Word Networks

Relation Extraction from Texts

Sometimes, text data are the only source of information about a network. Most of the respective cases are instances of one or more of the following types of networks:

• Networks that are inaccessible or unobservable for the data collector. Prominent examples are covert networks, such as groups engaged in organized crime and secret societies, and cognitive models, which are representations of the knowledge and information that individuals hold in their minds.
• Networks that have ceased to exist at the time of data collection. Examples are former cultures and bankrupt companies.
• Large-scale networks for which gathering data within the appropriate network boundaries by using traditional network data collection methods, such as surveys, is prohibitively expensive, and which do not allow for sampling due to the skewed distribution of the number of links per node. Examples are sizable communities of practice, geopolitical entities, and the diffusion of behavior through segments of society.
• Virtual networks that do not necessarily feature an underlying, real-world social network and that are confined to the traces of behavioral data generated by the members of the network. Examples are open collaboration initiatives.

Text data that potentially contain relevant information about such networks include documents authored by members of these networks, such as diaries, narratives, interpersonal communication, mission statements, and annual reports, and material originating from outside the network, such as newswire data, reports from subject matter experts, and transcripts of court hearings. In these cases, relation extraction methods can be employed to identify the relevant pieces of information and their connections as they are explicitly or implicitly represented in the text data and to convert this information into the nodes and edges of a network. Across many relation extraction methods, triples comprising the subject, action, and object of an event or phenomena form the smallest structured unit in a network of words. Depending on the method and user's needs, these data can be enhanced, for instance, with spatial and temporal information, attributes, and weights.

[Page 959]Two examples illustrate the usage of relation extraction: First, information about the who, what, when, where, why, and how of an event can be converted into nodes of the type agent, task or event, date, location, motivation or sentiment, and means or resources, respectively. Connecting the dots results in a structural representation of a single event; that is, a network of words comprising nodes and edges. Applying this process to data on many events, such as long-term and large-scale news feeds, allows analysts to generate network data that can be used to investigate the structure, properties, evolution, and behavior of complex, dynamic, and sizable sociotechnical networks. Another example includes transcripts of narrations, interviews, and conversations, to which anthropologists, cognitive scientists, and social scientists, among others, apply relation extraction methods in order to identify the themes, intensions, and emotions addressed by the authors. Linking up these concepts according to how the authors of the documents had connected them can result in cognitive models. Such models are used, for instance, to study how students and members of teams coincide and differ in their understanding and perception of certain phenomena.

...