Up the Conversational Recom mender Systems' Biases

Abstract

The growing popularity of language models has sparked interest in conversational recommender systems (CRS) within both industry and research circles. However, concerns regarding biases in these systems have emerged. While individual components of CRS have been subject to bias studies, a literature gap remains in understanding specific biases unique to CRS and how these biases may be amplified or reduced when integrated into complex CRS models. In this paper, we provide a concise review of biases in CRS by surveying recent literature. We examine the presence of biases throughout the system's pipeline and consider the challenges that arise from combining multiple models. Our study investigates biases in classic recommender systems and their relevance to CRS. Moreover, we address specific biases in CRS, considering variations with and without natural language understanding capabilities, along with biases related to dialogue systems and language models. Through our findings, we highlight the necessity of adopting a holistic perspective when dealing with biases in complex CRS models.

1. Introduction

In recent years, conversational recommender systems (CRSs) [ 1, 2, 3, 4, 5, 6, 7, 8, 9 ] have garnered significant attention, reshaping personalized recommendations tion is notably supported by the successful integration of large language models (LLMs) like ChatGPT [ 10, 11 ], thereby driving the widespread deployment of LLMs in various applications. Such models have found substantial integration in prominent platforms, including Microsoft Bing1, which has invigorated dialogue search engines and recommender systems with unprecedented capabilities and paving the way for a new era of user engagement.

Although biases within recommender systems have

garnered significant attention [ 12, 13, 14, 15, 16, 17, 18 ], the examination of biases in conversational recommender systems remains a relatively unexplored domain [ 19 ].

Despite the conceptual alignment between conventional

recommender systems and their conversational counterparts, the latter exhibit heightened complexity and an increased potential for biases [ 20 ]. It is worth noting that some existing research has specifically addressed biases in conversational recommender systems (G. Farnadi) CEUR htp:/ceur-ws.org ISN1613-073

CEUR

Workshop Proceedings (CEUR-WS.org) https://blogs.bing.com/search/march_2023/ Confirmed-the-new-Bing-runs-on-OpenAI’s-GPT-4 [ 3, 4, 5, 6, 16, 20, 21 ]. However, no concerted efort has been undertaken to systematically categorize and analyze biases unique to conversational recommender systems, including how previously studied biases manifest within the conversational context. This paper aims to bridge intricate biases that characterize these complex systems.

This study conducts a systematic literature review to explore biases within conversational recommender systems. The methodology involves analyzing recent papers from top conferences in machine learning and information retrieval. Keyword searches within titles and abstracts identify relevant contributions that shed light on biases, including fairness concerns and bias amplification.

The paper begins by thoroughly investigating biases in classic recommender systems, establishing a foundation for understanding common biases, addressing similar notions to each of them, and most importantly, examining each bias in a conversational setting in Section 5. This step ensures a holistic grasp of biases across traditional recommendation systems. We then delve into each bias within CRSs. We start with focusing on CRSs without natural language understanding which uses basic dialogue systems for user interaction in Section 6. Furthermore, to capture diverse aspects and potential biases arising from natural language understanding, a dedicated literature review is conducted on the more complex CRSs that aim to understand natural language in Section 7. Following tial avenues for future research of our study. Finally, in

Section 9, we draw conclusions and wrap up the paper.

© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License that, in Section 8, we present the limitations and potenAttribution 4.0 International (CC BY 4.0).

2. Related Work

our knowledge, a dedicated survey paper focusing solely on the biases in conversational recommender systems is Although biases in general machine learning models, still missing from the literature. Consequently, this paper natural language processing (NLP), and recommender aims to embark on the initial journey towards gaining a systems (RS) domains have undergone extensive study, deeper understanding of the biases present in CRS and biases in conversational recommender systems (CRSs) their interactions, elucidating how biases from various have received limited attention. components within this complex system can either be

On a broad aspect, ML biases have been a subject of accentuated or alleviated. interest, with an example of the work by Mehrabi et al. [ 22 ] which provides a comprehensive survey on the challenges and approaches to mitigate biases in ML models. 3. Methodology Additionally, NLP biases have received attention, and for instance, Blodgett et al. [ 23 ] discussed the manifestation The primary objective of this study is to conduct a reand implications of biases in language models which can view of existing literature on potential biases in conbe seen as an important part of CRSs. versational recommender systems. To achieve this goal,

Moreover, several surveys have focused on biases and a systematic approach is adopted, focusing on papers fairness in recommender systems (RS) as a whole [12, published in prominent machine learning and informa24, 25, 13]. These surveys provide valuable insights into tion retrieval-related conferences from January 2019 to biases present in RS, which is a sub-module of CRSs, June 2023. These conferences include knowledge distherefore they can be used as valuable sources for specific covery and data mining (KDD), Special Interest Group investigation in a conversational setting. on Information Retrieval (SIGIR), ACM Conference on

Within the realm of CRS biases, recent research has Recommender Systems, User Modelling, Adaptation and shed light on diferent biases that can arise in such sys- Personalization (UMAP), International World Wide Web tems. For example, unintended biases are discussed by Conference (WWW), Neural Information Processing SysShen et al. [ 21 ]. Lin et al. [ 26 ] quantified biases in CRS tems (NeurIPS), International Conference on Machine by exploring the fairness of recommendations across dif- Learning (ICML). ferent user groups. Another study by Lin et al. [ 20 ] To identify relevant papers, we conducted keyword highlighted the importance of addressing biases in CRS searches within the titles or abstracts of the papers preto ensure equitable and inclusive recommendations. sented at these conferences. The selected keywords in

As for CRS surveys, Jannach et al. [ 19, 27, 9 ] have clude bias, dialog, conversation, chat, question, mitigat*, contributed significantly to the understanding of conver- recommend*, amplif*, fair*. These were chosen to align sational recommender systems. However, to the best of with key aspects within the scope of biases in conversational recommender systems and related notions such two types: Topic-guided CRSs, which utilize natural lanas dialogue systems and classic recommender systems. guage understanding, and Attribute-aware CRSs, which These keywords were thoughtfully chosen to reflect the rely on simplified input processing methods and attribute themes within the scope of biases in conversational rec- inquiries, as classified by Ren et al. [ 5 ]. ommender systems. Specifically, they align with key Another significant point to consider is that a tradiaspects such as recommender systems, dialogue systems, tional recommender system can also be illustrated by fairness considerations, and the amplification of biases. utilizing a subset of the modules featured in Figure 1. The chosen keywords collectively encompass a wide spec- By removing the Natural Language Module and the Diatrum of research that pertains to these facets, ensuring logue Management System (DMS) components, and by that our selection is representative of the relevant liter- establishing a direct communication path connecting the ature landscape. Subsequently, we filtered the results Recommender System and the User, we can achieve a once again based on their contributions to the trustwor- simplified structure. This approach facilitates a clearer thy intricacies of recommender systems, conversational understanding of the underlying nature and impact of recommender systems, or dialogue systems. biases discussed in Section 5.

