055429 | 0.16598518 | -0.07558561 | TN |
| gross | -0.08251506 | -0.05296010 | -0.01279388 | TN |
| klein | 0.01614168 | -0.02968299 | -0.06549737 | FP |
| nation | -0.12702281 | 0.02772727 | 0.22280701 | TN |
| staat | -0.02131612 | -0.00057563 | 0.07930892 | TN |
| land | -0.01411605 | 0.30458365 | 0.06202623 | TN |
| erde | -0.00483644 | 0.07872798 | -0.01798863 | TN |
| kind | 0.07643456 | 0.00363171 | -0.07258882 | TP |
| mensch | 0.00071798 | -0.12818525 | -0.07854126 | TP |
| mann | 0.22271588 | 0.03623078 | 0.02678832 | TP |
| frau | 0.37619667 | 0.05319752 | 0.02605344 | TP |
| hund | 0.10778271 | -0.08325005 | -0.08226270 | FP |
| katze | 0.02991319 | -0.04702012 | -0.04696872 | FP |
| maus | 0.01681365 | 0.00533434 | 0.01775033 | TN |
| schnell | 0.10562476 | 0.03038920 | -0.03581870 | FP |
| langsam | 0.02893238 | -0.06661929 | -0.02774489 | FP |
| leicht | -0.02133644 | -0.09002297 | -0.05527122 | FP |
| er | 0.39979196 | 0.12971292 | -0.02913975 | TP |
| sie | 0.30369458 | 0.07522399 | 0.01552288 | TP |
| es | 0.15317690 | 0.04054807 | -0.01511156 | FP |
| kreativ | -0.08148986 | -0.09344466 | 0.03209482 | TN |
| sehr | 0.08762989 | -0.02305989 | 0.02555214 | FP |
| morgen | 0.14761052 | 0.11134531 | -0.06228537 | FP |
| Recall | 100.000 | |||
| Precision | 50.000 | |||
| F-score | 66.667 |