Utility-Based Information Distillation - Text Mining: Classification, Clustering, and Applications - page 223

9.3.3 Novelty Detection Component
To avoid showing information that the user has already seen, the system
maintains a user history H ( t ), which contains all the spans of text h i that the
user highlighted (as feedback) during their past interactions with the system,
up to the current time t . Each passage in the relevance list (Section 9.3.2) is
compared to the user history for novelty assessment.
Denoting the history as

H ( t )= h 1 ,h 2 , ...,h t , (9.2)

the novelty score of a new candidate passage x is computed as:

cos( x,h i )

f ND ( x )=1 − max

i

(9.3)

∈ 1 ..t

where both candidate passage x and highlighted spans of text h i are
represented as TF-IDF vectors.
The novelty score of each passage is compared to a pre-specified threshold
(also tuned on a training set), and any passage with a score below this
threshold is removed from the relevance list .

9.3.4 Anti-Redundant Ranking Component
Although the novelty detection component ensures that only novel
(previously unseen) information remains in the relevance list , this list might
still contain the same novel information at multiple positions of the ranked
list. Suppose, for example, that the user has already read about a $100,000
reward for information about the escaped convicts. A new piece of news that
the award has been increased to $200,000 is novel since the user hasn't read
about it yet. However, multiple news sources would report this information
and we might end up showing (redundant) articles from all these sources
in a ranked list. Hence, a ranked list should also be made non-redundant
with respect to its own contents. We use a simplified version of the Maximal
Marginal Relevance method, originally developed for combining relevance and
novelty in text retrieval and summarization (6). Our procedure starts with
the current list of passages sorted by relevance (Section 9.3.2) and filtered
by Novelty Detection component (Section 9.3.3), and generates a new non-
redundant list as follows:
1. Take the top passage in the current list as the top one in the new list.

2. Add the next passage x in the current list to the new list only if

f AR ( x ) >t

where

p i ∈L new { cos( x, p i ) }
and L new is the set of passages already selected in the new list.

f AR ( x )=1 − max