4. Conversational Recommender

Systems

5. Biases in Classic Recommender

Systems

To explore biases within conversational recommender In this section, we examine biases in classic recommender systems (CRSs), a precise definition and understanding systems across three key aspects for each bias. For each of their model architecture are crucial. A Conversational bias, first, we define and discuss the bias, drawing from Recommender System (CRS) is intricate, with intercon- relevant literature for a strong foundation. Second, unnected components (Figure 1). Users interact with the der Similar Notions, we identify related notions that fall system to either provide answers (feedback) to the sys- within each bias. And third, in Through the CRS lens, tem’s questions (recommendations) or receive person- we analyze each bias in the context of conversational alized recommendations. Within the CRS, a language recommender systems, exploring specific works in this module assumes a pivotal role, consisting of two submod- domain. This approach ofers a wide-ranging perspecules: natural language understanding (NLU) and natural tive on biases in classic recommender systems within language generation (NLG). The NLU empowers the sys- conversational interactions. tem to understand user intentions, extracting insights from their input and prior profiling data. On the other 5.1. Popularity Bias hand, the NLG crafts coherent and contextually relevant responses in natural language. Alongside this, a recom- Popularity bias in recommender systems prioritizes popmender system undertakes user queries, harnessing avail- ular items, assuming they are more likely to interest able data to generate personalized recommendations. users. However, this bias can lead to a lack of diversity,

The dialogue management system (DMS) as the heart overshadowing lesser-known options. Recognizing and of the model, is in charge of orchestrating conversations mitigating popularity bias is crucial for developing recbetween the user and the system, ensuring logical flow ommender systems that ofer a wider range of choices and pertinence in each interaction. At each point of and promote serendipitous discovery [ 25, 28 ]. Besides the dialogue, The DMS decides when to finalize recom- investigating and mitigating this bias in a classic fashion, mendations or seek more information, guided by NLU there are some other works that focus on diferent setand the recommender system and by navigating the out- tings and notions. For example, Zhu et al. [ 17, 29 ] try puts through the NLG. Continual system enhancement is to solve the popularity bias in a dynamic environment achieved through the accumulation of conversation logs setting. Abdollahpouri et al. [ 30 ] see how diferent types into a database, leveraging these processed logs to train of users are afected by popularity bias and in the study both the natural language module and the recommender by Zhu et al. [ 17 ] the focus pertains to the challenge system. This iterative training process augments their of ranking a set of equally favored items based on their capabilities over time, refining user interactions. popularity.

It’s important to emphasize that the architecture out- Similar notions: lined above represents the general structure of a CRS.

However, certain CRSs do not incorporate natural lan- • Long-tail bias: The opposite of popularity guage understanding and generation into their opera- bias can also be the issue in a model, overtions; instead, they adopt simplified input processing recommending the niche items instead of already methods. This leads to the categorization of CRSs into established items [ 31 ].

• Filter bubbles, Echo chambers, Polarization: missing interactions [ 41, 42, 15 ].

These concepts constitute a significant area of • Exploitation Bias: It is investigated that users research within recommender systems and of- have a tendency to interact with or rate items ten arise as consequences of existing popularity that they personally prefer [ 43, 44 ]. bias. In the work by Michiels et al. [ 32 ], the filter bubble is characterized by a “decrease in any di- Through the CRS lens: Likewise, biases in conversational mension of diversity.” It’s essential to recognize settings concerning user logs have been examined, particthat this notion doesn’t solely originate from pop- ularly regarding the process of dataset curation. Szpektor ularity bias; various other biases can contribute as et al. and Yu et al. [ 45, 46 ] note that the limited number of well. Wang et al. [ 33 ] present a user-controllable individuals responsible for labeling conversation quality model to alleviate filter bubbles, which holds po- may introduce bias stemming from their personal prefertential for application in conversational settings. ences. Additionally, Pang et al. [ 47 ] highlight their focus on cultural diferences, which can result in divergent Through the CRS lens: Lin et al. [ 20 ] examine the preva- labelings of CRS datasets. lence of popularity bias by investigating three CRS mod- Looking back to the similar notions, conformity bias is els as baselines, revealing its existence in these systems also studied in the conversational recommender systems as well. It is speculated that this bias can be alleviated [ 48, 49 ]. Overall, the conversational nature of CRSs and through the conversational setting because the user is the inevitable increasing complexity can cause the model capable of criticizing the recommendations in order to to have more biases in the log than the classic models have a more niche recommended list of items. leading to a reduction of the quality of the datasets which needs to be addressed in future studies.

5.2. User Log Bias

5.3. Recommendation Evaluation Bias

Diferent underlying biases can lead to a biased log of user data. Therefore we define User Log bias as an umbrella In order to evaluate a recommender system, certain priorterm for the discrepancy between real-world representa- ities need to be addressed with respect to the needed spection of user-item interactions (the ground truth) and the ifications of the system in addition to common accuracy recorded data. This variation can happen due to many and ranking evaluation metrics. For example, serendipity reasons, such as oversimplifying the user logs, not hav- and diversity of the recommendations and long-term vs ing access to certain user data and imprecise user-item short-term fairness can all be taken into account as a interaction modeling. way to measure the recommender system [ 50, 51, 52 ].

There are some works that address this bias with difer- Therefore, it is important to establish metrics to assess ent approaches and diferent but similar bias definitions. the various aspects or qualities of a recommender system Frumermann et al. [ 34 ] investigate the real meaning be- that need to be evaluated. Also, the flaws of some of the hind rejected items and whether all the rejected items already established metrics have been challenged [ 53, 54 ] should be seen as the same. On the same issue, Nazari and some even propose a metric-less ofline evaluation et al. [ 35 ] make an efort to use user-implicit signals method [ 55 ]. Additional metrics such as evaluating fairand Xu et al. [ 36 ] try to leverage unclicked items in the ness in ranking [ 56 ], and recommendation uncertainty dataset in addition to the interactions. Lastly, Zhang et al. [ 57 ] are also addressed in the literature. Building on [ 37 ] focus on how we should tackle user inattentiveness fairness, De et al. [ 58 ] go beyond fairness metrics and to the items that are being interacted with but do not suggest that search engines can manipulate users while necessarily correlate with user satisfaction. maintaining top-notch fairness metrics. Lastly, Wang Similar Notions: et al. and Zhang et al. [ 59, 37 ] try to investigate user attention and how it afects recommendation models and the way users interact with it, directly influencing the ways we measure their capability.

Through the CRS lens: There can be various use-cases for conversational recommender systems and diferent measurements can and should be prioritized depending on them. This makes choosing the evaluation metrics for conversational recommender systems a challenging task.

