The Atticus Project was recently announced as an initiative to, among other things, build a world-class corpus of labelled legal contracts which could be used to train and/or benchmark text classifiers and question-answering NLP models. Their initial release contains 200 labelled contracts. I wanted to experiment with the data set and build a working classifier that I could use on contract data, so I set out to build a simple project to load the dataset, convert it into a format that Spacy can read, and then train some classifiers to see how the data set performs. This repository contains the code I used to train classifiers based on 1) Word2Vec embeddings and 2) a BERT-based transformer model.
If you are in a hurry to test out the classifiers and are not really interested in how they were trained, you can currently install the classifier directly from a package I'm hosting on my AWS bucket by typing:
pip install https://jsv4public.s3.amazonaws.com/en_atticus_classifier_bert-0.1.0.tar.gz
This should install Spacy, Spacy-transformers, a BERT model and the classifiers. Once you've installed the model, you can use it like this:
import spacy
nlp = spacy.load('en_atticus_classifier_bert')
clause = """The Joint Venturers shall maintain adequate books
and records to be kept of all the Joint Venture activities and affairs
conducted pursuant to the terms of this Agreement. All direct costs and
expenses, which shall include any insurance costs in connection with the
distribution of the Products or operations of the Joint Venture, or if the
business of the Joint Venture requires additional office facilities than
those now presently maintained by each Joint Venturer"""
cats = nlp(clause).cats
cats = [label for label in cats if cats[label] > .7] # If you want to filter by similarity scores > .7
print(cats) # Show the categories
As discussed below, the performance of the model is good enough to be interesting, but currently not good enough to really be production ready. I think this is primarily due to the dataset being relatively small and many clause categories having fewer than 20 examples. I wanted to release this as-is, however so others could experiment. As the Atticus Project corpus grows, these classifiers should get better. In my experience 50 - 100 examples is typically a good target to aim for, so doubling or tripling the Atticus Corpus will hopefully lead to much, much better performance.
I first experimented with using Spacy's OOTB Word2Vec models. This approach was very quick to train, but the performance was not very good. The f-score was about .6. I also tried using a different set of word embeddings released as "Law2Vec", and these improved performance marginally to an F-Score of ~.64. I've included the code to train these models in Word2VecModelBuilder.py. You can simply run that python script. The default settings will load Spacy's en_core_web_lg model and embeddings. You can also load the Law2Vec model if you download the vector file:
wget -O ~/Downloads/Law2Vec.200d.txt https://archive.org/download/Law2Vec/Law2Vec.200d.txt
Then you can use Spacy to convert this file into a Spacy-compatible model like so:
mkdir /models python -m spacy init-model en /models/Law2VecModel --vectors-loc ~/Downloads/Law2Vec.200d.txt
Then you can change the model argument (per the example above) to '/models/Law2VecModel'. You probably want to change the output_dir too. Once you've trained a new model, you can load the trained model with spacy.load(output_dir).
- Overview
- The transformer models encode a lot more contextual information about words than Word2Vec models, so I wanted to see if I could squeeze more performance out of the dataset using BERT. The good news was performance increased substantially using a BERT-based model. This is still probably not good enough for use in production, but it's good enough to yield some interesting insights, particularly if you set your similarity threshold very high.
- Training Results
Using a BERT-based model, the beta release of the Atticus training set yields an acceptable (but still not really production-ready) F-score of .735:
LOSS P R F 1.093 0.739 0.472 0.576 1.960 0.763 0.566 0.649 0.290 0.756 0.661 0.706 0.985 0.764 0.683 0.721 1.616 0.770 0.681 0.723 0.517 0.743 0.673 0.706 1.044 0.754 0.697 0.724 0.127 0.762 0.728 0.745 0.542 0.748 0.722 0.735 0.946 0.756 0.722 0.739 0.219 0.751 0.720 0.735 0.551 0.751 0.720 0.735
Training the BERT-based model takes a lot more computing power, and a CUDA-compatible graphics card is absolutely recommended. Using a Nvidia 1050 Ti, the above training took about three hours.
