=Paper=
{{Paper
|id=Vol-3294/long6
|storemode=property
|title=Meta-Intents in Conversational Recommender Systems
|pdfUrl=https://ceur-ws.org/Vol-3294/long6.pdf
|volume=Vol-3294
|authors=Yuan Ma,Tim Donkers,Timm Kleemann,Jürgen Ziegler
|dblpUrl=https://dblp.org/rec/conf/recsys/MaDK022
}}
==Meta-Intents in Conversational Recommender Systems==
https://ceur-ws.org/Vol-3294/long6.pdf
Meta-Intents in Conversational Recommender Systems
Yuan Ma, Tim Donkers, Timm Kleemann and Jürgen Ziegler
University of Duisburg-Essen, Forsthausweg 2, Duisburg, 47057, Germany
Abstract
We present a study investigating the psychological characteristics of users and their conversation-related preferences in a
conversational recommender system (CRS). We collected data from 260 participants on Prolific, using questionnaire responses
concerning decision-making style, conversation-related feature preferences in the smartphone domain, and a set of meta-
intents, a concept we propose to represent high-level user preferences related to the interaction and decision-making in
CRS. We investigated the relationship between users’ decision-making style, meta-intents and feature preferences through
Structural Equation Modeling. We find that decision-making style has a significant influence on meta-intents as well as
on feature preferences, however, meta-intents do not have a mediating effect between these two factors, indicating that
meta-intents are independent of item feature preferences and may thus be generalizable, domain-independent concepts. Our
results provide evidence that the proposed meta-intents are linked to the general decision-making style of a user and can
thus be instrumental in translating general decision-making factors into more concrete design guidance for CRS and their
potential personalization. As meta-intents seem to be domain-independent factors, we assume meta-intents do not affect
users’ various interests in concrete product features and mainly reflect users’ general decision-support needs and interaction
preferences in CRS.
Keywords
Decision-making style, Meta intents, Conversational UI design, Conversational recommender systems
1. Introduction CRS, in which the system guides the dialog and the user
answers, the difference is that SAUP requires the user to
Conversational recommender systems (CRS[1]) have answer the question directly, while SAUE allows the user
been gaining increased attention in research and indus- to not answer the question directly, instead providing
try in recent years [2, 3]. Generally, conversational tech- another preference or chit-chat. There are also lots of
niques can provide users with strong guidance to achieve challenges for system-initiated CRS, e.g. questions from
their goals combined with a high level of flexibility in CRS need to be formulated at an appropriate level of ab-
expressing their needs. Jannach et al. [4] distinguish straction, for example, asking either about the intended
between natural language-based, form-based, and cri- use of the product or about some specific technical fea-
tiquing approaches. Due to the advances in NLP tech- tures. Question relevant GUI widgets need to show a
niques in recent years, natural language-based CRS have suitable number of options. Dialog flow should follow
become subject of extensive research. Fu et al. [5] summa- the user’s likely mental decision process, providing suffi-
rized NLP-based CRS into 3 paradigms: System is Active, cient flexibility without becoming overly complex, and
User is Passive (SAUP), System is Active, User Engages recommendations should be presented in appropriate
(SAUE), System is Active, User is Active (SAUA). SAUA numbers and with an appropriate level of detail.
is a user-initiated paradigm of CRS, which provides the To address these challenges for CRS, a thorough under-
user with the greatest degree of flexibility, allowing the standing of user needs and their decision-making style is
user and the system to lead the conversation, and be able needed. Little research, however, has investigated the in-
to give appropriate feedback to the user’s questions. The fluence of psychological user characteristics and general,
appropriate feedback means answering user questions dialog-related preferences, in the context of CRS thus far
in a user-friendly style, but different users should have [6]. In this paper, we explore psychological characteris-
different preferences, e.g. preferring long sentences or tics of CRS users under two different objectives. First, we
short sentences, involving more technical details or not. aim at obtaining a deeper understanding of psychological
These are challenges for user-initiated CRS. characteristics of CRS users, based on responses from
SAUE and SAUP are system-initiated paradigms of questionnaire instruments. Here, we distinguish between
stable individual traits including personality factors [7]
4th Edition of Knowledge-aware and Conversational Recommender Sys- and decision-making style [8], and, second, task-oriented
tems (KaRS) Workshop @ RecSys 2022, September 18–23 2023, Seattle, characteristics that represent general user preferences
WA, USA.
