Work with us
What are you looking for?
Master thesis
- Who can apply
- Master students of the University of Bologna. Most come from the Master’s degree in Artificial Intelligence, but other degrees are welcome.
- What it is
- A master thesis is a research activity on one of our research proposals, supervised by the lab members that the proposal names.
Project work
- Who can apply
- Students of the University of Bologna enrolled in the Master’s degree in Artificial Intelligence.
- What it is
- A project work is a 3 CFU activity. You can choose one of our research proposals or propose a topic of your own.
Visiting student or researcher
- Who can apply
- Master and PhD students enrolled at another institution, and researchers affiliated with another institution.
- What it is
- A student visit typically lasts between three and six months.
Master or PhD programme
- Who can apply
- Students who intend to enroll in a master or PhD programme at the University of Bologna and to work with the lab.
- What it is
- Admission follows the official procedures of the University of Bologna.
Project partnership
- Who can apply
- Companies, public institutions, and research groups that want to start a research project with us.
- What it is
- We take part in national and international research projects.
Choose a proposal
Open a proposal to read its description and supervisors, then select Choose this proposal.
Write to us
Tell us how many exams you still have to pass and when you expect to finish them, when you can start, and the degree session in which you intend to graduate. We reserve at least six months for a thesis, so the session must be at least six months after your start, and not before the end of your exams. You may start before your last exam.
For inspiration
Work done with students.
Read the course pages
The course page and the Virtuale page describe the project work and how it is assessed.
Choose a topic
Select Choose this proposal on one of our research proposals, or write a topic of your own in the next step.
Write to us
Describe the topic and your intended approach.
For inspiration
Work done with students.
Find a research topic
Our research proposals show the topics we currently work on.
Write to us
Agree with us on the visit and its research topic. Tell us how the visit is funded: by your home institution, or by the lab, and in that case how much funding you need. Researchers attach a curriculum vitae, a research statement, and a proposal for the research activity during the visit.
Complete the formalities
Students complete the required formalities with the administrations of their home institution and of the University of Bologna.
See our recent projects
The Projects page lists all of them.
Write to us
Describe your organisation and your idea.
Contacts
Context
Hate speech often relies on implicit content: meaning that a text suggests without stating it. It also relies on subtle reasoning nuances. Argument mining is the problem of automatically detecting and extracting arguments from text. An argument usually combines a premise and a claim. The premise is a fact, and it supports the claim, a subjective conclusion.
Objective
This proposal applies argumentative reasoning to hate speech, to make its implicit content explicit. The resulting hate speech detection systems should be more interpretable and more user-friendly.
Contacts
Context
Argument mining automatically detects and extracts arguments from text. An argument usually combines a premise and a claim. The premise is a fact, and it supports the claim, a subjective conclusion. Spoken arguments also carry speech information, such as prosody, which is the rhythm and intonation of speech. A spoken argument comes as recorded speech together with its transcript, the written words of that speech.
Objective
This proposal aims to detect arguments with speech information in addition to text. Speech information enhances the set of features used to find premises and claims.
Contacts
Context
Selective rationalization explains the prediction of a classification model with highlights of the input text. These highlights, called rationales, are a subset of the input that a user can read. A rationale should also describe faithfully how the model reaches its prediction. A rationale is a local explanation: it explains the prediction for one example.
Objective
Examples that belong to the same class might share similar rationales. This proposal aims to go from local explanations to a global explanation, such as a knowledge base, by aggregating and summarizing extracted rationales.
Contacts
Context
Selective rationalization trains a model to explain each prediction with highlights of its input text. These highlights, called rationales, are a part of the input that a user can read and that faithfully shows how the model classifies. In the Select-then-Predict Pipeline (SPP), a generator selects the rationale and a predictor classifies it. Training an SPP can get stuck in poor solutions, called local minima, because of interlocking: a suboptimal interplay between generator and predictor.
Objective
A mixture of experts (MoE) trains several models, called experts, on the same data, and each expert specializes in a subset of it. This proposal studies whether an MoE model for selective rationalization can address interlocking.
Contacts
Context
Selective rationalization explains a prediction with a rationale: a short part of the input text that a user can read and that faithfully shows what the prediction is based on. The Select-then-Predict Pipeline (SPP) builds this into the model: a generator selects the rationale, and a predictor classifies the rationale only.
Objective
Large language models (LLMs) can select a rationale and predict a label in a single answer when a prompt asks them to. This proposal evaluates how well they perform selective rationalization through prompting, compared with traditional SPP models.
Contacts
Context
Selective rationalization explains the prediction of a classification model with highlights of the input text, called rationales. A rationale should be readable by a user and should faithfully describe how the model reaches its prediction. Several techniques transform text into structured content, such as parse trees and abstract meaning representation graphs.
Objective
This proposal applies rationalization to structured content and constrains the structures that a rationale can take. These structural constraints depend on the application scenario and define which types of structures the rationalization system can extract. With tree kernels, functions that measure the similarity between two trees, the allowed structures are different types of trees.
Contacts
Context
Legal documents are among the domains that would benefit most from natural language processing (NLP), the field that builds tools to analyze text automatically. Tasks in this domain usually require a specialist with background knowledge, and they are hard to transfer to an automatic tool. Argument mining is the task of detecting and extracting arguments from text automatically. An argument usually combines a premise and a claim. The premise is a fact, and it supports the claim, a subjective conclusion.
Objective
Arguments in legal documents are usually well-structured and follow specific domain rules. This proposal applies the most recent NLP techniques, and hybrid techniques that can use domain knowledge, to argument mining on legal documents.
Contacts
Context
Terms of Service and Privacy Policies are the documents that people accept when they use an online service. Some of their clauses can be unfair to the people who accept them. For several years, the lab has built tools that detect unfair clauses in such documents written in English, in the CLAUDETTE and PRIMA projects listed on the Projects page.
Objective
This proposal applies new methods to detect unfair clauses and to classify them. The current focus is on large language models (LLMs), and alternative techniques are also of interest.
Contacts
Context
What counts as abusive or hate speech depends on the socio-cultural context: the culture, norms, and laws of the people who read a text. The same text might be deemed offensive in one culture, allowed in another, and legally prosecutable in a third. A detection model is a machine learning model that flags a text as abusive or hate speech. It learns from labeled examples.
Objective
This proposal evaluates how different definitions of abusive and hate speech affect detection models. The aim is to promote awareness when developing accurate detection models.
Contacts
Context
The standard way to train a machine learning model is to give it an annotated dataset: data paired with class labels. Annotators build this dataset after training on annotation guidelines, the rules that define how to assign the classes. They label the data with classes from a given class set. The model learns from the labels alone, without the guidelines, so it can learn any mapping from data to classes that fits the given data.
Objective
This proposal gives the annotation guidelines directly to the model. During training, the model receives the guidelines and unlabeled data, but no class labels.