Fabrizio Gotti’s NLP Projects

 

←  Back to home page

 

I have the pleasure of tackling exciting real-world challenges in NLP, surrounded by the best researchers in academia. Here are some of these projects.

2024
Principal developer: Fabrizio Gotti

In a preliminary study, I explored the use of Large Language Models (LLMs) to support Retraite Québec’s customer relationship management employees in answering customer emails. The project involved implementing and comparing leading LLMs: OpenAI Assistants with File Search, Google Vertex Agents, and Microsoft’s Copilot. I also implemented bespoke solutions to investigate how to customize LLMs to the task, using OpenAI and Llama 3 (on-prems). The goal was to enhance the efficiency and accuracy of responses to customer inquiries by implementing retrieval augmented generation (RAG). By leveraging these advanced AI models, the study aimed to improve customer service interactions and reduce response times. The findings indicated potential benefits in using LLMs for this purpose, paving the way for further research and development in this area.

Implementation: Python scripts to implement retrieval-augmented generation of emails, notably in order to implement an artificial “service script”. APIs from OpenAI, Google, Microsoft (Copilot and Custom question answering). On-premises: Llama 3, Weaviate’s Verba.

2023–2024
Principal researchers: Jian-Yun Nie and François Michaud (Université de Sherbrooke)
A project helmed by Fondation Sunny

Socially Assistive Robots are considered potential solutions to support the independence and well-being of older adults. This project, initiated by the not-for-profit organization Fondation Sunny, aims at creating such a robot, capable of empathetic conversations with seniors. The robot’s hardware and software were designed and tested by the Interdisciplinary Institute for Technological Innovation (Université de Sherbrooke), and the RALI laboratory was tasked with implementing empathetic conversations with large language models, both on-premises (with Llama) and in the cloud (OpenAI). The project also involved the Research Centre on Aging (RCA) of the Université de Sherbrooke.

In the context of this project, I mainly worked on fine-tuning and evaluating empathetic large language models using Facebook’s Llama 2 and 3, and implemented an API to serve this model.

Photo credit: Interdisciplinary Institute for Technological Innovation, Université de Sherbrooke.

2023–2024
Principal researcher: David Décary-Hétu

The aim of this project is to identify unknown new psychoactive substances (NPS) that are offered for sale to Canadians. To achieve this goal, this project aims to find discussions and advertisements online that are related to NPS, both on the internet and on the dark web.

This project involves four main activities: 1) the development of algorithms to detect new psychoactive substances in online conversations; 2) the monitoring of marketplaces; 3) the monitoring of discussion forums and; 4) the detection of new psychoactive substances using machine learning algorithms.

Since 2020
Principal researcher: Jian-Yun Nie

Since 2020, my colleagues and I have been building CLIQ-ai, the Computational Linguistics in Québec consortium, bringing together experts from academia and industry in Quebec, whose collaboration advances natural language processing technologies. An initiative of IVADO, this research group counting 10 academics and 13 industry partners, will ultimately contribute to the advancement of NLP by investigating how knowledge engineering and reasoning can be used in NLP models when data is scarce. We will focus on three research axes: Information and Knowledge Extraction (e.g. Open Information Extraction), Knowledge modeling and reasoning and applications of this research, mainly in question answering and dialogue systems (“chatbots”). I act as a scientific advisor in NLP and help write the grant applications to fund this consortium.

Website of the consortium

“Development of a weather text generator”

August 2021, in collaboration with Environment Canada and the Centre de Recherches Mathématiques

Weather forecasts include many ways to express possible forecasts. However, the texts of the forecasts are very structured and very limited in their formulation. In addition, there is only one “good” standard way of reporting the weather considering a given set of concepts. In order to continue to provide quality weather forecast and information services to Canadians, the MSC wishes to develop a weather forecast text generator, in English and French, which uses meteorological concepts representing, in coded form, the weather forecast.

