1. Melusine

_images/melusine.png https://img.shields.io/pypi/v/melusine.svg https://travis-ci.org/MAIF/melusine.svg?branch=master Documentation Status

1.1. Overview

Melusine is a high-level Python library for emails classification and feature extraction, written in Python and capable of running on top of Scikit-Learn, Keras or Tensorflow. It was developed with a focus on emails written in french.

Use Melusine if you need a library which :
  • Supports both convolutional networks and recurrent networks, as well as combinations of the two.

  • Runs seamlessly on CPU and GPU.

Melusine is compatible with Python 3.6 (<=2.3.2), Python 3.7 and Python 3.8.

1.2. Release Notes

1.2.1. Release 2.3.5

Bug fix:
  • PR139: Fix config search with a recursive paramater.

1.2.2. Release 2.3.4

New features:
Bug fix:
  • PR 124: fixing purge of dict_attr keys while saving bert models (train.py)

  • Issue 126: fixing initialisation of bert_tokenizer for cross validation (train.py)

1.2.3. Release 2.3.2

  • Compatibility with python 3.7 and 3.8

  • Optional dependencies (viz, transformers, all)

  • Specify custom configurations with environment variable MELUSINE_CONFIG_DIR

  • Use any number of JSON and YAML files for configurations

(instead of just one config file)

Bug fix:
  • Fixed bug when training transformers model without meta features

1.2.4. Release 2.3

New features:
  • Added a class ExchangeConnector to interact with an Exchange Mailbox

  • Added new tutorial tutorial14_exchange_connector to demonstrate the usage of the ExchangeConnector class


1.2.5. Release 2.0

New features:
  • Attentive Neural Networks are now available. :tada: We propose you an original Transformer architecture as well as pre-trained BERT models (Camembert and Flaubert)

  • Tutorial 13 will explain you how to get started with these models and attempt to compare them.

  • Validation data can now be used to train models (See fit function from NeuralModel for usage)

  • The activation function can now be modified to adapt to your needs (See NeuralModel init for usage)

1.2.6. Release 1.10.0

  • Melusine is now running with Tensorflow 2.2

1.2.7. Release 1.9.6

New features:
  • Flashtext library is now used to flag names instead of regex. It allows a faster computation.

1.2.8. Release 1.9.5

New features:
  • An Ethics Guide is now available to evaluate AI projects, with guidelines and questionnaire. The questionnaire is based on criteria derived in particular from the work of the European Commission and grouped by categories.

  • Melusine also offers an easy and nice dashboard app with StreamLit. The App contains exploratory dashboard on the email dataset and a more specific study on discrimination between the dataset and a neural model classification.

1.3. The Melusine package


This package is designed for the preprocessing, classification and automatic summarization of emails written in french. 3 main subpackages are offered :

  • prepare_email : to preprocess and clean the emails.

  • summarizer : to extract keywords from an email.

  • models : to classify e-mails according to categories pre-defined defined by the user.

2 other subpackages are offered as building blocks :

  • nlp_tools : to provide classic NLP tools such as tokenizer, phraser and embeddings.

  • utils : to provide a TransformerScheduler class to build your own transformer and integrate it into a scikit-learn Pipeline.

An other subpackage is also provided to manage, modify or add parameters such as : regular expressions, keywords, stopwords, etc.

  • config : contains ConfigJsonReader class to setup and handle a conf.json file. This JSON file is the core of this package since it’s used by different submodules to preprocess the data.