Envelope-Open yuan.ma@uni-due.de (Y. Ma); tim.donkers@uni-due.de
when interacting with a CRS, such as obtaining detailed
(T. Donkers); timm.kleemann@uni-due.de (T. Kleemann); information about items or comparing products. We call
juergen.ziegler@uni-due.de (J. Ziegler) the latter characteristics meta-intentions (or meta-intents
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0). for short) since they describe user goals that are more
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
general and high-level than the search goals typically ex- The disadvantage is that the question sequences/paths
tracted through intent detection methods in CRS. Psycho- are hand-crafted, not enough freedom, and a lower-level
logical factors are an important resource underutilized by personalization.
current CRS. We therefore propose an initial framework Critiquing-based CRS will first recommend options
that includes the psychological factors in CRS design, and and then elicit users’ feedback in the form of critiques
also describes our core research target, the meta-intents [16], It help users to efficiently refine their preference
in it (Figure 1). As a second objective, we investigate by providing more options, but on the other hand, it can
the relations between psychological user characteristics be frustrating for novice users because they are over-
and users’ interests in product topics in a conversational whelmed by so many parameter options without really
scenario. understanding what those parameters really mean. Ma
In this paper, we describe a study analyzing these ques- et al. [17] proposed mixing language, GUI elements to im-
tions and present its results. Our contribution is: we prove user experience in CRS. However, it poses a greater
provide insights about CRS users’ decision-making style challenge to the design of the CRS as well.
(rational vs. intuitive) and its influence on the different Currently, very limited research has as yet studied
meta-intents that we propose, as well as about the re- users’ psychological influence in CRS and related design
lation between decision style, meta-intents, and users’ questions. For example, Papenmeier et al. [18] investi-
interest in product-specific features/topics. For this pur- gated human advisory dialogs, identifying some recur-
pose we present an analysis using Structural Equation ring strategies such as funneling to successively narrow
Modeling. down the space of potential items. Kleemann et al. [19]
investigated user behavior and personal characteristics
when using a advisor in combination with other deci-
2. Related work sion aids, and studied various supporting methods’ (chat-
bot, advisor, filter, recommendation) popularity, utiliza-
Conversational Recommender Systems (CRS) have be-
tion, and switching rate between each other [20]. Atas
come a rapidly growing and popular research area be-
et al. [21] summarize that preferences are determined
cause they provide a flexible, human-like multi-turn dia-
and adapted is influenced by various factors such as per-
log for preference elicitation, which is essential for gen-
sonality traits, emotional states, and cognitive biases.
erating personalized recommendations [9]. Jannach et al.
To provide design guidance for CRS and to potentially
[4] distinguish three types of CRS, differing in the style
adapt them to the individual user, a deeper understanding
and structure of the interaction used: natural language-
of the psychological factors influencing users’ decision
based, form-based, and critiquing-based.
making and interaction behavior in CRS is required. For
NLP-based CRS have received considerable interest
recommender systems in general, the influence of psycho-
recently due to the advancements in natural language
logical characteristics on users’ preference construction
processing. They typically use a question-answer format
and decision making has been shown repeatedly [21]. Lex
Zhang et al. [10]. As the mainstream, it has developed
et al. [22] distinguish between factors related to cognition,
vigorously in recent years, e.g. Sun and Zhang [1] import
personality, and emotion. The influence of psychological
the end-to-end reinforcement learning model to CRS,
characteristics such as the Big Five personality factors
Zhang et al. [11] combine contextual bandits method
(e.g. [23, 24]), Need for Cognition [25], or cognitive bi-
to improve preference elicitation and recommendation
ases [26] has been studied in several works. However,
performance. Zhou et al. [12] utilize knowledge graph-
these studies mostly aim at better understanding user
based as an external knowledge to enhance CRS, Li et al.
preferences with respect to the recommended items and
[13] unifying items and features in same arm space, use
at improving their accuracy. In contrast, the relationship
bandits method to facilitate cold-start problem in CRS,
between psychological factors and the design of advisory
Zhou et al. [14] extract topic threads from their dataset
dialogs in CRS remains an underexplored area. Especially
and leverage it to increase utility and user acceptability
theories related to human decision-making styles appear
of CRS. NLP-based CRS provides the greatest freedom,
to be promising points of departure for studying this
allowing users to express freely, while misunderstandings
relation. The distinction between rational and intuitive
can also usually happen and lead to user frustration.
decision-making styles [27] or cognitive styles such as
Form-based CRS present questions and answer in
the need for cognition may influence users’ assessment
a GUI style, leading users through a predefined dialog
of CRS. More domain-specific theories such as Shopping
structure. This type of CRS has many advantages as
Orientation [28, 29], distinguishing between task-focused
they provide guidance to the users, avoid errors, and
and experiential shopping are also of interest. However,
can incorporate domain knowledge. Especially usage-
none of these approaches has yet been applied to CRS.
related questions are important for users who have only
User goals and preferences when interacting with a
limited knowledge about technical item properties [15].
CRS may be located on different levels of abstraction.
Figure 1: CRS framework that combines psychological factors (Decision-making style and meta-intents ) and conventional
CRS
Low-level preferences refer to concrete properties of the able product quickly is more important than ex-
desired item (often called intents in CRS, specifically Add ploring all options.)
Details [30]). Jameson et al. [31] suggest high-level fac-
tors (such as economy and safety) but these factors are • Diversity orientation (When shopping online,
related to the product itself, not to the way users prefer I tend to explore a diverse range of products that
to interact with a CRS. On a more abstract level, meta- might interest me.)
level preferences that relate to the conversation and type • Goal focus (I usually have a clear idea of what
of questions in a CRS have, to our knowledge, not been I want before visiting an online shop. I often
studied yet. only make up my mind once I see the available
choices.)