For example, Lin et al. [ 20 ] propose that the Success Rate metric does not indicate how much the recommender system is benefiting each of its individual users. When evaluating a topic-guided conversational recommender • Conformity Bias: It is a cognitive bias which is defined as users’ disinclination towards negatively rating an item because of the item’s high ranking or popularity [ 25, 38 ]. • Exposure Bias: Since each user is randomly exposed to a subset of items during her lifetime, her profiling is biased towards the items that they have already been exposed to [ 14, 39, 40 ]. Also, it indicates that unobserved items do not always represent negative preference [ 25 ]. • Selection Bias: It is a similar notion to exposure

bias, as it is defined as the un-randomness of the system, specific metrics could come into play, such as and how experienced they are in interacting with a recpsychologically inspired measures and those assessing ommender system [ 69 ] and the model should be robust conversation quality and engagement [ 60 ]. These metrics in diferent interactions and be able to extract the users’ also address sub-objectives like accurate user satisfaction needs without emphasizing their demographic attributes. estimation [ 2, 61 ]. Lastly, it is very important to evalu- In a conversational setting, user interaction empowers ate the natural-language-understanding CRSs in order them to navigate and refine recommendations, allowing to measure the gap between the understanding of the them to mitigate attribute biases in the final list, similar NLU and how much this understanding is utilized by the to how they can address popularity bias. Nevertheless, recommender systems [ 62 ]. One of the fundamental rea- additional research is essential to examine which users’ sons is explored in the study by Zho et al. [ 8 ], where they attribute biases can be efectively mitigated through coninvestigate the disparity between the natural language versation, and also to understand which biases, such as representation of a potentially recommended item and gender or race bias, might be exacerbated due to the its lack of precise alignment. inherent biases of various components within CRS.

In the realm of CRS evaluation, the incorporation of new aspects amplifies complexity. Evaluating diferent 5.5. Position Bias modules and the complete CRS requires meticulous consideration, making CRSs more susceptible to recommen- It happens as users tend to interact with items in higher dation evaluation biases. positions of the recommendation list regardless of the items’ actual relevance so that the interacted items might 5.4. Attribute Bias not be highly relevant [ 25, 70 ].

Similar Notions: Attribute bias is a concerning issue in recommender systems. These systems can inadvertently amplify existing societal biases by making recommendations based on certain sensitive attributes that each user can have, such as gender, race, or age. This bias can lead to unfair and discriminatory outcomes, as individuals from certain demographics may be systematically excluded or receive less favourable recommendations. Addressing demographic bias in recommender systems is essential for ensuring equal and equitable treatment for all users, regardless of their demographic characteristics.

Similar notions: • Demographic bias and minority bias can also

be utilized to refer to Attribute bias [ 63, 64 ]. • User Activity Bias: There are several papers discussing the disparate impact of active users on the model [ 65 ], and how diferently the model interacts with them [ 66 ]. • Sentiment Bias: It is investigated that the more the users have positive interactions with the system they are more likely to get higher quality recommendations [ 16 ].

Through the CRS lens: Due to the potential usage of natu

ral language processing in conversational models, other attributes of users can be exposed to the model and can be exploited. For instance, in a voice dialogue system, the accent of the user is a sensitive attribute that should ideally not influence the system’s decision-making [ 67 ]. Also in the work by Cogswell et al. [ 68 ], it is discussed that the modality of presenting the data can afect the minorities’ ability to perceive it. On a similar issue, different people interact diferently with regard to their age • It is in accordance with lead bias in a work by Zhu et al. [ 71 ] for news recommender systems that show how the ‘lead’ part of the news in the recommender systems can bias a user’s behavior towards the item.

Through the CRS lens: In a conversational setting, there are options to counter position bias, such as providing explainability for each recommended item and framing the recommendations using natural language. These strategies can potentially mitigate the impact of position bias. It’s important to note that this bias is intertwined with Framing bias (Section 7), as the ranking can be seen as a framing of data presentation and position bias can be considered a form of framing bias as well. Nevertheless, it is crucial to acknowledge that explanation methods which rely on an additional or surrogate model to provide justifications for why specific items are recommended to the user and ranked higher, are also susceptible to biases. Consequently, these explanations might not accurately reflect the performance of the original model.

5.6. Personalization Bias

Personalization bias in recommender systems refers to the tendency of these systems to continuously recommend similar content based on a user’s preferences, potentially limiting their exposure to diverse perspectives and new experiences. Balancing personalization with serendipity is crucial to mitigate this bias and ensure users are presented with a broader range of recommendations [ 25 ].

Similar Notions: • User preference Amplification: Kalimeris et al. [ 72 ] talk about how even relevant and highquality recommendations can lead to user preference amplification, therefore, decreasing the users’ exposure to diverse content. • Feedback Loop: User tends to follow the recommendations and the recommendations become the user’s interests themselves, leading to ampliifed biases. Moreover, in the long run, and with repetition of this loop, the amplified biases become the ground truth as the user logs are utilized to train other models [ 73 ].

6. Biases of Attribute-aware

Conversational Recommender

Systems

Conversational recommender systems introduce a new set of biases that can impact recommendations and user experiences. While traditional biases in recommender systems have been extensively studied, the conversational nature of these systems introduces new biases in distinct ways. In this section, we conduct investigations on the existing biases of conversational recommender systems that do not have natural language understanding and interact with the user based on inquiring about attributes that they want to make decisions upon, such as [ 62, 1, 75, 76, 77 ].

6.1. Anchoring Bias

Through the CRS lens: Similar to the Popularity bias, in

a conversational setting, users’ ability to navigate the recommendations after receiving them gives them the option to reduce the attribute biases in the final list of recommendations as well.

With the inclusion of additional data such as conversations in user interactions, the model’s tendency to overemphasize the previous dialogues may increase, potentially exacerbating personalization bias. Therefore, the presence of diverse data sources should be carefully managed to strike a balance between personalization and diversity in recommendations.

