The classical emphasis posed by natural language processing (NLP) technologies on linguistic representation plays a crucial role in big data scenarios, as semantic descriptions are not only useful for triggering novel inferences and data-driven workflows, but mostly as they are crucial for data interperability across heterogeneous sources.
The promises of deep neural networks are opening a wide spectrum of applications also in NLP. Controversial issues emerge with respect to the readability of representations and models so that concerns about autonomy and ethics have been already raised.
In this talk, research on mathematical models of linguistic structures, the so called lexical, tree and semantic kernels, and their role in training neural networks and kernel machines will be discussed. On the one side, they provide sound and robust mechanisms for feature engineering in complex learning tasks (such as Question Answering, ranking or entity linking tasks). Their mimicking cognitively plausible structures provides semantically meaningful similarity metrics and transparent representations useful for learning over complex data streams. On the other side, they give rise to modular representations, as combining two or more kernels gives still rise to kernel functions: in this way it is possible to account for different semantic dimensions (e.g. lexical, grammatical or encyclopedic) during the training for an individual task.
In the talk, recently proposed kernelized deep learning architecture, based on linearization methods, called Nystrom method, will be presented. Their beneficial impact in the training of deep networks within complex kernel spaces, as extensively tested across different semantic tasks, will be discussed.
Roberto Basili is Associate Professor at the Faculty of Engineering, University of Roma, Tor Vergata, since May 2003, where he currently teaches in courses such as “Data Mining, “Web Mining and Retrieval”, “Database Systems” for the Computer Science and Computer Engineering curriculum. He is co-editor in chief of the Italian Journal of Computational Linguistics in coordination with Simonetta Montemagni (ILC, CNR, Pisa), Member of the Board of Trustees of the Italian Association for Artificial Intelligence (AI*IA), since 2005, co-Founder and Member of the Board of Trustees of the Italian Association for Computational Linguistics (AILC), since September 2015. His research focused since 90’s on Artificial Intelligence methodologies and technologies, in the area of Machine Learning, and Natural Language Processing as well as on the engineering of Distributed and Web-oriented Natural Language Processing and Information Retrieval Systems. He studied mathematical models of machine learning, as explanations and core automation models for the development of linguistic capabilities in intelligent agents.
In this area, he has contributed with the definition of several algorithmic techniques for the optimization of semantic inference tasks (such as classification and pattern recognition from texts and streams of unstructured data, adaptive ranking in search engines as well as sentiment detection and analysis in Social Web data). He is author of more than 150 publications on international journals (h-Index: 30), proceedings of International Conferences and Workshops.