3. User characteristics and • Openness for guidance (I appreciate it if a shop
meta-intents recommends products I might like.)
To investigate differences in CRS users’ psychological • Interest in detail (I usually gather as much in-
properties, we hypothesized that decision-making style formation as possible about products that I want
might influence users’ usage and interaction in CPA. Ac- to buy. I am interested in detailed information
cordingly, we applied instruments to measure these prop- about products.)
erties, using the Decision Styles Scale (DSS) [27] for dis- • Human-like (I would like a human-like conver-
tinguishing rational and intuitive decision-making. sation with an advisor system such as a chatbot.)
While general decision-making styles, e.g. rational and
intuitive, apply to arbitrary decision contexts, we also • Comparison orientation (Comparing the fea-
aimed at capturing users’ preferences at a more specific, tures of different candidate products is important
yet still abstract level. These meta-intents should bridge for me.)
the gap between item-level intents and general decision-
making style, and should also relate to the design and • Scope of choice (When the system recommends
question-asking style in CRS. They might also be rele- products, I rather like to see a longer list rather
vant for more general recommendation scenarios. We than a short one.)
postulated the following set of meta-intents (with sam-
The CRS framework we propose incorporates psycho-
ple questionnaire items in parentheses), partly related
logical level factors and preferences that relate to the
to general usage factors such as efficiency, effectiveness,
items and their properties (topic preferences and value
and user guidance. We see this list as a first step towards
preferences) as shown in Figure 1. We first introduce
defining factors relevant for users’ decision-making pro-
what each part represents. The decision style shown on
cess in CRS which is neither complete nor final.
the left side as main characteristic factors that might influ-
• Efficiency orientation (For me, finding a suit- ence meta-intents and users’ feature preferences which
are in the middle part of the figure. Here we use the term
topics instead of features to emphasize that in the CRS,
the user’s preference is not only about product features
but also the user experience, usage, and other higher ab-
stract level topics. For example, asking user questions
about the resolution of the main camera (feature level),
or taking good pictures (usage level), or the quality of the
main camera (assessment level) all belong to topic prefer-
ence elicitation. Users’ interest in product features/topics
is abbreviated as topic preference below.
The right part of the Figure 1 refers to a conventional Figure 2: Our structural equation model that includes 3 parts,
CRS model which can be, for example, CRM model [1] stable psychological traits (decision-making style), the pro-
or EAR model [32] (which are 2 popular CRS models posed psychological traits (meta-intents) and topic preference
that include conversation function and recommendation (smartphone domain).
function and utilized deep neural networks). The top ele-
ment of the middle part is meta-intents which solves the
problem how to ask and respond and can be used to guide
4.1. Method
the interaction style of CRS. Topic preference is related
to the interactive content (ask which topic) and can be We first presented participants with a scenario in which
used to improve the preference elicitation process. Value they were supposed to buy a new smartphone, and then
preference which stands for the users’ personalized pref- started our questionnaire. The smartphone domain was
erence value for one specific feature/topic. The middle chosen because it requires a sufficiently complex deci-
part, topic preference and value preference are also known sion process, involving a variety of decision criteria. For
as intents detection which is an active area of research in most people it is also a well-known, real-life task that
NLP-based CRS. We decouple intents detection into two requires understanding the product features at least to
elements here for studying the impact of decision-making a certain extent. Furthermore, it has a large number of
style on it. Our framework proposes that psychological feature options. To measure psychological characteris-
characteristics can be treated as additional knowledge tics, we applied the existing Decision Style Scale (DSS)
to improve CRS design, so in this paper, we apply SEM questionnaire [27] as well as a self-developed question-
to analyze how does psychological characteristic impact naire on meta-intents (Section 3), both with 5-point Likert
these factors and our research can be boiled down to two scales. To measure topic preference, we collected a total
questions: of 27 topics in the smartphone domain, including 4 dif-
ferent levels: usage-level, general-level, technical-level,
• Does decision-making style significantly influ- and professional-level, as shown in Table 1. There were
ence users’ meta-intents and topic preference? short descriptions for some less well-known topics in our
questionnaire, e.g. network sensitivity (signal strength,
• Do meta-intents have a mediation effect between
how easy is it to connect to a mobile network). We asked
decision-making style and topic preference?
participants to rate each topic on a 5-point Likert scale ac-
cording to their interest (1: don’t care, 5: very interested
4. Study in), along with an unknown option, in case participants
did not understand the topic’s meaning.
