Information extraction

In the context of the ROLAQUAD project (an NWO IMIX programme), I am working on domain-specific information extraction. The application domain of the project is general (i.e. non-expert) medical information written in Dutch. Within the project, machine learning components are developed for analysing unstructured texts on multiple levels of semantic annotation. These levels are:

token level

Domain-specific named entitiy recogntion; for example, identifying occurrances of illnesses or treatments.

sentence level

Detecting and identifying sentence topics, roughly corresponding to domain-specific relations; for example, whether a sentence gives information on the cause or prevention of an illness, or on the side-effects of a treatment..

section level

High-level classification of the topic(s) of a document section. Possible section topics are those commonly found in medical encyclopedias, such as definition, cause, and treatment.

One of the aims of the ROLAQUAD project is to exploit (partially) redundant information stored on these levels of semantic annotation. For example, in the correct annotation, the sentence below is assigned the sentence topic causes and the word influenza is labelled as referring to an illness.

Influenza is a contagious infection of the airways caused by the influenza virus.

The presence of an illness in the sentence may be highly informative for determining whether the sentence expresses a causes relation. Likewise, knowing a sentence to have causes as one of its topics, strongly points at the presence of at least one illness in the sentence. Effectively exploiting this redundancy for classification may require architectures more complex than a sequential pipeline, in which the semantic annotation levels are predicted one after the other. Developing architectures optimised for predicting multiple interacting levels of information is central to the ROLAQUAD project.