Learning-Oriented Question Recommendation Using Bloom’s Learning Taxonomy and Variable Length Hidden Markov Models - Transactions on Large-Scale Data-and Knowledge-Centered Systems - page 30

and, therefore, represent a potential source for learning and establishment of
new fields of study. Such systems have become not just a source of informa-
tion retrieval, but also a medium for online information seeking and knowledge
sharing [ 15 ], a means towards a unique learning experience. However, the expo-
nential growth in the data volume of QA systems has made the users access to
the desired information more dicult and time-consuming [ 15 ].
Current QA systems integrate traditional content-based recommendation
engines with the goal to identify the most suitable user to answer a question
[ 11 , 12 ], but little research aims at filtering out for the user the questions/answers
that might be of interest [ 15 ]. The drawback of such approaches is that the rec-
ommender does not take into account explicitly the user's learning goals or
learning process, neither the order in which questions are selected. The goal of
this paper is to improve the learning experience of the user in the role of question
asker.
Recent research in the field of query recommendation for search engines are
based on query search graphs that aim at extracting interesting relations from user
query logs [ 3 , 4 ]. Some of these graphs are constructed based on relations between
queries, which are explored and categorized according to different sources of infor-
mation (e.g., words in a query, clicked URLs, links between their answers). Other
techniques rely on the co-occurrence frequency of query pairs, which are part of the
same search mission [ 5 , 7 - 9 ]. However, these approaches do not take into account
the user's search goal. A recent attempt to tackle this issue is presented in [ 10 ].
In [ 10 ], the authors propose a general approach to context-aware search using
a variable length hidden Markov model (vlHMM). This work is motivated by
the belief that the context of a users query, i.e. the past queries and clicks in the
same session, may help understand the users information need and improve the
search experience substantially. Cao et al. [ 10 ] develop a strategy for parameter
initialization within the vlHMM learning, which can reduce the number of para-
meters to be estimated in practice. Additionally, they devise a method for dis-
tributed vlHMM learning under the map-reduce model. Within this context, the
authors also argue that by considering only correlations between query pairs, the
model cannot capture well the users search context. In order to achieve general
context-aware search, a comprehensive model is needed that can be used simulta-
neously for multiple applications (e.g. query suggestion, URL recommendation,
document re-ranking). They propose a novel model to support context-aware
search and develop ecient algorithms and strategies for learning a very large
vlHMM from big log data. The experimental results show that this vlHMM-
based context-aware approach is effective and ecient.
Despite the extensive research in this area and the successful application of
such methods, they are not suitable for QA systems for at least two reasons. First,
the recommendation items are represented by questions as well-formed grammat-
ical units endowed with semantic content, whereas search queries are usually a
collection of keywords. Secondly, most QA systems are used with the purpose
of learning (e.g., find an explanation for a particular phenomenon, understand a