To investigate CRS users’ psychological characteristics,
both at the level of decision-making style and meta- 4.2. Participants
intentions, as well as possible relations with their topic
1
preference, we conducted an online survey in which par- We recruited 278 participants using Prolific , a tool com-
ticipants were presented a scenario involving the pur- monly used for academic surveys [33], of whom 275 fin-
chase of a new smartphone and answered questionnaires ished the study. In our analysis, we only considered par-
concerning their product-related preferences as well as ticipants who passed 3 inner attention test questions (e.g.
their psychological characteristics. We hypothesized that , It’s an attention test, please select strongly agree), leav-
general traits (decision-making style) significantly influ- ing us with 260 participants. 143 of the 260 participants
ence meta-intents and topic preference. We also assumed were female. Their age ranged from 19 to 75 (M = 38.42,
that meta-intents might have a mediating effect between SD = 12.60). We pre-selected Prolific users based on the
decision-making style and users’ topic preference . following criteria to maximize quality: (1) participants
should be fluent in English; (2) their success rate should
1
https://www.prolific.co
Table 1
The collection of total 27 user-interested topics in smartphone domain and 4 categories.
smartphone taking photo taking watching videos
Usage multi apps
game and video selfies and documents
price network
General brand color size weight robustness voice quality
performance ratio sensitivity
latest headphone good front number of battery life and biometric
Technical 5G dual SIM
technology jack 35 camera main cameras charging speed unlock
screen main camera operating
Professional RAM ROM localization CPU and GPU
resolution resolution system
be greater than 95 %. The average duration of the sur-
vey was 5.56 minutes (SD = 2.48) and each participant
received compensation of 0.75£ if they successfully com-
pleted the survey.
5. Results
We applied Structural Equation Modeling (SEM) to our
dataset for estimating and testing the causal effects
of three main variables: Decision-making style, meta-
intents and topic preference. Since DSS is a well-
established, validated questionnaire and meta-intents are
captured with single-items (each factor has only one ques-
tion), we could directly incorporate both in our proposed
model (see Figure 2). Concerning topic preference, on
the other hand, a total of 27 topics (items) with assumed Figure 3: Unknown number of user-interested topics. X-
commonalities have been asked, e.g. taking photo video axis stands for the unknown number, Y-axis stands for user-
(usage), number of main cameras (technical) and main interested topics. Orange bar indicates the unknown number
camera resolution (professional) should involve correlated of 4 categories.
rating patterns, hence presumably loading onto the same
factor. Therefore, we do not treat all 27 topics as sin-
gle independent variables but apply Exploratory Factor mine the number of factors. We ran the EFA recursively
Analysis (EFA) to extract conjoint latent variables that such that after the first epoch, resulting in dropping one
can subsequently be fed into our proposed SEM model. item, a second run finally met the requirements. We
filtered out seven topics in total: screen resolution (fac-
5.1. EFA on topic preference tor loadings < 0.4), price performance ratio (single item
factor), biometric unlock (single item factor), multi apps
The scores for topic preference are derived from a set of (factor loadings < 0.4), headset jack 35 (single item factor),
260 valid participants’ ratings of 27 smartphone topics. take photo and video and smartphone game (Cronbach’s 𝛼
63 of them tagged at least one topic as unknown (see < 0.6). Finally, we extracted six factors from 22 topics. We
Figure 3). Thereby, we found that the more technical the name these factors according to the topics they represent:
topic, the fewer people could grasp its meaning. Finally, camera, reliability, novelty, design, memory storage, and
only 197 participants’ data could be used for the EFA technical. The cumulative variance of 6 factors is 66.35 %,
analysis. all of them having a factor loading over 0.4, commonality
First, we performed prerequisite tests for EFA, the over 0.49, Cronbach’s 𝛼 over 0.60. Details are shown in
Kaiser-Meyer-Olkin (KMO) value is .796 (> 0.7) and Table 2.
Bartlett’s test is significant (< .001), which both indi-
cate that our data meets the requirements for performing
EFA. Next, we used Principal Component Analysis (PCA)
to extract factors, with Varimax rotation and Kaiser Nor-
malization, taking eigenvalue > 1 as the threshold to deter-
Table 2
Final EFA results of total 27 topics (df =197). The first column represents the kept topics in the smartphone domain and
Cronbach’s 𝛼 values of factors. The first column represents communities of topics and the founded latent factors. The bold
font indicates the values are greater than 0.5.