Our solution, developed during the week of the workshop, employed a deep learning sequence-to-sequence (seq2seq) model that translates meteorological data into text, focusing on a subset of the problem: the text for the temperature forecast. We achieved a very promising BLEU score of 76%. See our presentation here.

Website for the 2021 edition of the workshop

2020-2022, in collaboration with Polytechnique Montréal and IVADO
Leads: Nancy Laramée, Lévis Thériault, Lilia Jemai, Fabrizio Gotti

IVADO is a Québec-wide collaborative institute in the field of digital intelligence, dedicated to transforming new scientific discoveries into concrete applications and benefits for all of society. As part of their mission, they match researchers in artificial intelligence with industry projects and challenges.

With IVADO’s Lilia Jemai, we led a team of fantastically talented Polytechnique students to integrate NLP techniques in order to create a custom search engine capable of matching the academics’s research interests with the project descriptions. This “match-making” tool relied on bibliometric data retrieved from Microsoft Academic (now defunct, replaced with OpenAlex) and Google Scholar, for all researchers in Canada and elsewhere.

The chosen implementation, in React and Python, leverages the spaCy NLP framework and is deployed on IVADO’s Google Cloud Platform (Docker containers). Here is a screenshot of the application.

“Detection of recurring defects in civil aviation”

August 2021, in collaboration with Air Canada and the Centre de Recherches Mathématiques
Lead at Air Canada: Keith Dugas, Manager, Connected Operations

Transport Canada mandates per the Canadian Aviation Regulation (CAR 706.05 and STD 726.05) that an Air Operator Certificate (AOC) holder must include in its maintenance control system procedures for recording and rectification of defects, including the identification of recurring defects. Air Canada wished to detect recurring defects automatically that meets and exceeds Transport Canada requirements for both MEL and Non-MEL defects.

Extensive pre-processing was necessary because of the complexity of the data, including the (partial) resolution of the multiple acronyms. Text classification vastly outperformed direct clustering. The task and the data deserved much more work than possible during the workshop week, but the preliminary results are promising.

Website for the 2020 edition of the workshop

Website for our challenge

2017-2020, in collaboration with Érudit, the leading digital dissemination platform of HSS research in Canada
Leads: Philippe Langlais and Vincent Larivière

CO.SHS is dedicated to supporting research in the humanities and social sciences and the arts and letters in multiple ways. Financed by the Canada Foundation for Innovation as part of the Cyberinfrastructure initiative (read the funding announcement), the project is overseen by Vincent Larivière, associate professor of information science at the École de bibliothéconomie et des sciences de l’information, holder of the Canada Research Chair on the Transformations of Scholarly Communication and scientific director of Érudit.

The Allium prototype, developed by the RALI (Philippe Langlais and Fabrizio Gotti, Université de Montréal), aims to increase information discoverability in the digital library Érudit through the open information extraction (OIE) and end-to-end named entity recognition. It uses the semantics of the plain text of journal articles to explore the deep syntactic dependencies between the words of a sentence. The direct extraction of indexed facts allows to expand user experience and offers the possibility of browsing based on concepts, themes and named entities in context. The implementation of external links pointing to knowledge bases such as Wikipedia further enriches the browsing experience.

See the prototype in action on YouTube.

2014-2017, in collaboration with Little Big Job Networks Inc.

Lead: Guy Lapalme

The RALI at Université de Montréal and its industrial partner LittleBigJob Networks Inc. develop a platform to improve the process of recruiting managers and high level technicians by exploiting Big Data on many aspects: improvement in harvesting and combining public data from social networks; improvement in the matching process between candidates and job offers; prediction of the success of a given candidate at a new position. This project relies on Big Data analysis gathered by the industrial partner both on the web and from its business partners while complying with the privacy of the candidates and making sure that internal recruitment strategies of industry are not revealed.

2014-2016, in collaboration with the Observatoire de linguistique Sens-Texte, OLST

Lead: Marie-Claude L’Homme

EcoRessources is a comprehensive platform that brings together online dictionaries, glossaries and thesauri focusing on the environment. Subjects covered include climate change, sustainable development, renewable energy, threatened species, the agri-food industry.

