Supervised learning for relationship extraction from textual documents

Abstract

Information Extraction (IE) is the task of automatically extracting structured information from unstructured data, aiming to facilitate the use of said data by other applications. A typical sub-problem is the extraction of relationships from textual documents, which aims at identifying and classifying the relationships expressed between entities mentioned in the texts. In order to extract relationships from a raw text, it is important to pre-process the data, organizing the textual contents into useful data structures, with techniques from Natural Language Processing. Furthermore, since relationships are expressed between entities, it is mandatory to identify the entities using an entity extraction method, which is another sub problem of IE.

Assigning a relationship type to a pair of entities can be seen as a classification problem. Therefore, supervised machine learning techniques can be applied. In this thesis, we used Support Vector Machines (SVM), which we trained with basis on online methods similar to Pegasos. Two specific modelling choices have been tested. The first one is a simple online solution that trains SVM models considering a single kernel.

The second approach is based on the idea of online multiple kernel learning. With existing datasets and common pre-processing tools, we formulated a benchmark, which was then used to evaluate kernel-based methods. We then implemented state-of-the-art kernels, specifically designed for relationship extraction.

The results show that a multiple kernel learning solution obtains the best performance, and that multiple kernel learning solutions can perform better than heuristic solutions learning with linear combinations of the same kernels.