This repository contains the implementation of the models in Athiwaratkun et al., Probabilistic FastText for Multi-Sense Word Embeddings, ACL 2018. The paper can be accessed on Arxiv here.
Similar to our previous work in Athiwaratkun and Wilson, Multimodal Word Distributions, ACL 2017, we represent each word in the dictionary as a Gaussian Mixture distribution that can extract multiple meanings. We use FastText as our subword representation to enhance semantic estimation of rare words or words outside the training vocabulary.
The BibTeX entry for the paper is:
@InProceedings{athi_multift_2018,
author = {Ben Athiwaratkun, Andrew Gordon Wilson, and Anima Anandkumar},
title = {Probabilistic FastText for Multi-Sense Word Embeddings},
booktitle = {Conference of the Association for Computational Linguistics (ACL)},
year = {2018}
}
We provide
(1) scripts to train the multi-sense FastText embeddings. We give instructions on how to train the model in 1.
(2) Python scripts to evaluate the trained models on word similarity in 2. Our scripts allows the subword model to be loaded directly into a Python object which can be used to other tasks.
(3) pre-trained model and evaluation script in 3. This section includes intructions on how to load a pre-trained FastText model (single sense) into our format which allows loading as Python object.
1.1 Compile the C++ files. The step requires a compiler with C++11 support such as g++-4.7.2 or newer or clang-3.3 or newer. It also requires make which can be installed via sudo apt-get install build-essential
on Ubuntu.
Once you have make and a C++ compiler, you can compile our code by executing:
make
This command will generate multift, an executable of our model.
1.2 Obtain text data for training. We included scripts to download text8 and text9 in data/.
bash data/get_text8.sh
bash data/get_text9.sh
In our paper, we use the concatenation of ukWaC and WaCkypedia_EN as our English text corpus. Both datasets can be requested here.
The foreign language datasets deWac (German), itWac (Italian), and frWac (French) can be requested using the above link as well.
1.3 Run sample training scripts for text8 or text9.
bash exps/train_text8_multi.sh
After the training is complete, the following files will be saved:
modelname.words List of words in the dictionary
modelname.bin A binary file for the subword embedding model
modelname.in The subword embeddings
modelname.in2 The embeddings for the second Gaussian component.
modelname.subword The final representation of words in the dictionary. Note that the representation for words outside the dictionary can be computed using the provided python module based on the files *.in and *.in2.
2.1 The provided python module multift.py can be used to load the multisense FT object.
ft = multift.MultiFastText(basename="modelfiles/modelname", multi=True)
Note that the first time it loads the model can be quite slow. However, it saves the .npy files for later use which allows the loading to be much faster.
We can query for nearest neighbors give a word or evaluate the embeddings against word similarity datasets.
2.2 The script eval/eval_model_wordsim.py calculates the Spearman's correlation for multiple word similarity datasets given a model. We provide examples below.
python eval/eval_model_wordsim.py --modelname modelfiles/multi_text8_e10_d300_vs2e-4_lr1e-5_margin1 | tee log/eval_wordsim_text8.txt
python eval/eval_model_wordsim.py --modelname modelfiles/multi_text9_e10_d300_vs2e-4_lr1e-5_margin1 | tee log/eval_wordsim_text9.txt
Sample output of the text8. sub and sub2 correspond to the Spearman's correlation of the first and second Gaussian components. We can see that having two components with potentially disentangled meanings improve the Spearman's correlation for word similarity.
Below is the output for word similarity evaluation on text8.
Dataset sub sub2 sub-maxsim
0 SL 26.746543 10.859781 27.699946
1 WS 65.720245 34.064534 66.638705
2 WS-S 70.978888 36.393886 70.474031
3 WS-R 62.104036 34.294508 60.655725
4 MEN 63.523683 34.983555 68.341550
5 MC 57.521141 37.650201 66.266134
6 RG 57.971173 50.623423 58.956847
7 YP 33.104499 10.518915 38.473373
8 MT-287 65.407845 45.664255 70.200717
9 MT-771 52.972120 33.979479 56.735716
10 RW 36.954597 3.707983 33.639994
A sample script eval/eval_text9_model_nn.py show the nearest neighbors of words such as rock, star, and cell where we observe multiple meanings for each word.
python eval/eval_text9_model_nn.py | tee log/eval_text9_model_nn.txt
Nearest Neighbors for rock, cluster 0
Top highest similarity of rock cl 0
['rock,:0', ..., 'bedrock:0', 'rocky:0', 'rocks,:0', '[[rock:0', ...
