Monday, 10 May, 2021 - 14:00

UFAL PhD conference - Part II

Peter Polák
Hashem Sellat
Vadim Gudkov (ÚFAL MFF UK)


Peter Polák: Babel Fish: Multilingual Speech Recognition and Translation


Automatic Speech Recognition (ASR) is a task of transcribing recordings, i.e., transforming speech to text. Closely related is the Speech Translation (ST), which additionally translates the transcription into another language. Our goal: Babel Fish: a multilingual ST system transcribing speech from many languages to a single target language.

While the modern goal (and our long-term aim) are end-to-end systems that handle the speech translation task at once, we start by breaking the problem into two pieces: ASR followed by Machine Translation (MT). This has two advantages: enables us to concentrate on many challenges separately and leverage non-speech data. One of the challenges is a high variability of speech data. The input quality may vary for different reasons, such as by input devices (studio, a computer microphone), background noise, accented speech (non-native speakers) or disfluent speech. We also deal with the problem of the selection of modelling units between ASR and MT (graphemes, phonemes, ...). Finally, we address another major issue, the lack of domain- or language-specific training data.

In our presentation, we will summarize current approaches for ASR and ST. We will talk about some of the open issues and outline experiments that aim to address these issues.




Hashem Sellat: Dialect adaptation of neural machine translation



Neural Machine translation is the process of using neural models to produce good quality translations. However, training such models require very large amounts of training resources (parallel data), which is typically available only for languages spoken by many people. A significant research effort has been recently put into the development of methods that allow training NMT models for language pairs with limited (or not existing) training data (unsupervised, zero-shot, semi-supervised). 

In this project, we focus on machine translation between English and Arabic. Arabic is a language spoken by more than 400 million people and provides substantial resources for language technologies including machine translation. However, virtually all such (written) resources are in the formal Modern Standard Arabic. We aim at developing an NMT system between English and Arabic dialects (taking the Levantine dialect as a prototype). First, we will exploit existing resources for MSA and state-of-the-art methods to deliver a state-of-the-art system for MSA-English translation. Then we will focus on methods for adaptation/fine-tuning/multitask learning to build a system that translates between the Levantine dialect and English, with or without using MSA as an intermediate language. 

I will talk about the state-of-the-art techniques and researches in deep learning to approach the problem. Also, the challenges related to dialects (Syrian dialect as an example) and collecting enough parallel data to train these models. 




Vadim Gudkov: Multilingual Contextualized Processing of Basic Syntactic-semantic Elements



The focus of my work is the notion of the syntaxeme - the minimal semantic-syntactic entity. As characterized by the founder of the syntaxeme-based approach to syntax, G.A. Zolotova, the syntaxeme is both the elementary carrier of meaning and a constructive component of more complex formations. Thus, a sentence can be regarded as a sequence of such functional elements. 

Although this theory is not being actively utilized in the domain of NLP, I believe that modern syntactic parsing, which greatly depends on language modelling and semantic representation, can be enriched by the adoption of the syntaxeme-based method. My research based on the Russian prepositional syntaxemes has shown that prepositional constructions can be effectively classified according to the function performed by the syntaxeme.

In my talk I will mainly focus on my research on the semantic classification of Russian prepositional syntaxemes, the potential areas of expansion of this approach to other structural entities as well as the potential of application of my research to other Slavic languages.




***The talks will be streamed via Zoom. For details how to join the Zoom meeting, please write to sevcikova et***