1.4. Getting started : 30 seconds to Melusine

1.4.1. Importing Melusine

To use Melusine in a project:

import melusine

1.4.2. Input data : Email DataFrame

The basic requirement to use Melusine is to have an input e-mail DataFrame with the following columns:

  • body : Body of an email (single message or conversation history)

  • header : Header of an email

  • date : Reception date of an email

  • from : Email address of the sender

  • to (optional): Email address of the recipient

  • label (optional): Label of the email for a classification task (examples: Business, Spam, Finance or Family)







Thank you.\nBye,\nJohn

Re: Your order

jeudi 24 mai 2018 11 h 49 CEST




To import the test dataset:

from melusine.data.data_loader import load_email_data

df_email = load_email_data()

1.4.3. Pre-processing pipeline

A working pre-processing pipeline is given below:

from sklearn.pipeline import Pipeline
from melusine.utils.transformer_scheduler import TransformerScheduler
from melusine.prepare_email.manage_transfer_reply import check_mail_begin_by_transfer, update_info_for_transfer_mail, add_boolean_answer, add_boolean_transfer
from melusine.prepare_email.build_historic import build_historic
from melusine.prepare_email.mail_segmenting import structure_email
from melusine.prepare_email.body_header_extraction import extract_last_body
from melusine.prepare_email.cleaning import clean_body

ManageTransferReply = TransformerScheduler(
    (check_mail_begin_by_transfer, None, ['is_begin_by_transfer']),
    (update_info_for_transfer_mail, None, None),
    (add_boolean_answer, None, ['is_answer']),
    (add_boolean_transfer, None, ['is_transfer'])

EmailSegmenting = TransformerScheduler(
    (build_historic, None, ['structured_historic']),
    (structure_email, None, ['structured_body'])

Cleaning = TransformerScheduler(
    (extract_last_body, None, ['last_body']),
    (clean_body, None, ['clean_body'])

prepare_data_pipeline = Pipeline([
  ('ManageTransferReply', ManageTransferReply),
  ('EmailSegmenting', EmailSegmenting),
  ('Cleaning', Cleaning),

df_email = prepare_data_pipeline.fit_transform(df_email)

1.4.4. Phraser and Tokenizer pipeline

A pipeline to train and apply the phraser end tokenizer is given below:

from melusine.nlp_tools.phraser import Phraser
from melusine.nlp_tools.tokenizer import Tokenizer

tokenizer = Tokenizer (input_column='clean_body', output_column="tokens")
df_email = tokenizer.fit_transform(df_email)

phraser = Phraser(
_ = phraser.fit(df_email)
df_email = phraser.transform(df_email)

1.4.5. Embeddings training

An example of embedding training is given below:

from melusine.nlp_tools.embedding import Embedding

embedding = Embedding(

1.4.6. Metadata pipeline

A pipeline to prepare the metadata is given below:

from melusine.prepare_email.metadata_engineering import MetaExtension, MetaDate, Dummifier

metadata_pipeline = Pipeline([
  ('MetaExtension', MetaExtension()),
  ('MetaDate', MetaDate()),
  ('Dummifier', Dummifier())

df_meta = metadata_pipeline.fit_transform(df_email)

1.4.7. Keywords extraction

An example of keywords extraction is given below:

from melusine.summarizer.keywords_generator import KeywordsGenerator

keywords_generator = KeywordsGenerator()

df_email = keywords_generator.fit_transform(df_email)

1.4.8. Classification

The package includes multiple neural network architectures including CNN, RNN, Attentive and pre-trained BERT Networks. An example of classification is given below:

from sklearn.preprocessing import LabelEncoder
from melusine.models.neural_architectures import cnn_model
from melusine.models.train import NeuralModel

X = df_email.drop(['label'], axis=1)
y = df_email.label

le = LabelEncoder()
y = le.fit_transform(y)

pretrained_embedding = embedding

nn_model = NeuralModel(architecture_function=cnn_model,

nn_model.fit(X, y)
y_res = nn_model.predict(X)

1.5. Glossary

1.5.1. Pandas dataframes columns

Because Melusine manipulates pandas dataframes, the naming of the columns is imposed. Here is a basic glossary to provide an understanding of each columns manipulated. Initial columns of the dataframe:

  • body : the body of the email. It can be composed of a unique message, a history of messages, a transfer of messages or a combination of history and transfers.

  • header : the subject of the email.

  • date : the date the email has been sent. It corresponds to the date of the last message of the email has been written.

  • from : the email address of the author of the last message of the email.

  • to : the email address of the recipient of the last message.