Factors
memory
Topics Commonalities camera reliability novelty design storage technical
taking photo and video .74 .85 -.01 .01 -.06 .10 .04
good front camera .74 .79 .23 -.04 .02 .08 .07
main camera resolution .58 .75 .10 .04 .03 .32 -.03
taking selfies .49 .71 .13 .11 .20 .01 .03
number of main cameras .55 .64 .06 .33 .27 .08 .16
network sensitivity .66 .17 .76 -.02 -.09 .21 .07
robustness .61 .14 .69 .10 .19 -.04 .13
voice quality .77 .05 .64 .30 .05 -.10 .17
battery life and charging speed .65 .10 .60 -.23 .11 .35 -.18
5G .56 .08 .11 .75 .06 .15 .08
dual SIM .65 .03 -.06 .71 .19 .13 .03
latest technology .56 .26 .28 .56 .30 -.01 .04
color .68 .05 .05 .17 .85 .04 .04
brand .69 .21 -.02 .06 .68 -.06 .27
size and weight .65 .01 .23 .22 .64 .26 -.01
ROM .67 .24 .15 .14 .09 .83 .03
RAM .71 .15 -.01 .22 .02 .77 .23
operating system .76 .02 .02 -.08 .16 .08 .85
localization .72 .26 .27 .35 .12 .09 .60
CPU and GPU .81 .02 .25 .47 .01 .22 .58
Cronbach’s 𝛼 .83 .60 .66 .64 .80 .68
5.2. SEM on decision-making style, 5.2.1. Part A: decision-making style and
meta-intents, topic preference meta-intents
Finally, based on our data, we constructed a SEM, which Part A focuses on the influence of decision-making style
contains decision-making style, meta-intents and topic on meta-intents. After removing non-significant effects,
preference, as shown in Figure 4. Decision-making style five of eight meta-intents factors remain. We found the
(ovals) and topic preference (ovals) are estimated from factor rationality having significant influences on five
several directly measurable questionnaire items. In order meta-intents factors with the greatest impact on interest
to display the relationships between our main factors in details (0.61) and comparison oriented (0.47). Besides
as clearly as possible, we leave out the factor loadings these relationships, rationality also has positive influ-
of concrete questionnaire items. Decision-making style ences on diversity orientation (0.22) and scope of choice
and meta-intents are latent variables in this framework, (0.26), but a negative influence on efficiency orientation
however, since the meta-intents are measured by a single (-0.29). In contrast, for the intuitiveness factor only a
question, we use rectangles to represent them. The ar- single significant (positive) effect on efficiency orientation
rows indicate significant influences, with the value above (0.34) could be identified.
depicting standardized regression coefficients, while non-
significant connections have been removed for clarity. 5.2.2. Part B: decision-making style and topic
As the entire SEM is quite large, in order to analyze it preference
methodically, we split it into two parts with the green
Part B focuses on the influence of decision-making style
rounded rectangle representing Part A, and the yellow
on topic preference. After cleaning the non-significant
rounded rectangle representing Part B respectively.
effects, four of six topic preference factors remain. We
found the rationality has positive and significant influ-
ences on camera (0.27), memory storage (0.31), and techni-
cal (0.26). The intuitiveness has positive and significant
influences on camera (0.46), reliability (0.34), memory
storage (0.21), and technical (0.29). While decision style 6. Discussion
showed opposite effects at the MI level (for efficiency-
orientation), here they did not show this pattern, only 6.1. Part A: decision-making style and
differing in the impact coefficient. The biggest difference meta-intents
was observed for the camera factor, for which the intu-
itiveness has a larger standardized regression coefficient We found that rationality has more significant influences
(0.46) than the rationality (0.27). on meta-intents than intuitiveness, and that both have
From these results, we can answer the first research opposite effects on efficiency orientation. This implies
question posed earlier: decision-making style has a sig- that the more rational people are, the less they seem to
nificant influence on some meta-intents and topic prefer- care about efficiency. At the user interaction level in
ence factors. CRS, efficiency may be determined by interaction time as
well as the number of clicks and keystrokes needed for
Part A Interest in Diversity Efficiency Comparison Scope of
typing text. In personalized CRS design, this factor has a
details orientation orientation orientation choice
guiding role for the length of the dialogue, the amount of
0.61 0.22 -0.29
0.47
0.26
Camera
information displayed per output, and when to display
0.27 the recommended products. Rationality also has posi-
0.46
Rational
tive influence on diversity orientation, interest in details,
Reliability
0.31
0.34
comparison orientation, and scope of choice. Diversity ori-
0.34 entation indicates that the user would like to see a diverse
0.26
0.21
Memory
storage range of items in the recommendation list. Interest in de-
Intuitive
tails provides insights into how much content should
0.29
Part B
Technical be shown when displaying product features and other
information, such as customer comments. Comparison
Figure 4: Structural equation model including 3 parts, stable orientation suggests that users would like to see products,
psychological traits (Decision-making style), proposed psycho- their features and customer assessments side by side,
logical traits (meta-intents) and user interested topics (smart- e.g. in a comparison function, to take a decision. Scope
phone domain). of choice can inform us about choosing an appropriate
length of the recommendation list and probably also the
length of features lists shown for a product. In sum, these
5.2.3. Overall model findings provide some insights for the design of CRS with
respect to dialog structure, design of questions and an-
The overall model fit is shown in Table 3. The subsub- swers, and the presentation of recommendations. If data
section 5.2.1 and 5.2.2 claim that decision-making style on the user’s decision style were available, e.g. through
significantly impacts both meta-intents and topic prefer- classifying their interactive behavior, the findings can
ence, which meet the prerequisite for testing mediating also provide a basis for personalizing the CRS.
