Using toponym co-occurrences to measure relationships between places: review, application and evaluation

While there is consensus that network embeddedness of cities is of great importance for their development, the precise effect is difficult to assess because of a lack of consistent information on relations between cities. This paper presents, applies and evaluates a rather novel method to establish the strength of relationships between places, a method we refer to as ‘the toponym co-occurrence method’. This approach builds the urban system on the basis of co-occurrences of place names in a text corpus. We innovate by exploiting a so far unparalleled amount of data, namely the billions of web pages contained in the commoncrawl web archive, and by applying the method also to small places that tend to be ignored by other methods. The entire settlement system of the Netherlands is consequently explored. In addition, we innovatively apply machine learning techniques to classify these relations. Much attention is paid to solving biases deriving from place name disambiguation. Gravity modelling is employed to assess the resulting spatial organization of the Netherlands. It turns out that the gravity model fits very well with the pattern of relationships between places as found in digital space, which contributes to our assessment that the toponym co-occurrence method is a solid proxy for relationships in real space. Using the method, it is established that the relationships in the Randstad region, by many considered a coherent metropolitan entity, are actually somewhat less strong than expected. In contrast, historically important, but nowadays small cities in the periphery tend to have maintained their prominent position in the pattern of relationships. Suburban, relatively new places in the shadow of a larger city tend to be weakly related to other places. Several suggestions to further improve the method, in particular the classification of relationships, are discussed.