In a conversational setting, the model has the option to

utilize the user information throughout the conversation, even through diferent sessions depending on the system. Therefore, it is a challenge to how this information and the user’s behavior and feedback to diferent recommendations should be utilized. Anchoring bias, or User History bias, happens when the previous recommenda5.7. Context Bias tions and the user dialogue history in the conversation can anchor subsequent recommendations, leading the In a work by Zheng et al. [ 74 ], the concept of ”con- system to focus on a particular subset of items and potext bias” is explored as a comprehensive framework tentially overlooking other options, therefore it is vital for analyzing a collection of recommended items. The to catch the dynamics of user profile while being able to study highlights that users’ decision-making processes make use of the information. There have been some studcan be influenced by a combination of biases when pre- ies related to this bias. It is investigated that the dynamic sented with options that possess diferent attributes and nature of conversational systems can amplify the impact potentially unique biases for each item. For instance, of anchoring bias [ 78, 4 ]. Also, Ren et al. [ 5 ] investigate when browsing a news website, users may encounter the diferences between old and new user preferences various forms of content, such as text and video (modal- and the way they change. ity bias), alongside popularity-driven recommendations (popularity bias). Understanding how this set of biases, 6.2. Attribute Selection Bias collectively referred to as context bias, impacts decisionmaking requires a broader and more in-depth investiga- Attribute selection bias, which can also be called User tion. Preference Assumption bias, refers to a phenomenon in Through the CRS lens: In conversational settings, context conversational recommender systems where the system bias can play a vital role as the conversational nature of becomes biased towards the attributes that users priorthe system makes it more complicated and the dynamics itize when making decisions [ 20 ]. In each step of the of the model and diferent existing biases should also be recommendation process, the system may assume that addressed in the same context. In a conversation, the the user wants to base their choices on specific attributes, concept of context assumes a broader and more intri- thereby influencing the recommendations accordingly. cate definition compared to its application in traditional This bias can impact the diversity and fairness of the recommender systems confined to lists of items. There- recommendations by potentially overlooking alternative fore, the potential exacerbation of context bias becomes attributes that users might value but are not explicitly exa relevant consideration when applied to CRSs. pressed. By primarily focusing on a subset of attributes, the system may limit the scope of recommendations, potentially hindering serendipitous discoveries and failing to provide a comprehensive and personalized experience. 6.3. Human-AI Interaction Bias user queries [ 89, 90 ]. However, if the system fails to address this bias efectively, it may lead to the recomAI-conversation bias refers to a type of bias where users mendation of low-quality items that do not align with alter their conversational behavior and speech patterns the user’s preferences. Consequently, users may need to when interacting with a conversational recommender provide feedback or criticize the initial recommendations system that they know is powered by an AI bot [ 79 ]. to prompt the system to refine its list of suggested items When users are aware that they are conversing with an [ 46, 7, 91 ]. Addressing Defective Query bias requires artificial intelligence rather than a human, they may con- the development of robust and adaptive conversational sciously or unconsciously modify their language, tone, or recommender systems that can handle uncertainties, disstyle of communication. This bias can arise from various ambiguate user queries, and incorporate user feedback to factors, including a perceived need to simplify language, improve the quality and relevance of recommendations. adapt to the system’s limitations, or conform to social norms associated with human-AI interactions. As a result, the quality and naturalness of the conversation may 7.2. Cognitive Biases be afected, potentially leading to a less engaging and authentic user experience [ 80 ].

6.4. Modality Bias

Modality bias in a conversational recommender system setting refers to the tendency of multi-modal text generation models, such as the multi-modal GPT-4 model [ 81 ], to heavily rely on textual input while paying less attention to non-textual signals, such as visual cues or signals [ 82 ]. This bias can limit the system’s ability to efectively incorporate and leverage non-textual signals, leading to a potential loss of valuable information and a less holistic understanding of user preferences. Addressing modality bias involves developing more balanced and comprehensive models that can efectively capture and utilize both textual and non-textual cues.

7. Biases of Topic-guided

Conversational Recommender

Systems

In this section, our focus is to investigate the biases of natural language understanding models and their potential implications when integrated into conversational recommender systems with natural language modules. These types of models can be considered as a general type of the existing CRSs in the literature [ 8, 83 ].

7.1. Defective Queries Bias

Defective query bias in a conversational recommender system setting occurs when users intentionally manipulate the system or unknowingly express ambiguous statements, making it dificult for the system to comprehend the conversation and generate useful recommendations [ 3, 84, 85, 86, 87, 88 ]. This poses a challenge as the system struggles to fully grasp user intent and preferences. In such situations, the system may need to ask clarifying questions to obtain additional context and disambiguate Cognitive bias in language models within a conversational recommender system setting involves the evaluation of large language models in relation to the cognitive biases observed in humans. These biases have the potential to directly impact the system’s natural language generation module, influencing it to make irrational decisions depending on whether or not it is afected by these cognitive biases. Detecting and understanding the presence of cognitive biases in language models is crucial to ensure that the recommendations provided are fair and unbiased. By addressing and mitigating these biases, conversational recommender systems can strive to deliver more objective and rational recommendations that are not influenced by human cognitive biases [ 92 ]. Similar Notions: • Framing bias: Framing in conversational recommender systems refers to how information presentation influences user perception and decisionmaking. By addressing framing biases through transparent and balanced recommendations, systems can enhance fairness and efectiveness [ 93, 94 ]. • Uncertainty-aversion bias: This bias arises from users’ negative inclination toward uncertain recommendations [ 95 ]. It can impact the efects of explainability techniques on user behavior. Moreover, it can also be studied on the efects of uncertainty in manually labeling datasets. • User Trust Bias: User Trust Bias It is been studied that having a conversational interface will increase the trust rate of users [ 96, 6 ]. Karduni et al. [ 97 ] argue that the way the faces are shown in news posts afects how much the users trust the platform. Also, the proactivity of the bot is discussed in the works by Kraus et al. [ 98, 99 ], Zhu et al. [100] and Lei et al. [101] and it is verified that it afects human trust.

7.3. Unintended Bias

Unintended bias in a conversational recommender system setting refers to the unintentional influence that certain factors or characteristics may have on the recommendations provided [ 21 ]. These biases can arise from societal stereotypes, historical data biases, or implicit associations present in language modeling and recommendation algorithms. Unintended biases can result in unequal treatment or favoritism towards certain groups or preferences, leading to potentially unfair or biased recommendations.

7.4. Persona Bias

Rises From RS CRS RS Training RS DMS RS RS DMS DMS User NLU User NLG NLG (NLU) NLU (NLG)

First Afects DMS Logs RS DMS NLG DMS User RS NLG NLU DMS NLU User User (DMS) DMS (DMS) There is Persona bias in a dialogue system, and therefore in a topic-guided CRS, refers to the detrimental discrep- Table 1 ancies in responses that arise when diferent personas The various biases and their origins are listed, along with their are adopted. There are a few works in dialogue mod- efects on diferent modules of the CRS. Each of the biases els and recommendation systems that try to adapt the can be traced back to a route in the architecture of a CRS generation process by conforming to the user’s persona illustrated in Figure 1. The start of route would be where the [102, 103, 104, 105, 106, 107? ]. These biases manifest bias rises from and the initial efecting point of the bias would in various ways, such as variations in the level of ofen- be the end of the route. This would enable us to navigate siveness or agreement with harmful statements within through the system in order to investigate the interplay of the the generated responses. The adoption of diferent de- diferent biases and their efects on each other. mographic personas can lead to unequal or biased treatment of users based on their demographic attributes [108].

With a similar outlook, Melchiorre et al. [ 18 ] investigate how diferent user personalities afect the recommendations that they receive in a music recommender system. address and mitigate biases in recommendation systems efectively.

In addition to addressing the above limitations, in future research endeavours, an exploration of the intersection between these biases presents a promising avenue.