effects (meta-intents as mediator). However, we found
no significant influence of meta-intents on topic prefer-
ence. Applying Bootstrap testing (2000 iterations) for
6.2. Part B: decision-making style and
indirect effects (decision-making style → meta-intents topic preference
→ topic preference) yielded no significant indirect effect, Concerning Part B, we notice that the intuitiveness has
preventing further mediation analysis and answering larger standardized regression coefficients (0.46) on cam-
the second research question: meta-intents do not act era than the rationality (0.27) implying that intuitive
as mediators between decision-making style and topic people are more interested in camera functionality than
preference. This finding provides some indication that rational people. This gives us some pointers for CRS de-
meta-intents are independent of the specific product do- sign in this specific domain. When eliciting preferences
main, in this case smartphones. (or detecting intents), camera is a topic of interest to the
user with high intuition. From a more general point of
Table 3 view, the discrepancies between rational and intuitive de-
The overall fitness indices of the proposed structural equation cision makers suggest that the former are more focused
model. on the low-level technical specifics of a product domain
𝜒 2 /df GFI AGFI TLI NFI CFI RMSEA (such as the CPU which is installed in a smartphone),
evaluation while the latter are more attracted to information about
1< & <3 >0.8 >0.8 >0.9 >0.9 >0.9 <0.08 immediately experiential properties (such as the quality
standard
proposed SEM 1.809 .846 .794 .849 .768 .877 .064 of a shot photo).
Our findings provide insights into intent detection for
CRS in preference elicitation. Supposing data on users’ translating general decision-making factors into more
decision-making styles are available, the survey results concrete design guidance for CRS and their potential per-
can provide a basis for personalizing the preference elic- sonalization. At the same time, we also point out three
itation process, e.g. to help choosing which features to limitations of this experiment: 1. Meta-intents is a new
ask and the order in which they are asked. At the same concept for which we used only one or two questionnaire
time, we want to point out a limitation here. Unlike the items per intent. We plan to develop the instrument fur-
high abstract level of meta-intents which can be applied ther and validate the meta-intents with a larger number
to various fields, the findings here are based on a specific of questions. 2. The domain of this experiment is limited
field (smartphone). Still, we provide an idea for utilizing to smartphones, and comparative experiments in several
decision-making knowledge to enhance the preference fields will be necessary in the future. 3. The integration
elicitation process of CRS in a specific domain. of meta-intents into specific CRS models and the real
impact of MI on user interaction needs to be explored in
6.3. Overall model future work.
After applying SEM to the overall model, we found that
the meta-intents factors do not significantly impact topic References
preference, which means our proposed meta-intents seem
to be independent of product features and topics, and [1] Y. Sun, Y. Zhang, Conversational recommender
do not have mediating effects between decision-making system, in: The 41st International ACM SIGIR Con-
style and topic preference. Meta-intents appear to be in- ference on Research & Development in Information
dependent of our concrete product domain, but whether Retrieval, SIGIR ’18, Association for Computing
it is truly domain-independent remains to be verified by Machinery, New York, NY, USA, 2018, p. 235–244.
multi-domain research. doi:1 0 . 1 1 4 5 / 3 2 0 9 9 7 8 . 3 2 1 0 0 0 2 .
We believe our results provide some initial valuable [2] Z. Fu, Y. Xian, Y. Zhang, Y. Zhang, WSDM 2021
insights that can help be better designe and possibly Tutorial on conversational recommendation sys-
personalize CRS. In particular, we would like to point tems, WSDM ’21, Association for Computing Ma-
out that experiences with a CRS should be viewed from chinery, New York, NY, USA, 2021, p. 1134–1136.
multiple perspectives. Differences in individual decision doi:1 0 . 1 1 4 5 / 3 4 3 7 9 6 3 . 3 4 4 1 6 6 1 .
behavior seem to be related not only to attitudes toward [3] C. Gao, W. Lei, X. He, M. de Rijke, T.-S.
specific product features, but also to expectations of the Chua, Advances and challenges in conversa-
interaction process as a whole. When designing a CRS, tional recommender systems: A survey, AI Open
therefore, consideration should be given not only to per- 2 (2021) 100–126. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j .
aiopen.2021.06.002.
sonalizing the recommended products, but also to adapt-
ing the agent’s mode of communication. If both aspects [4] D. Jannach, A. Manzoor, W. Cai, L. Chen, A survey
are taken into account in an appropriate manner, it can be on conversational recommender systems, ACM
assumed that positive transfer effects could arise which, Comput. Surv. 54 (2021). doi:1 0 . 1 1 4 5 / 3 4 5 3 1 5 4 .
taken together, enrich the user’s overall experience. [5] Z. Fu, Y. Xian, Y. Zhang, Y. Zhang, Tuto-
rial on conversational recommendation systems,
in: Fourteenth ACM Conference on Recom-
7. Conclusion mender Systems, RecSys ’20, Association for Com-
puting Machinery, New York, NY, USA, 2020,
In this paper, we propose a set of meta-intents factors, p. 751–753. URL: https://doi.org/10.1145/3383313.