Nearest Neighbors for rock, cluster 1
Top highest similarity of rock cl 1
['(band)]],:0', '(band)]]:0', 'songwriters:0', 'songwriter,:0', 'songwriter:0', ...
We provide a pre-trained English model in .tar.7z and .zip format.
- 7z option (15 GB download, long extraction time). The .tar.7z file contails .npy files of vectors and .words for the dictionary file.
wget https://s3.amazonaws.com/probabilistic-ft-multisense/multift-english/mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.tar.7z -P modelfiles/
cd modelfiles
7z x -so mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.tar.7z | tar xf - -C .
cd ..
- .zip option (30 GB, faster extraction time).
wget https://s3.amazonaws.com/probabilistic-ft-multisense/multift-english/mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.zip -P modelfiles/
cd modelfiles
unzip mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1.zip
cd ..
Evaluate the downloaded model using:
python eval/eval_model_wordsim.py --modelname modelfiles/mv-wacky_e10_d300_vs2e-4_lr1e-5_mar1 --multi 1 | tee log/eval_wordsim_multift300_eng.txt
Below is the expected output:
Dataset sub sub2 sub-maxsim
0 SL 37.338851 17.488524 39.605635
1 WS 65.360961 28.699307 76.114355
2 WS-S 71.203755 29.717879 80.114966
3 WS-R 62.165077 35.096523 75.345368
4 MEN 73.473605 30.225000 79.653409
5 MC 77.280821 35.603027 80.930131
6 RG 78.106458 21.979871 79.811171
7 YP 54.495191 13.808997 54.929330
8 MT-287 66.591830 26.882829 69.437580
9 MT-771 58.283664 36.931110 69.678762
10 RW 47.873384 -3.224643 49.358893
FastText (www.fasttext.cc) provide model files in two formats: .bin and .vec. Note that the model based on .bin can calculate the representation for any given word that might not be in the directionary. This has the advantage over using .vec files which contains pre-calculated vectors of words in the training dictionary.
We additionally provide a functionality to convert the .bin FastText objects to our format, which can then be loaded into a Python object using multift.py.
One can convert a .bin file into our format via:
./multift output-model -multi 0 modelfiles/downloaded_model.bin
The model in our format will be saved in modelfiles/downloaded_model.in, modelfiles/downloaded_model.out, etc. This model can be loaded with our Python script using:
ft = multift.MultiFastText(basename=modelfiles/downloaded_model, multi=False)
Note that the above two steps will generate extra .npy files which allow for much faster loading at subsequent times.
The following script downloads the Wiki English embeddings from www.fasttext.cc, converts it to the python-readable format, and evaluate it on word similarity datasets.
https://s3-us-west-1.amazonaws.com/fasttext-vectors/wiki.en.zip -P modelfiles/
cd modelfiles
unzip wiki.en.zip
cd ..
./multift output-model -multi 0 modelfiles/wiki.en.bin
python eval/eval_model_wordsim.py --modelname modelfiles/wiki.en --multi 0 | tee log/eval_wordsim_ft_wiki-eng.txt
Output:
Dataset modelfiles/wiki.en-sub
0 SL 38.033168
1 WS 73.880820
2 WS-S 78.111959
3 WS-R 68.201896
4 MEN 76.367311
5 MC 81.197154
6 RG 79.983827
7 YP 53.327914
8 MT-287 67.934178
9 MT-771 66.892286
10 RW 48.092870
Note: The C++ code is adapted from the FastText library (https://github.com/facebookresearch/fastText).