GSoC-FrameNetBr

This repository contains code for evaluating the semantic similarity between a sentence given in source (natural) language and another sentence given in target language using FrameNet open source technology. In this project, we experiment with three languages precisely : English, Deutsche and Portuguese.

Pre-requisites :

1. Linux system
2. Anaconda, Python 3.x

Instructions :

1. Create an environment in your local server using the given environment.yml file with the following command :
   conda env create -f environment.yml
   The first line of the yml file sets the new environment's name.
2. Activate the environment using : conda activate <env_name>
3. In that environment, open jupyter notebook to access the given .ipynb files in this repo
4. Run the xml_parsing.ipynb script first.

Since the size of the binary files of the pretrained embeddings used in the next script create_feat_embeddings.ipynb are huge, storage of these resources in local machines and running this script can be a problem. Hence it is recommended to skip steps 5, 6 & 7. The outputs of steps 5 & 6 which will be required in the execution of some scripts later is already provided in the /resources folder of this repo.

5. The next script create_feat_embeddings.ipynb, uses FastText pretrained embeddings to create dictionaries of embeddings for the different features used to evaluate the semantic similarity. For the pretrained embeddings, download the .bin files for the required languages from this website.

6. The above script outputs .txt files which stores the out-of-vocabulary(oov) words for each respective language. To compute the embeddings for the oov words in each language,

• Install FastText in your environment folder using :

$ wget https://github.com/facebookresearch/fastText/archive/v0.2.0.zip
$ unzip v0.2.0.zip
$ cd fastText-0.2.0
$ make

• For installing and running FastText properly, ensure you have a (g++-4.7.2 or newer)

• Then run the following command using the pretrained embedding you had previously downloaded.

$ cat oov_words_en.txt | ./fasttext print-word-vectors cc.en.300.bin >> en_oov.bin

This will give a .bin file of the oov words.

7. The above script uses Google's Universal Sentence Encoder to obtain sentence embeddings. Although the dependencies should be installed by the requirements file, in case of any problem, please install Tensorflow v.1.12 explicitly and follow the instructions given in this link.

8. The script bert-embeddings.ipynb uses pretrained cased multilingual BERT embeddings to generate embeddings for lexical units and sentences. The pytorch interface for BERT by Hugging Face has been used to access and experiment with the pretrained embeddings. It contains a method which maps the lexical units to the actual words in the sentences and maps with an error of < 5%. It auto-generates a resource file called bert_embeddings.pickle which contains all the required embeddings, ready for use.

9. The script similarity.ipynb evaluates the semantic similarity on the basis of different features like FrameNet v.1.7 frames, etc. The already computed results and visualisations can be found inside the folder /resultsof this repo.

10. The script score_generator.ipynb finds the weighted average scores and brings together all the features of the model in a format suitable for the regression model.

11. Baseline models have been experimented with in the script baselines.ipynb and semi supervised models, in the script iterative-learning.ipynb .

Please run the scripts in the order they have been presented here. Have a look at TaskChecklist.md to have an idea of the workflow and the tasks that were/weren't accomplished during the period.

For a more detailed explanation of the project, please visit this blog post.