Delving into the intricate interplay of these biases of the system could ofer a more comprehensive comprehension. Additionally, leveraging the insights from Table 1, a deeper understanding of how certain biases might synergize within specific system architectures could shed light on the propagation of bias and its efects on user interactions and decision-making processes.

8. Limitations and Future Work

When conducting a survey on CRS biases, it’s crucial to acknowledge some limitations that might impact the scope and depth of the findings. This paper acknowledges the potential relevance of natural language processing conferences as valuable sources of related references on biases [109]. However, these conferences were not extensively explored here. Additionally, focusing solely on papers published after 2019 might overlook earlier 9. Conclusion works providing historical context and a deeper understanding of CRS biases’ evolution. While the survey This survey paper establishes a non-linear taxonomy of aims for wide-ranging coverage, the references might biases in conversational recommender systems. The innot encompass all relevant literature on each bias. Nev- vestigation first covers biases in classic recommender ertheless, the survey prioritizes breadth to establish a systems and their adaptability to conversational recfoundational understanding. Furthermore, lacking for- ommender systems, highlighting their nature in conmal bias definitions and evaluation metrics may afect versational settings. Subsequently, biases in conversaresults, making their inclusion desirable for better consis- tional recommender systems are explored in two parts: tency. Lastly, the study’s limitations include the absence attribute-aware systems and topic-guided systems. By of experimentation on datasets, models, and presented providing this taxonomy, the paper aims to aid remitigation methods. Despite these limitations, the sur- searchers and developers in efectively addressing biases, vey ofers valuable insights into CRS biases, paving the ensuring fairness, transparency, and user trust in these way for more in-depth research and robust approaches to systems’ direct impact on user decision-making.

Acknowledgments

This research was partially supported by the Canada CIFAR AI Chair program (Mila), the Facebook research award and the Natural Sciences and Engineering Research Council of Canada (NSERC).

References

[1] W. Lei , G. Zhang , X. He , Y. Miao , X. Wang , L. Chen , T.-S. Chua, Interactive path reasoning on graph for conversational recommendation , in: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , 2020 , pp. 2073 - 2083 .

[2] Y. Deng , W. Zhang , W. Lam, H. Cheng, H. Meng, User satisfaction estimation with sequential dialogue act modeling in goal-oriented conversational systems , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2998 - 3008 .

[3] S. Li , B. P. Majumder , J. McAuley , Self-supervised bot play for conversational recommendation with justifications , arXiv preprint arXiv:2112.05197 ( 2021 ).

[4] S. Li , R. Xie , Y. Zhu , X. Ao , F. Zhuang , Q. He , User-centric conversational recommendation with multi-aspect user modeling , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 223 - 233 .

[5] Z. Ren , Z. Tian , D. Li , P. Ren , L. Yang , X. Xin , H. Liang , M. de Rijke , Z. Chen , Variational reasoning about user preferences for conversational recommendation , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 165 - 175 .

[6] M. Radensky , J. A. Séguin , J. S. Lim , K. Olson , R. Geiger, “ i think you might like this”: Exploring efects of confidence signal patterns on trust in and reliance on conversational recommender systems , in: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , 2023 , pp. 792 - 804 .

[7] H. Li , S. Sanner , K. Luo , G. Wu, A ranking optimization approach to latent linear critiquing for conversational recommender systems , in: Proceedings of the 14th ACM Conference on Recommender Systems , 2020 , pp. 13 - 22 .

[8] K. Zhou , W. X. Zhao , S. Bian , Y. Zhou , J.-R. Wen , J. Yu , Improving conversational recommender systems via knowledge graph based semantic fusion , in: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , 2020 , pp. 1006 - 1014 .

[9] D. Jannach , L. Chen, Conversational recommendation: A grand ai challenge , 2022 . arXiv: 2203 . 09126 .

[10] P. P. Ray , Chatgpt: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope, Internet of Things and Cyber-Physical Systems ( 2023 ).

[11] S. Dai , N. Shao , H. Zhao , W. Yu , Z. Si , C. Xu , Z. Sun , X. Zhang , J. Xu, Uncovering chatgpt's capabilities in recommender systems , 2023 . arXiv: 2305 . 02182 .

[12] D. Jin , L. Wang , H. Zhang , Y. Zheng , W. Ding , F. Xia , S. Pan , A survey on fairness-aware recommender systems , 2023 . arXiv: 2306 . 00403 .

[13] S. Milano , M. Taddeo , L. Floridi, Recommender systems and their ethical challenges , Ai & Society 35 ( 2020 ) 957 - 967 .

[14] S. Gupta , H. Wang , Z. Lipton , Y. Wang , Correcting exposure bias for link recommendation , in: International Conference on Machine Learning, PMLR , 2021 , pp. 3953 - 3963 .

[15] H. Liu , D. Tang , J. Yang , X. Zhao , H. Liu , J. Tang , Y. Cheng, Rating distribution calibration for selection bias mitigation in recommendations , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2048 - 2057 .

[16] C. Lin , X. Liu , G. Xv, H. Li , Mitigating sentiment bias for recommender systems , in: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2021 , pp. 31 - 40 .

[17] Z. Zhu , Y. He , X. Zhao , Y. Zhang , J. Wang , J. Caverlee , Popularity-opportunity bias in collaborative ifltering , in: Proceedings of the 14th ACM International Conference on Web Search and Data Mining , 2021 , pp. 85 - 93 .

[18] A. B. Melchiorre , E. Zangerle, M. Schedl , Personality bias of music recommendation algorithms , in: Proceedings of the 14th ACM Conference on Recommender Systems , 2020 , pp. 533 - 538 .

[19] D. Jannach , A. Manzoor , W. Cai , L. Chen , A survey on conversational recommender systems , ACM Computing Surveys 54 ( 2021 ) 1 - 36 . URL: https://doi.org/10.1145%2F3453154. doi: 10 .1145/ 3453154.

[20] S. Lin , Z. Zhu , J. Wang , J. Caverlee , Towards fair conversational recommender systems , arXiv preprint arXiv:2208.03854 ( 2022 ).

[21] T. Shen , J. Li , M. R. Bouadjenek , Z. Mai , S. Sanner , Unintended bias in language model-driven conversational recommendation , arXiv preprint arXiv:2201.06224 ( 2022 ).

[22] N. Mehrabi , F. Morstatter , N. Saxena , K. Lerman , A. Galstyan , A survey on bias and fairness in machine learning , 2022 . arXiv: 1908 .09635.

[23] S. L. Blodgett , S. Barocas , H. Daumé III , H. Wallach, Language (technology) is power: A critical survey of” bias” in nlp , arXiv preprint arXiv: 2005 . 14050 ( 2020 ).

[24] Y. Wang , W. Ma, M. Zhang, Y. Liu, S. Ma , A survey on the fairness of recommender systems , ACM Transactions on Information Systems 41 ( 2023 ) 1 - 43 .

