Hadi Abdikhojasteh: Neural Methods for Document Ranking
Document ranking in information retrieval systems is the process of sorting the collection of documents based on the user input queries. Such systems aim to rank documents that are more relevant higher than the documents that are less relevant by assigning scores. This score indicates the quality of the assigned document and the search query. In cross-lingual information retrieval systems, documents and queries are written in different languages or even formats. In order to target this gap, a successful query expansion method has been introduced.
Recently, multiple neural language models have been widely exploited, such as BERT, ELMo and XLNet, that model the underlying data distribution and learn the linguistic patterns or features in the language. Inspired by these studies, novel query expansion models have been introduced for cross-lingual systems.
The talk will explore neural methods for documents ranking and novel researches for ranking and query expansion. Outline experiences and address our initial results in the multilingual information task for COVID-19.
Tomáš Nekvinda: Multi-domain dialogue systems
Task-oriented dialogue systems are typically handcrafted or trained from data for a small set of domains. Non-task-oriented chit-chat systems are trained to respond to open-domain utterances, but they have limited understanding and their responses are largely uncontrolled. My dissertation topic focuses on the exploration of novel and efficient ways of building multi-domain dialogue models that are able to jointly serve task-oriented and non-task-oriented dialogues. This may include, for instance, improving task detection, open-domain understanding, adding chit-chat capabilities to end-to-end task-oriented systems, etc. I will talk about the motivation, main obstacles, the recent research in this field, and my initial experiments.
Mateusz Krubiński: Multimodal Summarization
The goal of Automatic Summarization is to produce a concise summary of a given document. The output should contain the most relevant information within the original content, while also being short and easy to process.
In recent years there is a growing interest in multimodal approaches, which combine several sources of information, i.e. videos, texts and images. The aim is to produce a multimodal summary consisting of e.g. short textual summary and a cover photo. The area is still missing established benchmarking datasets and a single metric to measure effectiveness of such systems.
In this talk I will present several variants of this challenge, point out some important obstacles and briefly mention recent advances in other Vision+Language tasks such as Video Captioning or Visual Question Answering that may be useful for Multimodal Summarization.
***The talks will be streamed via Zoom. For details how to join the Zoom meeting, please write to sevcikova et ufal.mff.cuni.cz***