Author Archive: Svetlozara Lesseva

Short bio

---

Dr. Veselin Stoyanov is a Researcher with a track record of innovating in AI and NLP to solve real-world problems. He is currently the Head of AI at Tome building practical applications of LLMs. He was previously at Facebook AI, where he led the development of industry-standard Large LM methods such as RoBERTa, XLM-R, and MultiRay and their application to improve online experiences, e.g., reduce the prevalence of hate speech and bullying posts. He holds a Ph.D. from Cornell University and a Post-Doctoral degree from Johns Hopkins University.

Talk abstract

---

Large Language Models for the Real World: Explorations of Sparse, Cross-lingual Understanding and Instruction-Tuned LLM

Large language models (LLMs) have revolutionized NLP and the use of Natural Language in products. Nonetheless, there are challenges to the wide adoption of LLMs. In this talk, I will describe my explorations into addressing some of those challenges. I will cover work on sparse models addressing high computational costs, multilingual LLMs addressing the need to handle many languages, and work on instruction finetuning addressing the alignment between model outputs and human needs.

Short bio

---

Prof. Vito Pirrelli is Research manager at the National Research Council Institute for Computational Linguistics Antonio Zampolli since 2003, he is head of the Laboratory for Communication Physiology, and co-editor in chief of the Mental Lexicon and Lingue e Linguaggio. His main research interests focus on fundamental issues of language architecture and physiology, lying at the interdisciplinary crossroad of cognitive linguistics, psycholinguistics, neuroscience and information science.

Over the last 20 years, he has been leading a data-driven research program that uses artificial neural networks, language models and information and communication technologies to investigate language as a holistic dynamic system, emerging from interrelated patterns of sensory experience, communicative and social interaction and psychological and neurobiological mechanisms. This research program went beyond the fragmentation of mainstream NLP technologies of the early 21st century, allowing innovation to come out of research labs and address societal needs. Using portable devices and cloud computing to collect ecological multimodal language data, the Comphys Lab currently offers a battery of tools, resources and protocols that support language teaching and education assessment, cultural integration and early diagnosis and intervention of language and cognitive disorders.

In 2021, following a peer review by the relevant Class Committee, he was elected member of the Academia Europaea.

Talk abstract

---

Written Text Processing and the Adaptive Reading Hypothesis

Oral reading requires the fine coordination of eye movements and articulatory movements. The eye provides access to the input stimuli needed for voice articulation to unfold at a relatively constant rate, while control on articulation provides internal feedback to oculomotor control for eye movements to be directed when and where a decoding problem arises.

A factor that makes coordination of the eye and the voice particularly hard to manage is their asynchrony. Eye movements are faster than voice articulation and are much freer to scan a written text forwards and backwards. As a result, given a certain time window, the eye can typically fixate more words than the voice can articulate.

According to most scholars, readers compensate for this functional asynchrony by using their phonological buffer, a working memory stack of limited temporal capacity where fixated words can be maintained temporarily, until they are read out loud. The capacity of the phonological buffer thus puts an upper limit on the distance between the position of the voice and the position of the eye during oral text reading, known as the eye-voice span.

In my talk, I will discuss recent reading evidence showing that the eye-voice span is the “elastic” outcome of an optimally adaptive viewing strategy, interactively modulated by individual reading skills and the lexical and structural features of a text. The voice span not only varies across readers depending on their rate of articulation, but it also varies within each reader, getting larger when a larger structural unit is processed. This suggests that skilled readers can optimally coordinate articulation and fixation times for text processing, adaptively using their phonological memory buffer to process linguistic structures of different size and complexity.

Short bio

---

Prof. Joakim Nivre is Professor of Computational Linguistics at Uppsala University and Senior Researcher at RISE (Research Institutes of Sweden). He holds a Ph.D. in General Linguistics from the University of Gothenburg and a Ph.D. in Computer Science from Vaxjo University.

His research focuses on data-driven methods for natural language processing, in particular for morphosyntactic and semantic analysis. He is one of the main developers of the transition-based approach to syntactic dependency parsing, described in his 2006 book Inductive Dependency Parsing and implemented in the widely used MaltParser system, and one of the founders of the Universal Dependencies project, which aims to develop cross-linguistically consistent treebank annotation for many languages and currently involves nearly 150 languages and over 500 researchers around the world. He has produced over 300 scientific publications and has over 42,000 citations according to Google Scholar (February, 2024). He is a fellow of the Association for Computational Linguistics and was the president of the association in 2017.

Talk abstract

---

Ten Years of Universal Dependencies

Universal Dependencies (UD) is a project developing cross-linguistically consistent treebank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. Since UD was launched almost ten years ago, it has grown into a large community effort involving over 500 researchers around the world, together producing treebanks for 148 languages and enabling new research directions in both NLP and linguistics. In this talk, I will review the history and development of UD and discuss challenges that we need to face when bringing UD into the future.

Short bio

Hristo Tanev is a project officer and researcher at the Joint Research Centre of the European Commission. His research spans across various areas of computational linguistics and natural language processing, including event extraction, text classification, question answering, social media mining, lexical learning, language resources, and multilingualism.

He is a co-organizer of the Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text. He has carried out research in three research institutions: University of Plovdiv Paisii Hilendarski (Bulgaria), ITC-irst (now Fondazione Bruno Kessler), Trento, Italy, and the Joint Research Centre of the European Commission, Ispra, Italy. He is among the founders of SIG SLAV (Special Interest Group of Slavic Language Processing) at ACL.

 

Demo abstract

Ontopopulis, a System for Learning Semantic Classes

Ontopopulis is a multilingual terminology learning system which implements several weakly supervised algorithms for terminology learning. The main algorithm in the system is a weakly supervised one which takes on its input a set of seed terms for a semantic category under consideration and an unannotated text corpus. The algorithm learns additional terms, which belong to this category. For example, for the category “environment disasters” in Bulgarian language the input seed set is: замърсяване на водите, изменение на климата, суша. The highest ranked new terms which the system learns for this semantic class are : опустиняване, обезлесяване, озонова дупка and so on.

In the demo session we are going to show how the system learns different semantic classes in Bulgarian and English.