A single query allows users to quickly identify all the resources that contain a specific term and to directly access those that can provide additional information. Languages covered are English, French and Spanish. Try it!

Implementation: Custom-made web site built on Silex micro-framework, Bootstrap et al. The index is built offline from a collection of databases provided by our collaborators in various formats, then converted to XML. XSLT is used to perform the conversions and the queries.

2014-2016, in collaboration with personnel from DIRO and OLST

Lead: Philippe Langlais

Humans can understand each other because they share a code, i.e. language, as well as common knowledge. Machines do not possess this collection of facts, a lack that significantly hinders their capacity to process and understand natural language (like French). Various projects have been trying to fill this gap, like the Common Sense Computing Initiative (MIT) or NELL (CMU), typically for the English language. The defacto project is interested in creating a French platform to acquire common sense knowledge. The main originality of this project will be the implementation of an interactive tutoring environment where the user will help the computer interpret French text.

Implementation: I have written a preliminary study (in French) interested in the collaborative creation of linguistic resources in so-called “serious games”. With Philippe Langlais, I modified ReVerb to parse French and extracted simple facts from French Wikipedia. Top-level concepts were extracted from this data and put in relation with instances in the corpus. To properly visualize the results, this knowledge was plotted on an interactive graph created with Gephi and Sigmajs Exporter (with modifications).

  See a screenshot of the preliminary results here.

We devised an entity classifier based on the relations these entities are involved in. For this, we extracted millions of facts from the Erudit corpus and from French Wikipedia. Our results have been published at the 2016 Canadian Artificial Intelligence conference.

2014-2015, in collaboration with Pegasus Research & Technologies

Lead: Guy Lapalme

Suspicious Activity Reporting refers to the process by which members of the law enforcement and public safety communities as well as members of the general population communicate potentially suspicious or unlawful incidents to the appropriate authorities. SAR has been identified as one part of a broader Information Sharing Environment (ISE). The ISE initiative builds upon the foundational work by the US Departments of Justice and Homeland Security that have collaborated to create the National Information Exchange Model (NIEM).

The approach of the current project is to introduce artificial Intelligence technologies: 1) to enhance human-machine interactions, to get information into the system more rapidly and also to make it more readily available to the users; and 2) for advanced machine processing to data validation, fusion and inference to be performed on data collected from multiple sources.

Implementation: Custom-made Java library using software from Apache Xerces™ Project to parse IEPD specifications (XML Schema) in order to guide the creation of an intelligent user interface. The IEPDs used are based on NIEM.

This video shows the XSDGuide prototype we created.

An article published in Balisage 2015 describes the prototype.

Lead: Guy Lapalme

Zodiac is a system that will automatically restore accents (diacritical marks) in a French text. Even though they are necessary in French to clearly convey meaning (“interne” vs. “interné”) and to spell words correctly, diacritics are quite frequently omitted in various situations. It may therefore be useful to reinsert diacritical marks in a French text where these marks are absent. We therefore coded Zodiac, a statistical natural language tool providing this feature. It is based on a statistical language model and on a large French lexicon.

Implementation: In C++, with special care to ensure portability across Mac, Windows and Linux platforms. We used the ICU library to process Unicode text. The language model used is a trigram model trained with the SRILM library. The training corpus consists of 1M sentences pertaining to the political and news fields. The model is compiled in binary form with the C++ library KenLM in order to be loaded in memory in 0.1s. I also wrote a word add-in and a Mac service for Zodiac.

Try it!

RALI studied the creation of linguistic resources using cooperative approaches (games with a purpose, microsourcing, etc.) at the behest of the FAS. A report explaining the various possible strategies was written and is available on RALI’s website (in French).

Automatic tweet translation [2013]

In collaboration with NLP Technologies,  we benefited from an engage grant from the NSERC to study tweet translation. We focused on their collection and their particularities, in order to create an automated translation engine dedicated to tweets emanating from Canadian governmental agencies. An article published at LASM 2013 explains the project and the results it yielded.

