Chapter 6 Disseminating and Sharing Task-relevant
6.2 Adaptive task-based profiling approach
This section describes the adaptive task-based profiling approach which models workers’ dynamic information needs on the target task. The adaptation of profile is based on users’ access behaviors or relevance feedback on knowledge items. A modified relevance feedback technique is employed to adjust workers’ profiles.
6.2.1 Profile modeling and structuring
We propose an adaptive task-based profiling approach to model workers’ dynamic information needs via feedback analysis. Meanwhile, worker’s relevance feedback is modeled by a fuzzy linguistic approach, as described in the following.
Perception modeling through fuzzy linguistic approach. The fuzzy linguistic
approach is a technique for approximating human perception, and provides easier assess to qualitative problems. Linguistic assessment uses words rather than numbers.For example, the linguistic variable “Relevance” is defined to assess the degree of relevance between objects (such as document, task, etc.). Notably, a linguistic variable is characterized by a quintuple (S, E(S), U, G, M) as defined in Definition I of Appendix A [91]. The semantic meaning of a linguistic term can be formulated as a fuzzy number, which represents the approximate value of each linguistic term.
Profile structuring. Two kinds of profiles, feature-based task profile and
topic-based task profile, are maintained. Both profiles are used to represent a worker’s current information needs on the target task at hand.•
Feature-based task profile: The feature-based task profile of a task t
r is a feature vector of weighted keywords, denoted as tGr= <wkw1, wkw2,…,wkwn>.
The representation of feature-based task profile is the same as the profile that generated based on the task-relevance assessment result.
•
Topic-based task profile: The topic-based task profile of a worker u, denoted
as WPu = {<topicj, wp(topicj)>}, contains a set of topics (fields or tasks in domain ontology) with associated degree of relevance to the target task at a specific time period. wp(topicj) represents the relevance degree of topicjto the
target task at time p, from the aspect of u. Let FS denote the set of topics in field level and TS denote the set of topics in task level. Note that the category level is not considered since the topics in category are too general todifferentiate workers’ task needs. The associated degree of relevance indicates a similarity measure between a topic and the target task at a specific time period.
The similarity measure is initially obtained from a worker’s relevance assessment, and will be updated via a worker’s explicit feedback (e.g., relevance rating) or via analyzing a worker’s access behaviors and explicit feedback, will be addressed in Section 6.2.2.
A topic-based task profile represents a worker’s task-needs expressed as a set of relevant fields or tasks in domain ontology, and can be used to derive a worker’s personalized ontology (WPO) on the target task. An ontology threshold value
δ
can be defined by a worker to generate a worker’s personalized ontology on the target task by filtering out irrelevant fields or tasks with relevance degrees below the threshold value. Accordingly, WPOu = {<topicj, wp(topicj)> | wp(topicj) ≥δ
andtopic
j∈FS∪TS }. The result forms a worker u’s personalized ontology on the target task.6.2.2 Profile adaptation based on feedback analysis
Document feedback analysis. A temporal profile, denoted as
TGempu,p, is generated by the profile handler to represent a worker u’s current information needs on the target task. The temporal file is derived from the feature vectors of those documents accessed by worker u during time period p, as shown in the Eq. 6.1.
Du denotes the set of documents which had been explicitly rated by worker u in conducting the target task during the time period p. Au
(d
j) denotes worker u’s crisp
rating on the relevance of document dj to the target task. The crisp rating is derived from the linguistic rating according to the center of area (COA) method described previously. Duimp,p denotes the set of documents which had been browsed and accessed but not been rated by worker u during time period p. A linguistic rating“High” is given by default to represent the relevance degree of unrated documents (implicit feedback). CV(H~)u denotes the corresponding crisp value of relevance rating “High“ of worker u. Notably, our system will show the description of a document. Thus, we assume that a worker will read the description first to decide if
the document is relevant, and then access and browse the document. Accordingly, a linguistic rating “High” is assigned to unrated documents that had been accessed and browsed by the worker.
The similarity between the temporal profile and a topic tj in the domain ontology can be derived by cosine measure, namelysim(TKempu,p,tGj)
Profile adaptation. The new feature-based task profile of the target task, denoted as
+1
SGp
is generated based on Eq. 6.2, which is modified from standard Rocchio (1971) and Ide (1971) algorithms presented in Section 2.2.2. The modification considers the associated relevance degrees of relevant/irrelevant tasks to the target task and the temporal profile derived form the feedback analysis.
∑
∈denotes the feature-based task profile of the target task at time p.
Notably, Sp
G may be an initial task profile derived from the initial assessment result.
