Thesis proposals

Web data extraction: the Smallest Extraction Problem

The Smallest Extraction Problem (SEP) consists in finding the grammar that best describes a set of pages generated from the same HTML template, so as to separate the template from the data and extract the data without supervision. Theses can build on an existing implementation of the inference algorithms. Possible directions:

• Incremental model: update the grammar as new pages arrive and as the template changes, recasting the model in probabilistic terms.
• Efficient inference: make the A* search practical on larger sets of pages, also through a probabilistic formulation of the model.
• From index pages to detail pages: among the links of index and section pages, recognize those that lead to content, follow them and infer the model of the pages they reach.
• Grammars and LLMs: let an LLM make the semantic choices that the structure of the pages alone cannot settle, such as deciding whether a repeated text is data or part of the template, or choosing among grammars of similar cost.

References: paper (PVLDB, 2021) · companion website

Seed selection for large-scale crawling (URLBank)

In collaboration with Meltwater.

URLBank chooses the pages (seeds) from which a crawler starts at each round to discover new articles: instead of a manually maintained list, seeds are selected by analyzing how the links on the visited pages change over time. Theses extend the system, for instance with depth-2 crawling from the index pages of each site, and work on real production data.

Reference: paper (The ACM Web Conference, 2026)

Data extraction from page screenshots with multimodal models

In collaboration with Meltwater.

Extracting data from Web pages starting from an image of the rendered page, rather than from its HTML code, by means of multimodal models. A first thesis showed that the visual approach greatly reduces token usage and recognizes titles and other prominent elements well, but struggles with long texts and with mapping the results back to the DOM. Possible extensions:

• using the multimodal model on a few pages only, to generate extraction rules (for instance, XPath expressions) that are then applied at scale to all the pages of the same template;
• recognizing whole blocks of the page (the article body, the comments, the menus) instead of single fields, to make mapping to the DOM easier;
• hybrid architectures that combine visual and textual information;
• domains other than news, such as e-commerce or real estate listings, with lists, tables and nested data.

The computer-based exam system for Object-Oriented Programming

The Object-Oriented Programming exams are taken on a web-based system developed in-house and in use since 2006. Besides multiple-choice and open-ended questions, the system asks students to write Java code, which is compiled and checked automatically with JUnit tests.

Technologies: Java, Spring MVC, JSP, Tomcat, PostgreSQL, JUnit, Maven, git.

The first step is porting the system to current technologies (recent versions of Java, Spring Boot, JUnit 5), with isolated execution of students' code, versioning of questions and clearer diagnostics for the instructor. On this basis, theses can develop one of the following extensions:

• support for acceptance tests, in addition to unit tests;
• measuring the speed-up of concurrent algorithms written by students;
• automatic generation of quizzes and assignments with LLMs;
• a Moodle plugin, to integrate the system with course pages;
• managing the course's HQ (Homework & Quiz) track within the system: homework submission, mid-term quizzes and score calculation.