Semantic Networks: Properties, Representations, and Applications
Introduction
Large tech companies (Facebook, Google) and telecom providers (KPN, Ziggo) each collect and process tons of textual data on what their users say online. For these companies, the automated analysis of the text is an important tool with applications in internet search, product recommendation, and sentiment analysis. A systematic analysis of unstructured textual data is, however, extremely non-trivial.
Examples of text analysis tasks that have been considered in the literature are analogy finding, sentiment analysis, and sentence completion. Unfortunately, existing algorithms do not perform optimally, and understanding text computationally remains a difficult task.
One way to analyze text is through the lens of a semantic network: Semantic network is, in most general terms, a network of words or phrases that represent nodes that are connected in case the nodes are semantically related (e.g., co-occur in text, are synonyms or antonyms, etc.). Semantic networks represent knowledge or ‘meaning’—and they could allow computers to reason like a real person would [1]. In general, semantic networks are expected to play an important role in creating better artificially intelligent systems.
Track 1: Topological Properties of Semantic Networks
While semantic networks are extensively used in text analysis, their general topological properties are poorly studied. Early results demonstrate that semantic networks are sparse (number of links is comparable to the number of nodes), heterogeneous (number of links per node varies widely), and possess hierarchical structure [2]. A better understanding of the structural properties of semantic networks will be instrumental in understanding their design principles and will ultimately lead to better performance of text analysis algorithms.
Goal: In this project, you will learn to extract semantic network data from various sources and/or build semantic networks from raw text. You will catalog the extracted networks and analyze their structural properties.
Prerequisites: The successful student is expected to have good programming and data extraction skills, background in Probability and Statistics, and Network Science (EE4C06, CS4195 (or analogous) courses are recommended).
Track 2: Network Embedding Techniques
Text analysis can be performed through network embedding: network nodes can be mapped to points in certain (latent) space, such that related words are mapped close to each other in the space.
Then, new relations between words can be uncovered through the identification of unconnected word pairs in the geometric vicinity of one another. Another network embedding application is analogy solving, e.g., find X, such that X to Germany is like Paris to France. A basic geometric idea to the analogy problem is to find a vector connecting France to Paris in the latent space, then draw the same vector from Germany. The resulting vector should point to X, which is expected to be Berlin.
Although in the past decade, researchers developed a lot of network embedding techniques, most of these techniques are ad-hoc and are not systematically evaluated.
Goal: In this project, you will study existing network embedding techniques for semantic data analysis, learn basic principles of statistical inference, and conduct a systematic comparative analysis of the performance of network embedding techniques on different semantic tasks.
Prerequisites: The successful student is expected to have strong programming skills, have a background in Probability and Statistics and Network Science (EE4C06, CS4195, or analogous courses are recommended).
To apply:
Contact Maksim Kitsak,
References
[1] Sowa, J. F. (1987). Semantic networks. Wiley.
[2] Van Mieghem, P. (2014). Performance analysis of complex networks and systems. Cambridge University Press.