A web service published on our servers allowed our partners to closely follow the development of the prototype. A web interface (pictured below) offered a concrete demonstration platform of the prototype.

An engage grant awarded by the NSERC gave us the opportunity to collaborate with KeaText inc., a Montreal-based knowledge engineering firm. The project consisted in using the UIMA framework to create and interact with typed, semantic annotations in judicial corpora, for information extraction.

We wrote a Web application (see below) facilitating manual annotation, as well as a UIMA-based backend for storing and manipulating these annotations. These modules used UIMA in an innovative way and contributed to the adoption of these technologies by our commercial partner.

This research project will help researchers better understand the language used in text messages and how it helps build social networks. I worked on an automatic translator designed to translate SMS messages into proper French.

The project is described here and the results of our work is available here.

This Mitacs seed project explores new ways of customizing and translating the mass of daily information produced by Environment Canada. The project is explained here.

My efforts are focused on the statistical machine translation of weather alerts (English to French and French to English). A prototype of the system we are working on is available online.

We are in the process of overhauling the engine of TransSearch, our bilingual concordancer used by professional writers to consult large databases of past translations. This project is led in collaboration with our commercial partner, Terminotix Inc., which hosts the service.   

We will offer TransSearch users a new way to consult previous translations: instead of answering their queries simply by presenting them with pairs of sentences (one in the source language, the other in the target language), we want to highlight within these pairs the source and target words they are looking for.  This entails the computation of word alignment between the source and target language material.

A screenshot of the new TransSearch (click to enlarge):

We have worked with Montreal-based Druide Informatique Inc. to use artificial intelligence techniques to improve the precision of their popular French grammar checker Antidote. Our efforts were successful and integrated into the Antidote HD product. This is the topic of my master’s thesis.

In collaboration with the Université de Genève (LATL), we developed a topic-answering and summarizing system for the main task of DUC 2007. We chose to use an all-symbolic approach, based on FIPS, a multilingual syntactic parser. We used XML and XSLT to represent and manipulate FIPS’s parse trees. 

For a project funded by PRECARN, the RALI, NLP Technologies and Mrs. Diane Doray develop a technology for automated analysis of legal information in order to facilitate the information research in banks of judgments published by legal information providers. The system includes a machine translation algorithm for the French/English summaries and judgments.

The RALI helped create a new query engine for Quebec’s lexical corpora available on the website of the Secrétariat à la politique linguistique du Québec.

The RALI is working with Oralys Inc. to augment their IdeoVoice product with a speech-to-ideogram translation module to assist communication with hearing-impaired, dysphasic and autistic persons. Ultimately, it is our hope that this technology will bridge the language barrier using idea transference rather than word-for-word translation.

The Montreal-based company Lingua Technologies Inc., in partnership with the RALI and Transetix Global Solutions Inc. in Ottawa, started the R&D project 3GTM through funding by the Alliance Precarn-CRIM program at the beginning of 2005. The project aims at the development and the marketing of new computer-assisted translation software based on third-generation translation memories. This tool that we are helping to develop will be able to recycle previous translations intelligently.

My colleagues and I within the RALI are working to find solutions to the numerous scientific challenges of the project: statistical model creation, sub-sentential unit retrieval, integration of the translation context, etc. A French presentation of the project made for a seminar at the RALI is available here.

To test the many language tools and technologies at our disposal, and to improve them, the RALI also participates in the CESTA evaluation campaign. It proposes a series of evaluation campaigns of machine translation systems for different languages (the target language is French). This project, started in January 2003, is funded by the French Ministry of Research (Ministère de la Recherche français) under the Technolangue program.

The exploration of the vast and mysterious world of natural language processing also led me to research word alignment technologies and the statistical models that underlie them. Word-aligning Inuktitut and English proved particularly enlightening. Moreover, we are currently interested in exploiting new translation subsentential units called treelets, introduced by Quirk et al..