Biomimetic information retrieval with spreading-activation networks

Abstract

Information management systems act as a prosthetic scaffold for human memory. They retain and organize information objects to be conveniently recalled in support of knowledge-based tasks. We note a striking similarity between the functions of human memory and the processes in computational information retrieval. For this reason, we ask whether it is viable to purposely design information management systems biomimetically, i.e., in a manner inspired by biological systems. Based on a comparison of cognitive models of human memory and computational information retrieval algorithms, we propose the Principles of Mnemonic Associative Knowledge (P-MAK) to describe the necessary components of biomimetic systems: the constraints of computing machines, the properties of human memory, how semantic knowledge representations are constructed, and the contexts in which information is usefully retrieved. The goal of P-MAK is to describe systems that are simple, inspectable, comprehensible, and easy to use. Since human memory as described by cognitive network models is analogous to a large associative hypertext repository, P-MAK's principles suggest that networks would be an appropriate representation format. Therefore, we build a semantic similarity network from a document corpus using information retrieval (IR) algorithms, and describe how these processes are comparable to the functions of human semantic memory. To approximate an optimal link distribution, we introduce a novel link-pruning technique to tune the network to a small-world topology. We show in a user study that a semantic network based on cognitive models can improve user access to information. The ability to recall information in appropriate contexts is also a useful property of human memory. Based on models of human episodic memory, we propose a real-time, incremental temporal index that captures some of the regularity of human information behaviour. Temporal patterns are represented using a novel cue-event-object (CEO) model, in which observed events are related to a collection of cues. The cues describe time, place, or sensory qualities and are analogous to cognitive schemas . Cues are combined to represent an event, analogous to cognitive convergence zones. The model connects related cues, events, and objects together to encode the relations present in observed occurrences. The CEO model simulates cognitive reinforcement learning to build patterns of user information behaviour. If an object is used consistently at a given time, the links connecting cues, event, and object all grow stronger; otherwise, they decay and are forgotten. The resulting network structure can function as a recommender system by using spreading activation to retrieve objects at times and under circumstances where they have previously proven themselves useful. The model also allows users to pose queries such as when an event typically occurs, or what items are used at particular times. In a user-log experiment, we show that the CEO model quickly learns to make correct predictions of user behaviour, and increases in accuracy the more data that it is given.