[25] J. Chen , H. Dong , X. Wang , F. Feng , M. Wang , X. He , Bias and debias in recommender system: A survey and future directions, 2021 . arXiv: 2010 .03240.

[26] A. Lin , J. Wang , Z. Zhu , J. Caverlee , Quantifying and mitigating popularity bias in conversational recommender systems , in: Proceedings of the 31st ACM International Conference on Information & Knowledge Management , 2022 , pp. 1238 - 1247 .

[27] D. Jannach , Evaluating conversational recommender systems , Artificial Intelligence Review 56 ( 2022 ) 2365 - 2400 . URL: https://doi. org/10.1007% 2Fs10462 - 022 -10229-x. doi: 10 .1007/ s10462-022-10229-x.

[28] W. Brown , A. Agarwal, Diversified recommendations for agents with adaptive preferences , Advances in Neural Information Processing Systems 35 ( 2022 ) 26066 - 26077 .

[29] A. Zhang , J. Zheng , X. Wang , Y. Yuan , T.-S. Chua, Invariant collaborative filtering to popularity distribution shift , in: Proceedings of the ACM Web Conference 2023 , 2023 , pp. 1240 - 1251 .

[30] H. Abdollahpouri , M. Mansoury , R. Burke , B. Mobasher , The connection between popularity bias, calibration, and fairness in recommendation , in: Proceedings of the 14th ACM Conference on Recommender Systems , 2020 , pp. 726 - 731 .

[31] H. Abdollahpouri , R. Burke , B. Mobasher , Managing popularity bias in recommender systems with personalized re-ranking , arXiv preprint arXiv: 1901 . 07555 ( 2019 ).

[32] L. Michiels , J. Leysen , A. Smets , B. Goethals , What are filter bubbles really? a review of the conceptual and empirical work , in: Adjunct Proceedings of the 30th ACM Conference on User Modeling, Adaptation and Personalization , 2022 , pp. 274 - 279 .

[33] W. Wang , F. Feng , L. Nie , T.-S. Chua, Usercontrollable recommendation against filter bubbles , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 1251 - 1261 .

[34] S. Frumerman , G. Shani , B. Shapira , O. Sar Shalom , Are all rejected recommendations equally bad? towards analysing rejected recommendations , in: Proceedings of the 27th ACM Conference on User Modeling, Adaptation and Personalization , 2019 , pp. 157 - 165 .

[35] Z. Nazari , P. Chandar , G. Fazelnia, C. M. Edwards , B. Carterette , M. Lalmas , Choice of implicit signal matters: Accounting for user aspirations in podcast recommendations , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2433 - 2441 .

[36] Z. Xu , P. Wei , W. Zhang , S. Liu, L. Wang , B. Zheng , Ukd: Debiasing conversion rate estimation via uncertainty-regularized knowledge distillation , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2078 - 2087 .

[37] X. Zhang , S. Dai, J. Xu , Z. Dong , Q. Dai , J.-R. Wen , Counteracting user attention bias in music streaming recommendation via reward modification , in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022 , pp. 2504 - 2514 .

[38] Y. Zheng , C. Gao , X. Li , X. He , Y. Li , D. Jin , Disentangling user interest and conformity for recommendation with causal embedding , in: Proceedings of the Web Conference 2021 , 2021 , pp. 2980 - 2991 .

[39] A. Dash , A. Chakraborty , S. Ghosh , A. Mukherjee , K. P. Gummadi , When the umpire is also a player: Bias in private label product recommendations on e-commerce marketplaces , in: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency , 2021 , pp. 873 - 884 .

[40] J. McInerney , B. Brost , P. Chandar , R. Mehrotra , B. Carterette , Counterfactual evaluation of slate recommendations with sequential reward interactions , in: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , 2020 , pp. 1779 - 1788 .

[41] W. Zhang , W. Bao , X.-Y. Liu, K. Yang , Q. Lin , H. Wen , R. Ramezani , Large-scale causal approaches to debiasing post-click conversion rate estimation with multi-task learning , in: Proceedings of The Web Conference 2020 , 2020 , pp. 2775 - 2781 .

[42] Z. Ovaisi , R. Ahsan , Y. Zhang , K. Vasilaky , E. Zheleva, Correcting for selection bias in learning-torank systems , in: Proceedings of The Web Conference 2020 , 2020 , pp. 1863 - 1873 .

[43] J. Huang , H. Oosterhuis , M. De Rijke , H. Van Hoof , Keeping dataset biases out of the simulation: A debiased simulator for reinforcement learning based recommender systems , in: Proceedings of the 14th ACM Conference on Recommender Systems , 2020 , pp. 190 - 199 .

[44] T. Yang , C. Luo , H. Lu , P. Gupta , B. Yin , Q. Ai , Can clicks be both labels and features? unbiased behavior feature collection and uncertainty-aware learning to rank , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 6 - 17 .

[45] I. Szpektor , D. Cohen , G. Elidan , M. Fink , A. Hassidim , O. Keller, S. Kulkarni, E. Ofek , S. Pudinsky , A. Revach , et al., Dynamic composition for conversational domain exploration , in: Proceedings of The Web Conference 2020 , 2020 , pp. 872 - 883 .

[46] T. Yu , Y. Shen , H. Jin , A visual dialog augmented interactive recommender system , in: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , 2019 , pp. 157 - 165 .

[47] R. Y. Pang , J. Cenatempo , F. Graham , B. Kuehn , M. Whisenant , P. Botchway , K. Stone Perez , A. Koenecke , Auditing cross-cultural consistency of human-annotated labels for recommendation systems , in: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , 2023 , pp. 1531 - 1552 .

[48] A. Raul , A. P. Dharwadker , B. Schumitsch, Cam2: Conformity-aware multi-task ranking model for large-scale recommender systems , arXiv preprint arXiv:2304.08562 ( 2023 ).

[49] Y. Yang , C. Huang , L. Xia , C. Huang , D. Luo , K. Lin , Debiased contrastive learning for sequential recommendation , in: Proceedings of the ACM Web Conference 2023 , 2023 , pp. 1063 - 1073 .

[50] A. D'Amour , H. Srinivasan , J. Atwood , P. Baljekar , D. Sculley , Y. Halpern , Fairness is not static: deeper understanding of long term fairness via simulation studies , in: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , 2020 , pp. 525 - 534 .

[51] H. Wen , X. Yi , T. Yao , J. Tang , L. Hong , E. H. Chi , Distributionally-robust recommendations for improving worst-case user experience , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 3606 - 3610 .

[52] M. Mladenov , E. Creager , O. Ben-Porat , K. Swersky , R. Zemel , C. Boutilier, Optimizing long-term social welfare in recommender systems: A constrained matching approach , in: International Conference on Machine Learning, PMLR, 2020 , pp. 6987 - 6998 .