- Step 1 - Sign Up for Atticus Project Data and Download
- I've included the Atticus CSV in the repository for convenience, but you should go to the Atticus Project website and signup there. For one, they would like to collect user and contact info for people downloading their dataset. For another, you should go there to make sure you get the latest version of their dataset.
- Step 2 - Install Python Dependencies and SPACY BERT Model
First, install Python dependencies (I'm using LexNLP to tokenize test data, you do not need it to build the model):
pip install lexnlp spacy pip install spacy-transformers==0.5.2 pandas
Then, download the BERT transformer model:
!python -m spacy download en_trf_bertbaseuncased_lg
- Step 3 - Load Atticus Data and Format for Spacy
The Atticus dataset is a csv, so we can use Pandas to load and manipulate it. Since we're training classifiers and not answering questions, we only care about the columns containing text for a given classification. The columns with headers marked "...-Answer" are meant for question-answering and we don't want to train on this data. We also don't really want the filename column or the document title columns, which are the first and second columns respectively. The following function will load our Atticus CSV, filter out the ...-Answer cols, the filename col and the document title col. Then, it will format the data into Spacy's preferred training format and split the training set into two pieces - a training set and an evaluation set. The default is to split the total data set so 80% is used for training and 20% is used for evaluation.
Code:
def load_atticus_data(filepath='/tmp/aok_beta/Final Publication/master_clauses.csv'): """ Load data from the atticus csv (omitting the answer cols as we want to train classifiers not question answering). Data is returned in the Spacy training format: TRAIN_DATA = [ ("text1", {"cats": {"POSITIVE": 1.0, "NEGATIVE": 0.0}}) ] A list of headers is also returned so you can add these labels. FYI, the Filename and Doc name columns are dropped as well. """ # Load csv atticus_clauses_df = pd.read_csv(filepath) # Do a little post-processing data_headers = [h for h in list(atticus_clauses_df.columns) if not "Answer" in h] data_headers.pop(0) # Drop filename col (index 0 for col 1) data_headers.pop(0) # Drop doc name (orig col 2 (index 1) but now first col (index 0)) training_values = {i: 0 for i in data_headers} atticus_clauses_data_df = atticus_clauses_df.loc[:, data_headers] train_data = [] # Iterate over csv to build training data dict for header in atticus_clauses_data_df.columns: for row in atticus_clauses_data_df[[header]].iterrows(): value = row[1][header] if not pd.isnull(value): train_data.append((value, {'cats': {**training_values, header: 1}})) return train_data, data_headers def create_training_set(train_data=[{}], limit=0, split=0.8): """Load data from the Atticus dataset, splitting off a held-out set.""" random.shuffle(train_data) train_data = train_data[-limit:] texts, labels = zip(*train_data) split = int(len(train_data) * split) # Return data in format that matches example here: # https://github.com/explosion/spaCy/blob/master/examples/training/train_textcat.py return (texts[:split], labels[:split]), (texts[split:], labels[split:])- Step 4 - Build the Model
WARNING - running the training takes a looong time, even if you have a CUDA-compatible graphics card and it's properly configured in your environment
You can just run the BertModelBuilder.py with default settings. On my Nvidia 1050 Ti, it took about 3 - 4 hours to run the training. Unless you're adding additional data, I'd suggest you just use my pre-built models.
You can follow Spacy's excellent instructions here to package up the final model into a tar that can be installed with pip like this:
pip install local_path_to_tar.tar.gz
I've uploaded the package to my public AWS bucket, and you can install directly from there like so:
pip install https://jsv4public.s3.amazonaws.com/en_atticus_classifier_bert-0.1.0.tar.gz
Now you can load it just like this:
nlp = spacy.load('en_atticus_classifier_bert')
I plan to also upload this to PyPi as well so you can just do something like this:
pip install atticus_classifiers_spacy (DOESN'T WORK YET)
Another option, is you can load the pickled model in the pre-trained folder:
import pickle
import spacy
nlp = pickle.load(open("/path/to/BertClassifier.pickle", "rb"))
# Then you can use the spacy object just like normal:
clause = "Test clause"
cats = nlp(clause).cats
cats = [label for label in cats if cats[label] > .7] #If you want to look only at labels with similarity scores over .7
print(cats)