rbtsbg/nvc

Neural Vector Conceptualization

Neural Vector Conceptualization (NVC)

A new method for interpreting arbitrary word vector samples.

Accompanying code for the paper:

@inproceedings{Schwarzenberg_nvc_2019,
  title = {Neural Vector Conceptualization for Word Vector Space Interpretation},
  booktitle = {Proceedings of the NAACL-HLT 2019 Workshop on Evaluating Vector Space Representations for NLP (RepEval)},
  author = {Schwarzenberg, Robert and Raithel, Lisa and Harbecke, David},
  location = {Minneapolis, Minnesota, USA},
  year = {2019}
  }

Top: Neural Vector Conceptualization of listening. Bottom: Cosine Similarity.

Installation

Create and activate an environment with Python 3.6.

conda create --name NVC python=3.6
source activate NVC

Make sure, git-lfs is installed.

Note: when cloning the repository, 800 MB of data will be downloaded from the GitHub LFS server automatically (if git-lfs is installed).

If git lfs fails to download input_data.zip, please download the data from here.

After cloning the repository, install requirements.

pip install -r requirements.txt

Unzip the data:

unzip data/input_data.zip

Data

Word Vectors:

The current underlying word vectors were learned with the word2vec model (Mikolov et al., 2013). Please download the pre-trained word vectors to the data directory: https://code.google.com/archive/p/word2vec/ (GoogleNews-vectors-negative300.bin.gz.)

Microsoft Concept Graph

• For using the Microsoft Concept Graph data from scratch, please download the data here: https://concept.research.microsoft.com/Home/Download. This data dump does only comprise concepts, instances and associated counts, no probabilities or REP values. These need to be calculated before training. Therefore, see the script utils/ms_concept_graph_scoring.py in the notebook which calculates all needed probabilities and writes the data to a TSV file.

• The resulting file can be found in data/data-concept-instance-relations-with-rep.tsv. To get the REP values, follow the procedure as described in the notebook. The result will be a JSON file data/raw_data_dict.json.

• We recommend to use the preprocessed data data/raw_data_dict.json which includes all concepts and instances with their corresponding REP values in a JSON file.

Run and Replicate Experiments with the Notebook

The jupyter notebook demo_nvc.ipynb demonstrates how to use our (pre-)trained neural vector conceptualization (NVC) model to display the reported activation profiles.

Start the notebook:

jupyter notebook demo_nvc.ipynb 

or

EXPORT CUDA_VISIBLE_DEVICES=DEVICE_NUM jupyter notebook demo_nvc.ipynb

if you want to run it on a GPU.

You can run two versions of the notebook:

1. Use our pre-trained NVC model
2. Train a new model

Use the pre-trained model

If you run all cells of the notebook in the given order and without changing anything, our pre-trained NVC model is applied to a given filtered dataset and the results are reported at the end of the notebook.

Train a new model

1. Comment cell #4 and #5.
2. Uncomment cell #6:
   1. specify the data you want to use:
      1. either the same filtered data as above
      2. or a differently filtered data
   2. specify the file containing the word vectors
   3. specify the configuration file
3. Load the necessary modules: the embedding and model (cell #7)
4. In cell #9, load the data: nvc.load_data() is callable in three versions:
   1. nvc.load_data(path_to_data=path_to_filtered_data, filtered=True) use already filtered data.
   2. nvc.load_data(path_to_data=path_to_raw_data, filtered=False) use raw data and filter it according to the parameters set in the configuration.
   3. nvc.load_data(path_to_data=path_to_raw_data, filtered=False, selected_concepts=["city", "province"]) use raw data and filter it according to the parameters set in the configuration and according to a list of selected concepts.
5. Run nvc.train() (as in cell #14) to train a new model. The data split etc. is set in the configuration file.
6. The remaining cells display the activation profiles and report the results achieved by the model