PhD Research Seminar: Topic modeling, question answering, analysis of trading systems

First talk: Stability of Topic Modeling   
Speaker: Roman Derbanosov, third-year PhD student, Faculty of Computer Science

Topic modeling is a statistical method for analyzing a corpus of documents. The result of the modeling is a set of topics. One of the main problems of topic modeling is instability, i.e., convergence to different sets of topics from different initializations. Two main approaches to this problem are the study of mathematical origins of the problem and customization of basic algorithms to achieve better stability. We will give an overview of existing results and our contibution to both approaches.

Second talk: Еnriching Question Contexts with СonceptNet Knowledge for Improving Answer Accuracy
Speaker: Denis Smirnov, second-year PhD student, Faculty of Computer Science

Modern question-answering models can achieve near-human accuracy of answers to factual questions about a given piece of text in English. However, such models fail to achieve the same performance on datasets of questions that require commonsence knowledge (as background information) not present in the question context. This talk intoduces the problem of question answering with commonsence knowledge and describes experimental evaluation of a simple question context enrichment method based on collecting ConceptNet relations.

Third talk: Modeling and Analysis of Trading Systems Using Petri Nets   
Speaker: Julio Carrasquel, second-year PhD student, Faculty of Computer Science

Trading systems support the automatic exchange of securities (e.g., company shares) between market participants. Participants take many benefits from these systems. For example, investors buy securities with promising returns. Companies sell percentages of their holdings to obtain funding. This is why trading systems have become an important element in global finances, and thus their correct design and analysis are tasks of utmost importance. In the literature, there are some proposals for designing and validating trading systems, which are based on machine learning or transition systems. However, these approaches do not properly capture all the concurrent end-to-end processes being executed in such systems.

In our research, we consider Petri nets, a formalism for modeling/analysis of concurrent systems. We work with Petri net extensions that allow us to describe data aspects (e.g., how objects such as orders to buy/sell securities are manipulated) or agent aspects (e.g., how participants interact with the trading platform). In this talk, we will see how these extensions can be used for constructing formal specifications of the system and how we can compare them against execution traces of real systems, so that we can answer the question: Is the real system behaving as specified? In addition, we will present our latest work consisting of a new Petri net-based formalism that can be used to cover simultaneously data- and agent-related aspects of trading systems in single integrated models.