Layers 2 and 3 - Partition Words

Obviously, not all of the frequently occurring words are of use in characterizing the context. Terms such as ``the,'' ``and,'' and other common words occur frequently in many documents regardless of the context. Layer 1 learns frequently occurring words, including both context-determining words and common words with low information content. Layers 2 and 3 determine which of these words characterize the context.

The upper two layers (2 and 3) are continuously presented with the state of the units in the bottom layer. Each upper layer contains 500 units, which are paired such that units in both layers are sensitized to the same words. Layer 2 detects words that occur frequently in successive documents. Layer 3 detects which of the words in layer 2 tend to not occur for long periods of time. Each unit is associated with two values: excitement and priming. In layer 2, a unit's priming increases while its word is present in layer 1. The unit's excitement increases as a function of the priming. Excitement and priming will decay when the word is absent. In layer 3, almost the opposite happens: the priming increases while the word is absent in level 1. The excitement changes as a function of the priming.

Levels 2 and 3 build up user access profiles. When the excitement values of corresponding terms in layers 2 and 3 are multiplied together, the terms with higher values tend to be the terms which occur frequently for discrete periods of time in the user's document accesses. Such terms can characterize a user's information-seeking task, as is shown in the next section.