which can be seen as a bridge between classic psychologi- 3411548. doi:1 0 . 1 1 4 5 / 3 3 8 3 3 1 3 . 3 4 1 1 5 4 8 .
cal factors (decision-making style) and item-level intents, [6] B. Xiao, I. Benbasat, E-commerce product recom-
and which can be used as indicators of personalized UI mendation agents: Use, characteristics, and impact,
design for CRS. We also propose a CRS framework that MIS Q. 31 (2007) 137–209.
incorporates decision-making style, meta-intents factors, [7] S. D. Gosling, P. J. Rentfrow, W. B. Swann, A
and conventional CRS models. This work uses SEM to very brief measure of the big-five personality do-
validate the significant influence of decision-making style mains, Journal of Research in Personality 37 (2003)
on meta-intents and topic preference. We also found that 504–528. doi:1 0 . 1 0 1 6 / S 0 0 9 2 - 6 5 6 6 ( 0 3 ) 0 0 0 4 6 - 1 .
meta-intents seem to be domain-independent, dialog- [8] K. Hamilton, S.-I. Shih, S. Mohammed, The devel-
specific factors that do not have a mediating effect on opment and validation of the rational and intuitive
topic preference. Our results provide evidence that the decision styles scale, Journal of Personality As-
proposed meta-intents are linked to the general decision- sessment 98 (2016) 523–535. doi:1 0 . 1 0 8 0 / 0 0 2 2 3 8 9 1 .
making style of a user and can thus be instrumental in 2015.1132426.
[9] B. Lika, K. Kolomvatsos, S. Hadjiefthymiades, Fac- eling user interaction at the convergence of filter-
ing the cold start problem in recommender sys- ing mechanisms, recommender algorithms and ad-
tems, Expert Syst. Appl. 41 (2014) 2065–2073. visory components, in: Mensch Und Computer
doi:1 0 . 1 0 1 6 / j . e s w a . 2 0 1 3 . 0 9 . 0 0 5 . 2021, MuC ’21, Association for Computing Ma-
[10] Y. Zhang, X. Chen, Q. Ai, L. Yang, W. B. Croft, To- chinery, New York, NY, USA, 2021, p. 499–511.
wards conversational search and recommendation: doi:1 0 . 1 1 4 5 / 3 4 7 3 8 5 6 . 3 4 7 3 8 5 9 .
System ask, user respond, in: Proceedings of the [20] T. Kleemann, B. Loepp, J. Ziegler, Towards multi-
27th ACM International Conference on Information method support for product search and recom-
and Knowledge Management, CIKM ’18, Associa- mending, in: Adjunct Proceedings of the 30th
tion for Computing Machinery, New York, NY, USA, ACM Conference on User Modeling, Adaptation
2018, p. 177–186. doi:1 0 . 1 1 4 5 / 3 2 6 9 2 0 6 . 3 2 7 1 7 7 6 . and Personalization, 2022, pp. 74–79. doi:h t t p s :
[11] X. Zhang, H. Xie, H. Li, J. C. S. Lui, Toward building //doi.org/10.1145/3511047.3536408.
conversational recommender systems: A contex- [21] M. Atas, A. Felfernig, S. Polat-Erdeniz, A. Popescu,
tual bandit approach, CoRR abs/1906.01219 (2019). T. N. T. Tran, M. Uta, Towards psychology-aware
arXiv:1906.01219. preference construction in recommender systems:
[12] K. Zhou, W. X. Zhao, S. Bian, Y. Zhou, J.-R. Wen, Overview and research issues, Journal of Intelligent
J. Yu, Improving conversational recommender sys- Information Systems (2021) 1–23.
tems via knowledge graph based semantic fusion, [22] E. Lex, D. Kowald, P. Seitlinger, T. N. T. Tran,
in: Proceedings of the 26th ACM SIGKDD Interna- A. Felfernig, M. Schedl, Psychology-informed rec-
tional Conference on Knowledge Discovery & Data ommender systems, Foundations and Trends® in In-
Mining, 2020, pp. 1006–1014. doi:1 0 . 1 1 4 5 / 3 3 9 4 4 8 6 . formation Retrieval 15 (2021) 134–242. doi:1 0 . 1 5 6 1 /
3403143. 1500000090.
[13] S. Li, W. Lei, Q. Wu, X. He, P. Jiang, T.-S. Chua, [23] M. Tkalcic, L. Chen, Personality and Recommender
Seamlessly unifying attributes and items: Conver- Systems, Springer, Boston, MA, 2015, pp. 715–739.
sational recommendation for cold-start users, ACM doi:1 0 . 1 0 0 7 / 9 7 8 - 1 - 4 8 9 9 - 7 6 3 7 - 6 _ 2 1 .
