Computational Intelligence - May 2014 - 22

PWIs
Dp
D1
...
Importance
Ranking

Query
Expansion

Session
p
r1

Dn
Dt

...

Q
Relevance
Ranking

d1
d2
Ranked list of PWIs for ut
Figure 2 Diagram of context-aware personal information retrieval
framework. The framework takes a session and users' PWIs as input.
It works by first building a query Q from the session using the information in both the session and users' PWIs. A graph-based algorithm is
employed to derive the importance scores for each PWI of ut, which is
then combined with the relevance scores to obtain the final ranked list
of documents for ut.

Examples of S include conversations in SNSs, such as a
tweet posted on Twitter with replies from fellow followers, or a
question posted on a question answering website with answers
from other users.
In a Session ^S h, p denotes the initial post. U indicates the
set of users involved in S , i.e., U = {u p, u 1, f, u i, f, u n, u t} .
u p is the creator of p, u t is the targeting replier to whom we will
make recommendations, and u i ^u i g {u p, u t} and i = 1, ..., n )
is an existing replier. R = {r1, f, ri, f, rn} is the set of existing
replies, assuming that each user gives solely one reply.
D = D p , {D i} ni = 1 , D t is the entire collection of PWIs of all
the users within S , where D p, D t and D i (D i g {D p, D t}) are
the sets of PWIs of u p, u t and u i respectively. In addition,
D i = {d j} mj =i 1 is the PWIs of u i, where m i is the number of
u i 's PWIs.
The content of all the documents in S are represented by
the Vector Space Model [32], i.e., v d = [w 1, d, f, w k, d] T , in
which each term is weighted by its tfidf score [33].
IV. Context-Aware Personal Information Retrieval

text to build the query since they are the basic available context information in S. However, the posts in SNSs are usually short and ambiguous and thus are not sufficient to
characterize different properties of the session. Therefore, the
PWIs of the creator and the existing repliers are utilized to
obtain richer information. The main reason is that users participating in the same session usually share common interests
related to p, and they might have posted similar documents
in other SNSs before. Therefore, the PWIs of the creator and
the existing repliers can be used as the complementary context information.
In the second step, the shared interests among the targeting replier, the creator, and the existing repliers are considered, which forms the implicit-topical context. One of the
common interests of all the participating users is the topic
of the conversation they are involved in. By implicitly inferring the subset of documents on the topic in an unsupervised manner, the relevant PWIs of the targeting user can
be collected.
In the remaining parts of this section, we will describe the
two steps in detail and then combine the ranking results from
these two steps to obtain the final ranking scores for each PWI
of the targeting user.
B. Utilizing Participatory Context for Query Expansion

As aforementioned, the context of the session S consists of
an initial post p, existing replies, and the PWIs of the creator
and the existing repliers. To accurately model the context of
session S so as to build a comprehensive query for covering
the different properties of context, we need to extract as
many cues as possible from them.
In order to model the context, the query Q is built by
modeling the session at two levels. Formally, the initial p is
treated as the basic query. We first combine the replies of
existing repliers with p, since these replies are the responses
to p and thus can provide extra information about S . As different replies have different levels of correlation with p, the
replies are weighted according to their similarities with p.
The expanded query is calculated as follows:
v Q p +R = av p + (1 - a)

/ sim (v p, v r ) $ v r ,
i

(1)

ri ! R

In this section, we present the details of the CPIR method.
A. Overview

CPIR is mainly decomposed into two sub-steps: (i) query formulation and expansion: building a query by extracting sufficient information from the session; and (ii) PWIs ranking:
extracting the most relevant PWIs of the targeting replier
according to the query. The overview of CPIR is illustrated
in Figure 2.
In the first step, the participatory context is used to formulate and expand the query by considering both the replies
and the PWIs of all participating users. An intuitive idea is to
use the initial post p and the existing replies R as the con-

IEEE ComputatIonal IntEllIgEnCE magazInE | may 2014

where a (0 # a # 1) is a trade-off parameter to control the
contribution of replies. sim (., .) is the similarity metric to
measure the relevance between the initial post p and the
replier ri (ri ! R ).
Among the different similarity measures, it has been
shown that probabilistic methods like KL-divergence [34] can
obtain better results than vector space based measures [35]
[36], especially for short texts [37] like the PWIs discussed in
this paper. However, as the vocabulary in PWIs is sparse,
smoothing techniques are always introduced to take the entire
vocabulary into consideration to compare two distributions.
We introduce the translation-based language model [38] with

Table of Contents for the Digital Edition of Computational Intelligence - May 2014