[53] G. Hiranandani , W. Vijitbenjaronk , S. Koyejo , P. Jain , Optimization and analysis of the pap@ k metric for recommender systems , in: International Conference on Machine Learning, PMLR , 2020 , pp. 4260 - 4270 .

[54] K. Christakopoulou , M. Traverse , T. Potter , E. Marriott , D. Li , C. Haulk , E. H. Chi , M. Chen , Deconfounding user satisfaction estimation from response rate bias , in: Proceedings of the 14th ACM Conference on Recommender Systems , 2020 , pp. 450 - 455 .

[55] F. Diaz , A. Ferraro , Ofline retrieval evaluation without evaluation metrics , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 599 - 609 .

[56] A. Raj , M. D. Ekstrand , Measuring fairness in ranked results: An analytical and empirical comparison , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 726 - 736 .

[57] D. Cohen , B. Mitra , O. Lesota , N. Rekabsaz , C. Eickhof , Not all relevance scores are equal: Eficient uncertainty and calibration modeling for deep retrieval models , in: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2021 , pp. 654 - 664 .

[58] T. De Jonge , D. Hiemstra , Unfair: Search engine manipulation, undetectable by amortized inequity , in: Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , 2023 , pp. 830 - 839 .

[59] X. Wang , W. Zhu , C. Liu, Social recommendation with optimal limited attention , in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , 2019 , pp. 1518 - 1527 .

[60] A. Ghandeharioun , J. H. Shen , N. Jaques , C. Ferguson , N. Jones , A. Lapedriza , R. Picard , Approximating interactive human evaluation with self-play for open-domain dialog systems , Advances in Neural Information Processing Systems 32 ( 2019 ).

[61] W. Sun , S. Zhang , K. Balog , Z. Ren , P. Ren , Z. Chen , M. de Rijke, Simulating user satisfaction for the evaluation of task-oriented dialogue systems , in: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2021 , pp. 2499 - 2506 .

[62] X. Wang , K. Zhou , J.-R. Wen , W. X. Zhao , Towards unified conversational recommender systems via knowledge-enhanced prompt learning , in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022 , pp. 1929 - 1937 .

[63] R. Islam , K. N. Keya , S. Pan , J. Foulds , Mitigating demographic biases in social media-based recommender systems , KDD (Social Impact Track) ( 2019 ).

[64] Y. Ying , F. Zhuang , Y. Zhu , D. Wang , H. Zheng , Camus: Attribute-aware counterfactual augmentation for minority users in recommendation , in: Proceedings of the ACM Web Conference 2023 , 2023 , pp. 1396 - 1404 .

[65] F. Eskandanian , N. Sonboli , B. Mobasher , Power of the few: Analyzing the impact of influential users in collaborative recommender systems , in: Proceedings of the 27th ACM Conference on User Modeling, Adaptation and Personalization , 2019 , pp. 225 - 233 .

[66] Y. Li , H. Chen , Z. Fu , Y. Ge , Y. Zhang, User-oriented fairness in recommendation , in: Proceedings of the Web Conference 2021 , 2021 , pp. 624 - 632 .

[67] D. Harwell , The Accent Gap, https://www. washingtonpost.com/graphics/2018/business/ alexa -does-not- understand- your-accent/, 2018 .

[68] M. Cogswell , J. Lu , R. Jain , S. Lee , D. Parikh , D. Batra , Dialog without dialog data: Learning visual dialog agents from vqa data , Advances in Neural Information Processing Systems 33 ( 2020 ) 19988 - 19999 .

[69] Z. Zheng , Z. Qiu , H. Xiong , X. Wu , T. Xu , E. Chen , X. Zhao , Ddr: Dialogue based doctor recommendation for online medical service , in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022 , pp. 4592 - 4600 .

[70] M. Rufini , V. Bellini , A. Buchholz , G. Di Benedetto , Y. Stein , Modeling position bias ranking for streaming media services , in: Companion Proceedings of the Web Conference 2022 , 2022 , pp. 72 - 76 .

[71] C. Zhu , Z. Yang , R. Gmyr , M. Zeng , X. Huang , Leveraging lead bias for zero-shot abstractive news summarization , in: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2021 , pp. 1462 - 1471 .

[72] D. Kalimeris , S. Bhagat , S. Kalyanaraman , U. Weinsberg, Preference amplification in recommender systems , in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , 2021 , pp. 805 - 815 .

[73] M. Mansoury , H. Abdollahpouri , M. Pechenizkiy , B. Mobasher , R. Burke , Feedback loop and bias amplification in recommender systems , in: Proceedings of the 29th ACM international conference on information & knowledge management , 2020 , pp. 2145 - 2148 .

[74] Z. Zheng , Z. Qiu , T. Xu , X. Wu , X. Zhao , E. Chen , H. Xiong , Cbr: context bias aware recommendation for debiasing user modeling and click prediction , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2268 - 2276 .

[75] S. Li , W. Lei , Q. Wu , X. He , P. Jiang , T.-S. Chua, Seamlessly unifying attributes and items: Conversational recommendation for cold-start users , ACM Transactions on Information Systems (TOIS) 39 ( 2021 ) 1 - 29 .

[76] X. Ren , H. Yin , T. Chen , H. Wang , Z. Huang , K. Zheng , Learning to ask appropriate questions in conversational recommendation , in: Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , 2021 , pp. 808 - 817 .

[77] K. Xu , J. Yang , J. Xu , S. Gao , J. Guo , J.-R. Wen , Adapting user preference to online feedback in multi-round conversational recommendation , in: Proceedings of the 14th ACM international conference on web search and data mining , 2021 , pp. 364 - 372 .

[78] C. Qu , L. Yang , M. Qiu , W. B. Croft , Y. Zhang , M. Iyyer, Bert with history answer embedding for conversational question answering , in: Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval , 2019 , pp. 1133 - 1136 .

[79] A. Følstad , T. Araujo , E. L. -C. Law , P. B. Brandtzaeg , S. Papadopoulos , L. Reis , M. Baez , G. Laban, P. McAllister , C. Ischen , et al., Future directions for chatbot research: an interdisciplinary research agenda , Computing 103 ( 2021 ) 2915 - 2942 .

[80] E. Hosseini-Asl , B. McCann , C.-S. Wu , S. Yavuz , R. Socher , A simple language model for taskoriented dialogue , Advances in Neural Information Processing Systems 33 ( 2020 ) 20179 - 20191 .

[81] OpenAI , Gpt-4 technical report , 2023 . arXiv: 2303 . 08774 .

[82] Z. Tian , Z. Xie , F. Lin , Y. Song , A multi-view metalearning approach for multi-modal response generation , in: Proceedings of the ACM Web Conference 2023 , 2023 , pp. 1938 - 1947 .

[83] R. Sarkar , K. Goswami , M. Arcan , J. P. McCrae , Suggest me a movie for tonight: Leveraging knowledge graphs for conversational recommendation , in: Proceedings of the 28th International Conference on Computational Linguistics , 2020 , pp. 4179 - 4189 .

