We investigate how to provide personalized nudges to aid a user's exploration of linked data in a way leading to expanding her domain knowledge. This requires a model of the user's familiarity with domain concepts. The paper examines an approach to detect user domain familiarity by exploiting anchoring concepts which provide a backbone for probing interactions over the linked data graph. Basic level concepts studied in Cognitive Science are adopted. A user study examines how such concepts can be utilized to deal with the cold start user modelling problem, which informs a probing algorithm.
The recent growth of the Web of Linked Data1(LD), which provides access to big data
graphs representing domain entities and their relationships, has opened a new avenue
of research on developing computational models to facilitate data exploration by
layman users [
Our research brings a new dimension of personalization and adaptation to enhance
the benefits of linked data exploration, namely knowledge utility. We investigate how
to aid a user’s exploration of linked data in a way leading to expanding her domain
knowledge. This can have a broad implication for facilitating sense making while
exploring linked data. Learning is an inevitable part of exploratory search, as users are
discovering new connections and associations. Our earlier research has shown that
although linked data exploration can promote domain knowledge expansion
(‘serendipitous learning’ effect), not all paths can be beneficial. We derived empirically
strategies to nudge the user to beneficial paths [
Identifying the user familiarity with the domain entities (domain concepts or
instances) from the LDG is not a trivial task because LDGs usually include thousands
of entities at different levels of abstraction. This brings forth the well-known cold
start problem of user modelling, which is aggravated by the sheer number of LDG
entities. One way to address cold start is via a probing dialogue. While LDG can
provide a knowledge pool to implement a probing dialogue for user modelling (c.f. [
This paper examines an approach to detect user domain familiarity by using
anchoring concepts in the LDG around which a probing dialogue can be developed. We
adopt the Cognitive Science notion of basic level concepts (BLCs) – domain concepts
that are highly informative and can be easily retrieved from memory. An example of a
basic level concept in the Music domain is Guitar [
To get insights into how BLCs can be utilized to identify user's domain familiarity,
we conduct a user study that adopts earlier Cognitive Science methods which derive
BLCs in a specific domain [
We conducted a user study to examine how BLCs in a LDG can be utilized to model a user’s domain familiarity.
Dataset.We have used a dataset from the music domain which underpins a linked
data browser (MusicPinta) developed by us in an earlier research [
Ten online surveys were run adopting two strategies: (i) Strategy 1 – leaf instruments: eight surveys presented the leaf instruments – each survey presented 32 leaves and 8 additional images. (ii) Strategy 2 – category instruments: two surveys presentedthe category instruments- each survey showed 54 categories and 14 additional images. The image allocation in surveys was random. Every survey had 4 participants; each participant conducted one survey following an online link, including: • Pre-task questionnaire-collecting information about user profile (e.g. agegroup, nationality, and gender). • Free-naming task- Each image was shown for 10 seconds on the participant's screen and he/she was asked to type the name of the given object(s) in the image as quickly as possible. Figures 1-4 show example instrument images and participant’s answers from the study. For this task, we recorded accuracy (i.e. the participant answered correctly) and frequency (i.e. how many times a particular instrument name was mentioned correctly) of their accurate answers. • Post-task questionnaire- collected information about the participant's familiarity level for the six top level musical instrument categories (String Instruments, Wind Instruments, Percussion Instruments, Electronic Instruments, and Other Instruments). Participants were asked to rate their knowledge in these categories on a scale of 1 to 7 (1=No Knowledge and 7=Expert).
To extract BLCs from the MusicPinta dataset we considered accuracy and frequency
of the participants' answers [
Fig1. Leaf Violotta seen, named as Violin.
Fig2. Category Fiddle seen, named as Violin.
Fig3. Category Violins seen, named as Violins.
Fig4. Category Plucked String Instruments seen, named as String Instruments.
In each of the groups, LDG entities with frequency above 2 (i.e. they were named by at least two users) were included. The entities identified in Group 1 are derived from Strategy 1, while Group 2 and Group 3 give complementary output for the LDG entities derived by Strategy 2 (i.e. when the participants saw categories of instruments). Hence, the union of Group 2 and Group 3 gives the likely BLCs identified with Strategy 2, which is then intersected with the output from Strategy 1 to obtain the final BLC list. This included: Accordion, Bells, Bouzouki, Clarinet, Drums, Flute, Guitars, Harmonica, Harp, Saxophone, String instruments, Trumpet, Violins and Xylophone. 3
We compared the user survey answers and user's familiarity for the six top level musical instrument categories. The findings were used to derive probing heuristics.
When the user is able to name an instance (leaf instrument as seen in Strategy 1) instead of its corresponding BLC, she has high familiarity in the corresponding top level category. There were 27 cases (out of 41) where the participants could name leaf instruments rather than using their BLCs (as the majority of users did). For example, a participant named the leaf Electric cello instead of naming its BLC Violin. In 67% of these cases users had high familiarity with the top level instrument category.
When the user successfully names children of a basic level concept from images of the corresponding categories (as seen in Strategy 2), she has high familiarity in the corresponding top level category. There were 34 cases where the participants named children that belonged to the basic level. For example, one participant named the child Cello instead of naming it with its BLC Violin. In 62% of these cases participants had high familiarity with the top level instrument category.
When the user cannot name a basic level concept from the corresponding BLC images (as seen in Strategy 2), she has low familiarity in the corresponding top level category. There were 11 cases (out of 64) where participants were shown a BLC and were unable to name it. In these cases, the participants had indicated low or no knowledge with the top level instrument category.
When the user can name a basic level concept from the corresponding images for the BLC category (as seen in Strategy 2), she is likely to have high familiarity with the corresponding top level category.There were 43 (out of 64) cases where participants were shown a BLC and named it correctly. In half (58%) of these cases the participants had high familiarity with the top level instrument category of the BLC.
Based on the above heuristics, we propose a probing algorithm based on BLC.
Domain - a linked data graph G = (V , E) whereV = {v1,v2,...,vn} Set of Images I = {i1,i2,...,in} and a function image :V → I assigning an image i to each vertex v Set of Basic level concepts: B = {b1,b2,...,bk} User diagnosis is a mapping that overlays G’s vertices with a familiarity level – none, low, medium, high. familiarity :V → W whereV = {v1,v2,...,vn} and W = {none,low, medium, high} For every vertex from G, we define the following functions which are implemented with simple inferences using hierarchical relationships: - P(v) – returns all parent concepts for v - C(v) – returns all children concepts (including the leafs) for v - L(v) – returns the leaves (instances) for v - T(v) – top level categories for v
User model U = V ×W where d :V → W
Processing
//initialization
for all v ∈V do d(v) = none
for all b ∈ B do //BLC naming
show image(b) and ask to name it
if user_answer≠ b do //cannot name BLC
familiarity(b) = none
for all t ∈T (b) familiarity(t) = low
else do //names BLC
familiarity(b) = medium
for all t ∈T (v) familiarity(t) = medium
In this work, we examine the advantage of using basic level concepts to detect user
familiarity in a linked data graph. The user study identifies the BLCs in a Music
domain in a free-naming task and illustrates how these concepts can be utilized to detect
the user familiarity with a subset of entities from the LDG. Obviously, these findings
can only be applied if it is possible to automatically detect BLCs from a LDG.
Following the Cognitive Science definition of BLCs–domain concepts that carry the most
information, possess the highest category cue validity, and are, thus, the most
differentiated from one another are highly informative [
Our immediate future work is to tune the BLC algorithms and explore various
fusion methods to improve the precision results. We will then be able to implement the
probing algorithm and utilise it in developing the nudging strategies derived in [