PFE Book

2022
Contact:
team@icompass.digital
Please specify the project title with
your application

INDEX
About iCompass ………………………........……. Page 2
Licence Program ......…………………………. . Page 4
Engineering Program .....………………………. Page 8
Research Master’s Program …………………. Page 11
Engineering / Research Master’s…………. Page 12

1
About iCompass:

iCompass is a leading Artificial Intelligence star-

tup specialized in Natural Language Processing
products using the latest Deep Learning and
Reinforcement Learning technologies to develop
speech transcription services, automatic voice
generation and chatbot/voicebot products.

iCompass broke linguistic barriers in the Middle

East and Africa thanks to its newly developed
products that process Arabic dialects and African
languages, in addition to standard languages
such as English and French.

2
Our Vision:

We are committed to delivering the latest

R&D smart solutions, and providing a
deeper machine understanding of standard
languages and underrepresented dialects in
Africa and the Middle East.

Our Mission:

iCompass strives to provide a deeper machine

understanding of African and Arabic languages
and dialects bridging the gap digitally between
brands and their customers through automatic
Text to Speech, Speech to Text, and Voice to
Voice generation.

Our Values:

We value knowledge, integrity and diversity to

deliver innovative, anticipatory, and reliable
multilingual and multidialectal customer-centric
solutions.

3
Degree Licence Program (Bac+3)
Title Tunisian Dialect Identification from Text
Context This project is to alleviate the bottleneck in the
context of fine-grained Arabic dialect
identification.
Problem How to identify Tunisian dialect from other
Arabic dialects? The application is on noisy text
from social media.
Mission - State of the Art
- Dataset Annotation
- Dataset Cleaning
- Applying of Machine Learning Algorithms
- Discussing Results
Required - Fundamentals of Machine Learning
Expertise - Python Programming Language
Kick-o date January/February 2022

4
Degree Licence Program (Bac+3)
Title Tunisian Hate/Abusive Speech Detection
from Text

Context With the rise of social media, citizens of the

MEA region started using the internet as a
way to amplify their voices discussing diverse
and sensitive topics. In Tunisia, “Taboo" topics
have become popular and widely discussed
across social media platforms. However, on
the down side, it became easy to spread
abusive/hate propaganda against individuals
or groups.
Problem Compared to the increasing studies of AS/HS
detection in Indo-European languages, similar
research for Arabic dialects is still very limited
despite the existing need. This is mainly
attributed to the lack of publicly-available
AS/HS resources. We propose to solve this
problem using Machine Learning models.

Mission - State of the Art

- Dataset Annotation
- Dataset Cleaning
- Applying Machine Learning Algorithms
- Discussing Results
Required - Fundamentals of Machine Learning
Expertise - Python Programming Language
Kick-o date January/February 2022

5
Degree Licence Program (Bac+3)
Title Misogyny Detection in the Tunisian Dialect

Context According to DataReportal,The number of

social media users in Tunisia was equivalent
to 69.0% of the total population in January
2021.
This has led to a rise in the use of hate
speech including misogynistic behaviours.
Because of the increase of such interactions,
manual identification of misogyny speech
has become nearly impossible.
Problem Misogyny is a problem that has been obvious
since the earliest days of the internet but has
increased markedly in Tunisia since 2011. We
propose to solve this problem using Machine
Learning models.
Mission - State of the Art
- Dataset Annotation
- Dataset Cleaning
- Applying Machine Learning Algorithms
- Discussing Results
Required - Fundamentals of Machine Learning
Expertise - Python Programming Language
Kick-o date January/February 2022

6
Degree Licence Program (Bac+3)
Title Fake News Detection in the Tunisian Dialect

Context The spread of false and harmful information in

the Arab world on social media and formal
media (TVs, Radios, newspapers, etc.)
Problem Automatically detect fake news, propaganda,
misinformation and infodemic in Arabic
language and Arabic dialects without human
interference.
Mission - State of the Art
- Dataset Annotation
- Dataset Cleaning
- Applying Machine Learning Algorithms
- Discussing Results
Required - Fundamentals of Machine Learning
Expertise - Python Programming Language
Kick-o date January/February 2022

7
Degree Engineering Program
Title
Implementation of a Chatbot Dashboard

Context Conversation logs can be used to analyse user

and bot
behavior. Thus, a visualization tool could help
developers and/or users get better insights into
the chatbot performance.

Problem When people interact with chatbots, the size of

log files is growing continuously.
Therefore, the data that needs to be analyzed
and processed is growing.
Thus, a Business Intelligence system is crucial for
this task

Mission - Comparative study of different open source BI

frameworks.
- Understanding chatbot functionalities.
- Analyzing data and creating useful insights.
- Implementing an interface to provide
out-of-the�box visualizations, including
histograms, line charts, pie charts, and
more.

Required - Business Intelligence

Expertise - Data Analysis
- Data Visualization
- DataBase skills
- Good Programming Skills.