Trans. Inf. Syst. 39 (2021). doi:1 0 . 1 1 4 5 / 3 4 4 6 4 2 7 . [24] T. Z. Gizaw, H. Dong Jun, A. Oad, Solving cold-start
[14] K. Zhou, Y. Zhou, W. X. Zhao, X. Wang, J.-R. Wen, problem by combining personality traits and demo-
Towards topic-guided conversational recommender graphic attributes in a user based recommender sys-
system, in: Proceedings of the 28th International tem, International Journal of Advanced Research
Conference on Computational Linguistics, Interna- in Computer Science and Software Engineering 7
tional Committee on Computational Linguistics, (2017).
Barcelona, Spain (Online), 2020, pp. 4128–4139. [25] M. Millecamp, N. N. Htun, Y. Jin, K. Verbert, Con-
doi:1 0 . 1 8 6 5 3 / v 1 / 2 0 2 0 . c o l i n g - m a i n . 3 6 5 . trolling spotify recommendations: Effects of per-
[15] I. Kostric, K. Balog, F. Radlinski, Soliciting user sonal characteristics on music recommender user
preferences in conversational recommender sys- interfaces, in: Proceedings of the 26th Confer-
tems via usage-related questions, in: 15th ACM ence on User Modeling, Adaptation and Person-
Conference on Recommender Systems, RecSys ’21, alization, UMAP ’18, Association for Computing
2021, pp. 724–729. Machinery, New York, NY, USA, 2018, p. 101–109.
[16] L. Chen, P. Pu, Critiquing-based recommenders: doi:1 0 . 1 1 4 5 / 3 2 0 9 2 1 9 . 3 2 0 9 2 2 3 .
survey and emerging trends, User Modeling and [26] J. Zhang, Anchoring effects of recommender sys-
User-Adapted Interaction 22 (2012) 125–150. tems, in: Proceedings of the 5th ACM Conference
[17] Y. Ma, T. Kleemann, J. Ziegler, Mixed-modality in- on Recommender Systems, RecSys ’11, Association
teraction in conversational recommender systems, for Computing Machinery, New York, NY, USA,
in: Interfaces and Human Decision Making for Rec- 2011, p. 375–378. doi:1 0 . 1 1 4 5 / 2 0 4 3 9 3 2 . 2 0 4 4 0 1 0 .
ommender Systems 2021: Proceedings of the 8th [27] K. Hamilton, S.-I. Shih, S. Mohammed, The devel-
Joint Workshop on Interfaces and Human Decision opment and validation of the rational and intuitive
Making for Recommender Systems, 2021, pp. 21–37. decision styles scale, Journal of Personality As-
[18] A. Papenmeier, A. Frummet, D. Kern, “Mhm...” – sessment 98 (2016) 523–535. doi:1 0 . 1 0 8 0 / 0 0 2 2 3 8 9 1 .
conversational strategies for product search assis- 2015.1132426.
tants, in: ACM SIGIR Conference on Human Infor- [28] O. B. Büttner, A. Florack, A. S. Göritz, Shopping
mation Interaction and Retrieval, CHIIR ’22, Asso- orientation as a stable consumer disposition and
ciation for Computing Machinery, New York, NY, its influence on consumers’ evaluations of retailer
USA, 2022, p. 36–46. doi:1 0 . 1 1 4 5 / 3 4 9 8 3 6 6 . 3 5 0 5 8 0 9 . communication, European Journal of Marketing
[19] T. Kleemann, M. Wagner, B. Loepp, J. Ziegler, Mod- (2014).
[29] O. B. Büttner, A. Florack, A. S. Göritz, How shop- Springer, Boston, MA, 2015, pp. 611–648. doi:1 0 .
ping orientation influences the effectiveness of 1007/978- 1- 4899- 7637- 6_18.
monetary and nonmonetary promotions, European [32] W. Lei, X. He, Y. Miao, Q. Wu, R. Hong, M.-Y. Kan,
Journal of Marketing (2015). T.-S. Chua, Estimation-action-reflection: Towards
[30] W. Cai, L. Chen, Predicting user intents and sat- deep interaction between conversational and rec-
isfaction with dialogue-based conversational rec- ommender systems, in: Proceedings of the 13th
ommendations, in: Proceedings of the 28th ACM International Conference on Web Search and Data
Conference on User Modeling, Adaptation and Per- Mining, 2020, pp. 304–312. doi:1 0 . 1 1 4 5 / 3 3 3 6 1 9 1 .
sonalization, UMAP ’20, Association for Comput- 3371769.
ing Machinery, New York, NY, USA, 2020, p. 33–42. [33] E. Peer, L. Brandimarte, S. Samat, A. Acquisti, Be-
doi:1 0 . 1 1 4 5 / 3 3 4 0 6 3 1 . 3 3 9 4 8 5 6 . yond the turk: Alternative platforms for crowd-
[31] A. Jameson, M. C. Willemsen, A. Felfernig, M. d. sourcing behavioral research, Journal of Experimen-
Gemmis, P. Lops, G. Semeraro, L. Chen, Hu- tal Social Psychology 70 (2017) 153–163. doi:h t t p s :
man decision making and recommender sys- //doi.org/10.1016/j.jesp.2017.01.006.
tems, in: Recommender Systems Handbook,