Table of Contents


This web site is an entry point for NLP research at IDSIA. The NLP group at IDSIA has been established in 2019.

Together with our host Institute (IDSIA), we share a joint affiliation with the University of Applied Sciences and Arts of Southern Switzerland (SUPSI) and the Università della Svizzera Italiana (USI).

The dual nature of IDSIA (basic research and technology transfer) allows us to perform cutting edge state-of-the-art research, and at the same time requires us to collaborate with local and national companies in order to bring these technologies into practical usage.

Our Research

We combine a profound understanding of the nature of natural language (human language) with expertise in the most recent techniques in the field of Natural Language Processing (NLP), in particular transformer-based architectures, such as BERT and GPT-3.

We apply our expertise to basic research and applied projects in collaboration with industry, in many cases funded by the Swiss Innovation Agency (InnoSuisse). See below some selected examples of recent projects.

A particular area of research interest is biomedical text processing, in particular on text analytics for different textual domains, such as the scientific literature, clinical reports, and social media.

We are also working on applications of NLP deep learning models in the financial domain, in collaboration with the Swiss banking industry.

Team Members

Software Engineering

Former members

Our Projects


  • People: Sandra, Vani, Denis, Fabio

In this project, we focus on conversational recommender systems that allow users to specify their preferences through a sequence of dynamically customized interactions, as contrasted to traditional ones. In particular, we seek to improve an online recommendation platform of Stagend ( that aims at finding the most suitable performer ("an item") for a particular event specified by an event organizer ("a user"). In the first phase, an adaptive, Bayesian methods-based approach was used to sequentially update the model given a new piece of information, e.g. performer's answer to organizer's question. However, in a real-time setting, delayed/incomplete interactions (e.g. missing reply), can hamper the system efficiency.

To overcome this issue, and also to avoid unnecessary burden on performer (in cases when the answer is already available in performer's biography or previous events' conversations), we investigate the ways of enhancing the Bayesian approach with NLP methods. Specifically, we adopt a question-answering BERT-based approach to either provide a confident automated answer based on the existing information, or to indicate uncertainty and thus, the necessity of contacting the performer. Additionally, given that Stagend operates in multilingual markets, we benchmark different multilingual models such as multilingual BERT and XLM-RoBERTa, as well as compare these with separate language models per each of the target languages (DE + Swiss DE challenge, FR, IT, EN).


  • People: Sandra, Denis, Oscar, Fabio

SkillGym ( is a computer-based training system that enables in-role and prospective leaders to develop their communication skills by presenting them with realistic simulations of workplace situations. SkillGym walks the end user through a sequence of videos related to a specific management situation by showing a rich set of alternatives as text boxes. SkillGym also provides extensive feedback, which enables users to review a conversation step by step, and learn the implications of their behavior at each step.

Feedback from SkillGym users praises its engaging training environment. To make simulations even more realistic, our goal is to move from the existing point-and-click interface to a voice-based interface. Achieving this goal requires cutting-edge natural language understanding to interpret the user input in the context of the ongoing flow of the simulated interaction. Our proposed solution is to carry out feature extraction based on the output of a commodity speech-to-text engine so that a dialog state tracker can select the next video based on the user input. Notably, the user must be guided through textual hints to ensure that she provides input that is coherent with the training goals of SkillGym. Moreover, the dialog state tracker must handle all situations where the user input is not aligned with the training goals (e.g. off-topic comments, disambiguation).


  • People: Claudio, Fabio

In a collaborative project with a major pharma company we explored name entity recognition (NER) strategies applied to job/resume mining tasks. In the project we leveraged advanced NER approaches in order to identify job titles, organization names, and geographical locations which are the essential parts of a job mining task, such as recruiting, tracking job candidates and job recommendation. This process is currently based on the manual analysis of hundreds of CVs, often with no relevance for a specific position or a profile.

Despite the existence of many commercial providers of similar services, there are no publicly available datasets to evaluate the advertised algorithms. The existing pre-trained NER models such as spaCy models, and Stanford NER models were trained on blogs, news and media. Their performance drops significantly when applied on the sentences taken from the resumes, since titles, locations and organization names in a resume are often written in the manner of a heading.

Our approach outperforms pre-trained models by a significant margin. Our NER models have been integrated in a prototype system which demonstrates a more dynamic and flexible data analysis compared to baseline commercial solutions.

Social Media Mining for health

  • People: Joseph, Fabio

Social media platforms offer extensive information about the development of the COVID-19 pandemic and the current state of public health. In recent years, the Natural Language Processing community has developed a variety of methods to extract health-related information from posts on social media platforms. In order for these techniques to be used by a broad public, they must be aggregated and presented in a user-friendly way. We have aggregated ten methods to analyze tweets related to the COVID-19 pandemic, and present interactive visualizations of the results on our online platform, the COVID-19 Twitter Monitor.

How to find us

idsia-logo.jpeg We are based at the Dalle Molle Institute for Artificial Intelligence (IDSIA), near Lugano, Switzerland.


Click here to find our location on a map

Galleria 2
6928 Manno (Lugano)


Dr. Fabio Rinaldi
E-Mail: fabio AT
Tel: +41 (0)79 300 67 71
Fax: +41 (0)58 666 66 61

Author: Fabio Rinaldi

Created: 2020-10-06 Tue 19:18