The OG
denotes the aggregated relevant feature vector of the target task. The aggregation of irrelevant feature vectors is derived from Tn, which is the set of irrelevant tasks. The relevant feature vector OG
is derived based on Tr , the set of relevant tasks, and the temporal file generated from the feedback analysis. Herein, the set of Tn, and Tr are derived from worker’s feedback result. wp+1(tj) denotes the relevance degree (associated weight ) of task tj to the target task. TGempu,p
denotes the temporal profile derived from the feedback analysis. Meanwhile, α β γ, , are tuning constants. The parameter λ is used to adjust the relative importance of relevant tasks and the temporal profile. Note that there are two alternatives to derive
w
p+1(tj).• Explicit relevance feedback on tasks: wp+1(tj) denotes the worker u’s crisp rating on the relevance of existing tasks tj to the target task.
• Adjusting relevance degree of tasks by documents feedback.
The topic-based task profile will be adjusted based on the result of feedback analysis. Note that a topic-based task profile records topics (tasks or fields) with
associated relevance degree to the target task. The incremental analysis means the system learning the worker’s TRTs (set of relevant tasks) and TRITs (set of irrelevant tasks) on target task from the document feedback analysis. The result of document feedback analysis is to generate a temporal profile to represent a worker’s task-needs. Thus, the relevance degree between target task and topics (tasks or fields) are obtained by the similarity calculation between temporal profile and topics.
Moreover, since the relevance degree will be adjusted across time (increased or decreased), we named it as an incremental analysis procedure. The method to adjust associated relevance degree of each topic is addressed as follows.
The system will increase or decrease the relevance degree (associated weight) of a task tj (a topic in the task level of domain ontology) gradually, where wp+1(tj)=
w
p(tj)±∆ w. The adjustment ∆ w of a task t
j is derived based on the proportion of feedbacks and the similarity between the temporal profile and tj. If sim(TKempu,p,tGj)is above a relevance-adjustment threshold
θ , the system will increase the associated
weight of task tj. Meanwhile, if sim(TKempu,p,tGj)is below
θ
, the system will decrease the associated weight of task tj. The adjustment equation is given below.θ
where Nud denotes the number of documents accessed and browsed by worker u in conducting the target task prior to time p, while Nud,p denotes the number of documents accessed and browsed by worker u in conducting the target task during time p. Notably, wp+1(tj)=1, if wp(tj)+
∆ w >1; w
p+1(tj)=0, if wp(tj)-∆ w < 0. Moreover,
a field contains a set of tasks. Thus, the value of wp(fieldi) is set to the maximum value of wp(tj) for any task tj belongs to fieldi. Namely, the weight of fieldiwill be
adjusted at time p+1, where wp+1(fieldi) = maxtj∈fieldi(wp+1(tj)). Thus, it is a incremental analysis procedure to calculate the relevance degree of topics to target task over time. Meanwhile, the adjustment may change the information structure of a worker’s personalized ontology. The personalized ontology of worker u is adjusted by removing an irrelevant topic tj , if wp+1(tj) is below the ontology thresholdδ
, and adding a relevant topic tj , if tj did not exist at time p and wp+1(tj)≥δ
. Accordingly, this approach is different from the method of explicit feedback. The degree of relevance of a topic, wp(topicj), in the work profile is inferred from incremental analysis instead of explicit feedback on tasks, i.e., relevance rating.Furthermore, the feature-based task profile of the target task can be adapted based on the adjustment of topic-based task profile. The system generates a set of top-k relevant tasks (denoted as TRTs) and a set of top-k irrelevant tasks (denoted as TIRTs) based on the topic-based task profile. Relevant tasks are those tasks with associated relevance degree wp+1(tj) higher than the relevance-adjustment threshold
θ
, whereas irrelevant tasks are those tasks with wp+1(tj) lower thanθ
. Only the feature terms and the associated relevance degrees of topics in the task level are used to adjust the task profile of the target task. A field is a generic view of similar tasks;hence, the feature terms of fields are not as representative as the feature terms of tasks for the target task. The new feature-based task profile of the target task, denoted as SGp+1
is also generated based on Eq. 6.2.
The advantage of this method, i.e., adjusting relevance degree of tasks by document feedback, is that it dose not require workers to conduct tedious relevance feedback explicitly. However, this method need to continuously track and record workers’ access behaviors. Moreover, the time factor is also important to analyze the worker’s access behaviors. More recent access behaviors should give higher weight than earlier access behaviors in adjusting the relevance degree of tasks. Thus, further study is needed to investigate and evaluate this method.
6.2.3 K-Delivery: Delivering codified knowledge proactively
A worker’s feature-based task profile and topic-based task profile can properly reflect a work’s task-needs on the target task. The profiles can be used to further enhance the knowledge retrieval capability in the proposed system. Moreover, the adjustment of work profile across time will lead the system to refine the task profile based on the proposed profile adaptation approach. Accordingly, SGp+1
is used to retrieve relevant codified knowledge in the repository. Relevant task and document sets will be retrieved to (e.g. cosine measure). Figure 8 is the interface of knowledge delivery in which the system delivers task-relevant knowledge proactively based on the task profiles.
Fig. 8. Interface of knowledge sharing (α=0.9)