[84] Z. Lu , R. H. Kazi , L.-y. Wei, M. Dontcheva , K. Karahalios , Streamsketch: Exploring multi-modal interactions in creative live streams , Proceedings of the ACM on Human-Computer Interaction 5 ( 2021 ) 1 - 26 .

[85] T. Shen , Z. Mai , G. Wu, S. Sanner , Distributional contrastive embedding for clarification-based conversational critiquing , in: Proceedings of the ACM Web Conference 2022 , 2022 , pp. 2422 - 2432 .

[86] Z. Wang , Y. Tu , C. Rosset , N. Craswell , M. Wu , Q. Ai , Zero-shot clarifying question generation for conversational search , in: Proceedings of the ACM Web Conference 2023 , 2023 , pp. 3288 - 3298 .

[87] J. Liao , X. Zhao , J. Zheng , X. Li , F. Cai , J. Tang , Ptau: Prompt tuning for attributing unanswerable questions , in: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022 , pp. 1219 - 1229 . [99] M. Kraus , N. Wagner , W. Minker, Efects of proac-

[88] M. S. Mirzaei , K. Meshgi , S. Sekine, Is this ques- tive dialogue strategies on human-computer trust, tion real? dataset collection on perceived inten- in: Proceedings of the 28th ACM Conference on tions and implicit attack detection , in: Proceed- User Modeling , Adaptation and Personalization, ings of the ACM Web Conference 2022 , 2022 , pp. 2020 , pp. 107 - 116 . 2850 - 2859 . [100] Y. Zhu , J.-Y. Nie , K. Zhou , P. Du , H. Jiang , Z. Dou ,

[89] A. Montazeralghaem , J. Allan , P. S. Thomas, Large- Proactive retrieval-based chatbots based on relescale interactive conversational recommendation vant knowledge and goals, in: Proceedings of the system using actor-critic framework , in: Pro- 44th International ACM SIGIR Conference on Receedings of the 15th ACM Conference on Recom- search and Development in Information Retrieval, mender Systems , 2021 , pp. 220 - 229 . 2021 , pp. 2000 - 2004 .

[90] M. Aliannejadi , H. Zamani , F. Crestani , W. B. [101] W. Lei , Y. Zhang , F. Song , H. Liang , J. Mao , Croft, Asking clarifying questions in open-domain J . Lv , Z. Yang , T.-S. Chua, Interacting with noninformation-seeking conversations, in: Proceed- cooperative user: A new paradigm for proactive ings of the 42nd international acm sigir conference dialogue policy , in: Proceedings of the 45th Interon research and development in information re- national ACM SIGIR Conference on Research and trieval, 2019 , pp. 475 - 484 . Development in Information Retrieval, 2022 , pp.

[91] P. Yang , H. Huang , W. Wei , X.-L. Mao , Toward 212 - 222 . real-life dialogue state tracking involving negative [102] S. Fatahi , M. Mousavifar , J. Vassileva , Investigating feedback utterances, in: Proceedings of the 28th the efectiveness of persuasive justification mesACM SIGKDD Conference on Knowledge Discov- sages in fair music recommender systems for users ery and Data Mining , 2022 , pp. 2222 - 2232 . with diferent personality traits , in: Proceedings

[92] E. Jones , J. Steinhardt , Capturing failures of large of the 31st ACM Conference on User Modeling, language models via human cognitive biases , Ad- Adaptation and Personalization , 2023 , pp. 66 - 77 . vances in Neural Information Processing Systems [103] Z. Han , S. Zhang , X. Zhang , Persona consistent 35 ( 2022 ) 11785 - 11799 . dialogue generation via contrastive learning , in:

[93] M. Reiter-Haas , Exploration of framing biases in Companion Proceedings of the ACM Web Conferpolarized online content consumption , in: Com- ence 2023 , 2023 , pp. 196 - 199 . panion Proceedings of the ACM Web Conference [104] C. Xu , P. Li , W. Wang , H. Yang , S. Wang , C. Xiao , 2023 , 2023 , pp. 560 - 564 . Cosplay: Concept set guided personalized dia-

[94] M. Mulder , O. Inel , J. Oosterman , N. Tintarev , Oper- logue generation across both party personas, in: ationalizing framing to support multiperspective Proceedings of the 45th International ACM SIGIR recommendations of opinion pieces , in: Proceed- Conference on Research and Development in Inings of the 2021 ACM conference on fairness, ac- formation Retrieval , 2022 , pp. 201 - 211 . countability, and transparency, 2021 , pp. 478 - 488 . [105] J.-C. Gu , H. Liu , Z.-H. Ling , Q. Liu , Z. Chen , X. Zhu ,

[95] S. Park , J. Y. Park , H. Chin , J.-h. Kang, M. Cha , An Partner matters! an empirical study on fusing experimental study to understand user experience personas for personalized response selection in and perception bias occurred by fact-checking retrieval-based chatbots, in: Proceedings of the messages , in: Proceedings of the Web Conference 44th International ACM SIGIR Conference on Re2021 , 2021 , pp. 2769 - 2780 . search and Development in Information Retrieval,

[96] A. Gupta , D. Basu , R. Ghantasala , S. Qiu , U. Gadi- 2021, pp. 565 - 574 . raju, To trust or not to trust: How a conversational [106] Z. Ma , Z. Dou , Y. Zhu , H. Zhong , J.-R. Wen , One interface afects trust in a decision support system, chatbot per person: Creating personalized chatin: Proceedings of the ACM Web Conference 2022 , bots based on implicit user profiles , in: Proceed2022 , pp. 3531 - 3540 . ings of the 44th international ACM SIGIR confer-

[97] A. Karduni , R. Wesslen , D. Markant , W. Dou, Im- ence on research and development in information ages, emotions, and credibility: Efect of emo- retrieval , 2021 , pp. 555 - 564 . tional facial images on perceptions of news con- [107] A. Tigunova , A. Yates , P. Mirza , G. Weikum, Listent bias and source credibility in social media, tening between the lines: Learning personal atarXiv preprint arXiv:2102.13167 ( 2021 ). tributes from conversations , in: The World Wide

[98] M. Kraus , N. Wagner , N. Untereiner , W. Minker, In- Web Conference, 2019 , pp. 1818 - 1828 . cluding social expectations for trustworthy proac - [108] E. Sheng , J. Arnold , Z. Yu , K.-W. Chang , N. Peng, tive human-robot dialogue , in: Proceedings of the Revealing persona biases in dialogue systems, 30th ACM Conference on User Modeling, Adapta- arXiv preprint arXiv:2104.08728 ( 2021 ). tion and Personalization , 2022 , pp. 23 - 33 . [109] Y. Wan , W. Wang , P. He , J. Gu , H. Bai , M. Lyu,