Automatic extraction of multiple underlying causes from textual death records

Data sets containing natural language strings are increasingly becoming available as outputs from various international initiatives to digitise historical population records. Analysis of such records, for example, historical causes of death, is facilitated by classification to standard systems such as ICD-10.
Most death record systems include multiple causes of death, typically a primary cause and optionally a number of contributing secondary causes. For example, the primary cause of death might be heart failure, contributed to by some underlying chronic condition.
In modern records these separate causes are clearly differentiated by being entered in different fields on the recording form. In some historical records, however, there was no imposed structure, with the person recording the death having the freedom to use any form of language.
The Digitising Scotland project is in the process of transcribing all Scottish birth, death and marriage records from 1855 to 1973. Here we describe our approach to automatic extraction of multiple causes of death from the approximately 11M death records.