Columns added by Melusine:

  • is_begin_by_transfer : boolean, indicates if the email is a direct transfer. In that case it is recommended to update the value of the initial columns with the informations of the message transferred.

  • is_answer : boolean, indicates if the email contains a history of messages

  • is_transfer : boolean, indicates if the email is a transfer (in that case it does not have to be a direct transfer).

  • structured_historic : list of dictionaries, each dictionary corresponds to a message of the email. The first dictionary corresponds to the last message (the one that has been written) while the last dictionary corresponds to the first message of the history. Each dictionary has two keys :

    • meta : to access the metadata of the message as a string.

    • text : to access the message itself as a string.

  • structured_body : list of dictionaries, each dictionary corresponds to a message of the email. The first dictionary corresponds to the last message (the one that has been written) while the last dictionary corresponds to the first message of the history. Each dictionary has two keys :

    • meta : to access the metadata of the message as a dictionary. The dictionary has three keys:

      • date : the date of the message.

      • from : the email address of the author of the message.

      • to : the email address of the recipient of the message.

    • text : to access the message itself as a dictionary. The dictionary has two keys:

      • header : the subject of the message.

      • structured_text : the different parts of the message segmented and tagged as a list of dictionaries. Each dictionary has two keys:

        • part : to access the part of the message as a string.

        • tags : to access the tag of the part of the message.

  • last_body : string, corresponds to the part of the last message of the email that has been tagged as “BODY”.

  • clean_body : string, corresponds a cleaned last_body.

  • clean_header : string, corresponds to a cleaned header.

  • clean_text : string, concatenation of clean_header and clean_body.

  • tokens : list of strings, corresponds to a tokenized column, by default clean_text.

  • keywords : list of strings, corresponds to the keywords of extracted from the tokens column.

  • stemmed_tokens : list of strings, corresponds to a stemmed column, by default stemmed_tokens.

  • lemma_spacy_sm : string, corresponds to a lemmatized column.

  • lemma_lefff : string, corresponds to a lemmatized column.

1.5.2. Tags

Each messages of an email are segmented the in the structured_body columns and each parts are assigned a tag:

  • “RE/TR” : any metadata such as date, from, to etc.

  • “DISCLAIMER” : any disclaimer such as “L’émetteur décline toute responsabilité…”.

  • “GREETINGS” : any greetings such as “Salutations”.

  • “PJ” : any indication of an attached document such as “See attached file…”.

  • “FOOTER” : any footer such as “Provenance : Courrier pour Windows”.

  • “HELLO” : any salutations such as “Bonjour,”.

  • “THANKS” : any thanks such as “Avec mes remerciements”

  • “BODY” : the core of the the message which contains the valuable information.

1.6. Motivation & history

1.6.1. Origin of the project

MAIF, being one of the leading mutual insurance companies in France, receives daily a large volume of emails from its clients and is under pressure to reply to their requests as efficiently as possible. As such an efficient routing system is of the upmost importance to assign each emails to its right entity. However the previously outdated routing system put the company under ever increasing difficulties to fulfill its pledge. In order to face up to this challenge, MAIF has implemented a new routing system based on state-of-the-art NLP and Deep Learning techniques that would classify each email under the right label according to its content and extract the relevant information to help the MAIF counsellors processing the emails.

1.6.2. Ambitions of the project

Melusine is the first Open Source and free-of-use solution dedicated specifically to the qualification of e-mails written in french. The ambition of this Python package is to become a reference, but also to live in the French NLP community by federating users and contributors. Initially developed to answer the problem of routing e-mails received by the MAIF, the solution was implemented using state-of-the-art techniques in Deep Learning and NLP. Melusine can be interfaced with Scikit-Learn: it offers the user the possibility to train his own classification and automatic summarization model according to the constraints of his problem.

1.6.3. Why Melusine ?

Following MAIF’s tradition to name its open source packages after deities, it was chosen to release this package under the name of Melusine as an homage to a legend from the local folklore in the Poitou region in France where MAIF is historically based.

1.7. Credits

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.