Kick-o date January/February 2022

8
Degree Engineering Program
Title Build a Dialog System using a Deep Learning
Framework

Context Chatbot applications are growing rapidly.

Moreover, Frameworks provide Missions for
developing well� structured maintainable and
upgradable AI assistants. They save a lot of
time by reusing generic modules and libraries
to focus on other areas of applications.

Problem Understanding the functionalities of such

frameworks and developing task-oriented
systems is a challenging project.

Mission
- Understand the Framework architecture and
components
- Collect and Pre-process data
- Implement APIs
- Design and Implement Scenarios and Tasks
- Build and evaluate machine learning models

Required - Deep Learning

Expertise - Natural Language Processing
- Conversational AI
- Problem Solving
- Python Programming Language

Kick-o date January/February 2022

9
Degree Engineering Program
Title Data Augmentation for NLP

Context In deep learning, the more data we have, the

better performance we can achieve. However,
it is expensive and time-consuming to annotate
a large amount of training data. Therefore,
proper data augmentation is useful to boost up
the model performance.

Problem In computer vision, images can be augmented

easily by flipping, shifting, shearing, etc.
However, in Natural Language Processing, it is
hard to augment text due to the high complexity
of languages.
Mission - State of the art
- Familiarize with deep learning models
- Implement Similarity Functions
- Implement a system to generate synthetic
data for improving model performance
without manual effort
Required - Deep Learning
Expertise - Natural Language Processing
- Language Models
- Python Programming Language

Kick-o date January/February 2022

10
Degree Research Masters Program (Bac+5)
Research - Tunisian Dialect Detection
Projects - Misogyny Detection in the Tunisian
Dialect
- Tunisian Hate/Abusive Speech Detection
- Fake News Detection
- Arabic Speech Translation
- Audio Reconstruction using Deep
Learning
- Tunisian Dialect Translation

Context On social media, Arabic speaking people

tend to express themselves in their own
local dialects. More particularly, Tunisians
use the informal way called "Tunisian
Arabizi". Analytical studies seek to explore
text contents aiming to exploit them for
planning and prediction purposes such as
measuring the customer satisfaction and
automatic detection of hate/abusive
speech, fake news communities, etc.
However, analytical studies based on Deep
Learning are data hungry and the Tunisian
dialect is considered a low resourced
language.
Mission - State of the Art
- Dataset Annotation
- Dataset Cleaning
- Applying Deep Learning Algorithms
- Fine-tuning with state of the art
Language Models
- Discussing Results
Required - Fundamentals of Machine/Deep
Expertise Learning Algorithms
- Python Programming Language
- Knowledge about Language Models
Kick-o date January/February 2022

11
Degree Engineering / Master's Program (Bac+5)
Title Automatic Speech Recognition of
Code-Switching Speech

Context Conventional approaches in automatic

recognition of code-switching employ an
LID system for detecting the languages
present within an utterance.
Problem he use of LID systems prior to various ASR
language� specific models increases the
inference time. Our goal is to build a
code-switching acoustic model.
Mission - Comparative study of different state of
the art models.
- Build code-switching datasets.
- Build and evaluate machine learning
models.
- Build and evaluate a code-switching
language model.
- Testing and evaluation.

Required - Machine/Deep Learning Algorithms.

Expertise - Problem Solving
- Research Methodology
- Natural Language Processing
- Python Programming Language
- Automatic Speech Recognition

Kick-o date January/February 2022

12
Degree Engineering / Master's Program (Bac+5)
Title Detection of Intonation - Sentiment Analysis
in Speech
Context Most of the work in Sentiment Analysis
revolved around textual data with text
mining and deep learning techniques.
Audio sentiment analysis is still in an early
stage of research.

Problem We will perform Sentiment Analysis

directly on a speaker's speech to detect
the emotions in a conversation.

Mission - State of the Art

- Analyze different deep learning algorithms
- Find the best algorithm to perform this task
- Test & Results
Required - Knowledge of Deep Learning frameworks
Expertise (Keras, Tensorflow or PyTorch) and libraries
(scikit�learn..)
- Ability to write robust code in Python
- Experience in Natural Language Processing
using deep neural networks (RNN/CNN) is a
plus
- Experience in working with GPU, Cuda/-
CuDNN is a plus
- Good communication skills

Kick-o date January/February 2022

13
Degree (BAC+5)
Engineering Program/ Master's Program
Title Audio Reconstruction using Deep Learning

Context Text To Speech

Problem One of the text to speech components is the
reconstruction of an audio from an amplitude
spectrogram.
Mission - Understand and explain a deep learning
architecture being developed in the
university of Waseda, Japan
- Implementation of the architecture using
PyTorch or TensorFlow
Required - Deep Learning
Expertise - Math Theory of Optimization
- Signal Processing
- Python Programming Language
Kick-o date January/February 2022

14
Contact:
team@icompass.digital