Recommendation via Entering a Human Conversation

doi:10.21203/rs.3.rs-2990423/v1

Download PDF

Research Article

Recommendation via Entering a Human Conversation

https://doi.org/10.21203/rs.3.rs-2990423/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

We propose a novel way of the conversational recommendation where instead of asking user questions to acquire her preferences, the recommender tracks her conversations with other people, including customer support agents (CSA) and joins the conversation only when there is something important to recommend and the time is correct to do so. Building a recommender that joins a human conversation (CRS), we propose information extraction, discourse and argumentation analyses, as well as dialogue management techniques to compute a recommendation for a product and service that is badly needed by the customer, as inferred from the conversation. A special case of such conversations is considered where the customer raises his problem with a CSA in an attempt to resolve at, along with receiving a recommendation for a product with features addressing this problem. Performance of CRS is evaluated in a number of human-human and human-chatbot dialogues and demonstrates that CRS is efficient and less intrusive way to provide high relevance and persuasive recommendations.

In the last decade, there is a great expectation in the area of recommender systems implemented via conversational AI. This is due to popularity of virtual digital assistants such as Google Bard, Microsoft Bing, OpenAI ChatGPT, Amazon Alexa, and Siri. These assistants are capable of conversing with people in the form of a dialogue initiated mostly by people. In the emerging field of conversational recommendation systems (CRSs), a dialogue needs to be with mixed initiative and multi-party, where at some point the system interrupts conversing humans and starts to provide a recommendation relevant to the given discussion, expected to address a problem being communicated in this discussion.

Nowadays, the CRS target is a purchase service wherein the users can receive recommendations through conversations that capture their preferences. It would be amazing to get a music/movie recommendation or the apparel that one likes to have thorough conversations than browsing through the vast list of choices available on the e-commerce portal (Dalton et al., 2018). Along these lines, CRSs have evolved with personalized recommenders in various industries such as health care, entertainment, education, e-commerce, and services. An example is an intelligent customer assistant that gives conversation-based eyewear recommendations (Costello et al., 2006)

While most CRSs focus on user-CRS dialogue, we take a step forward and tackle user – user-opponent – CRS dialogue between multiple human agents, where CRS intervenes user – user-opponent conversation only when appropriate. This scenario has a number of advantages:

The user is not bored with routine questions required for personalization
The user is not destructed from normal activities chatting with friends or solving a problem with a customer support agent.
Recommendation are not forced upon a user but instead come naturally when CRS believes it is appropriate

Due to the popularity of texting and messaging and a recent advancement of deep learning technologies, a conversation-based interaction becomes an emerging usr interface. Over the last decade, recommendation and advertisement researchers and practitioners have advanced the frontiers of personalized recommendation by making them function in a conversational mode. While modern conversation platforms offer basic dialogue capabilities such as natural language (NL) understanding, entity extraction and simple dialogue management, there are still challenges in developing practical applications to support complex use cases such as dialogue-based recommendation (Thompson et al 2004, Christakopoulou et al 2016, Sun and Zhang 2018).

More and more precise and powerful recommendation algorithms and techniques have been proposed capable of effectively assessing users' tastes and predict information that would probably be of interest to them. Most of these approaches rely on the machine learning (ML) – based collaborative filtering and do not take into account the huge amount of knowledge, both structured and non-structured ones, such as prior user utterances in a dialogue describing the domain of interest for the recommendation engine (Anelli et al 2018). Collaborative filtering is a recommendation approach that filters out items that a user might like on the basis of accepted recommendations by similar users. It works by searching a large group of people and finding a smaller set of users with preferences and interests similar to a particular user.
Recommenders available in web portals such as Amazon or eBay are not always relevant and their recommendation is not necessarily timely. At the same time, conversational recommenders developed in the academic community and cited in this paper can be too intrusive, relying on routine efforts of a customer to explicitly share her preferences and intent, and also operates in too restrictive domains. What is needed is a non-intrusive, organic recommendation mode that does not require efforts from the user to share his preferences

In this paper we introduce a framework for recommendation by joining a conversation (CRS), a special case of a conversational recommendation with a focus on assisting with solving a current customer problem or a customer need to be communicated. In CRS scenarios, customers are expected to be fully aware of how and why a product or service being recommended would solve their issues.

When a user attempts to resolve a problem with a business, he is usually not in a mood to order another product or service from this business, or to spend extra on a product he has already purchased. However, frequently a recommendation for an additional product, or an extension of service helps this user save or benefit from this addition immediately or in the long run.

Typically, recommendation, advertisement and customer retention occur in different cycles of interaction with customers. However, in some customer support (CS) scenarios, recommending certain products may accelerate and simplify the problem resolution. For example, in the course of a dialogue where a customer support agent (CSA) explains how to avoid a non-sufficient fund fee (NSF), a linked saving account product can be recommended that would benefit both the user and a bank. This user would avoid NSF and the bank would have more funds in its possession.

In the course of a customer support dialogue, recommendation and advertisement need to be very relevant to customer needs and should assist in problem resolution in the way obvious to this user. In a conventional conversational recommendation, the system first gets information from the user about his needs and preferences and recommends a product after that. To manage such dialogue in an arbitrary domain, nontrivial dialogue management efforts from the system are required (Galitsky and Ilvovsky 2017, Narducci et al 2018, Sun and Zhang 2018, Galitsky 2019). Moreover, a user needs to be very patent and perform a routine activity of specifying his preferences. Neither of these is required in the CRS setting.

We consider two types of CRS scenarios:

User - Human CSA dialogue, where an automated advertisement agent tracks this dialogue and inserts its utterances with a recommendation;
User – Chatbot CS, where an automated advertisement agent and a chatbot is the same entity resolving a customer problem and providing product/service recommendations at the same time.

Both of these scenarios may occur in either problem-solving or general information acquisition settings (Kostelník 2018).

One of the main requirements for the advertising in the course of CS dialogue is that the relation to the product the user experiences problem with must be obvious, as well as the benefits to the user of relying on this new recommended product to overcome this problem.

We start with an example of casual conversation and demonstrate how an advertising utterance can naturally intervene (Fig. 1 on the top). An utterance of an CRS agent can be followed by additional factual questions CRS should be able to answer (Fig. 1 on the bottom).

A conversational advertising agent could have much commercial potential in comparison with a conventional advertising means such as random insertion in a sequence of conversation, as provided by a social advertising network (such as Facebook). Web portals like Amazon, eBay, JD, Alibaba and others are developing conversational recommender agents, but the research on this topic is very limited, and existing solutions are either based on single round conventional search or a traditional multi-round dialog system. They usually only utilize user inputs in the current session, ignoring users’ long term preferences, or just perform slot-filling, obtaining the parameters of interest from the user explicitly (Sun and Zhang 2018). Moreover, most such systems behave very differently from a human when asked for a recommendation (Galitsky 2018). Humans can quickly establish preferences when asked to make a recommendation for someone they do not know. (Christakopoulou et al 2016) address the cold-start conversational recommendation problem in an online learning setting, developing a preference elicitation framework to identify which questions to ask a new user to quickly learn their preferences. In this section, we formulate a broader advertising and recommendation problem learning user preferences implicitly from the previous utterances in an arbitrary problem-solving conversation, not just by asking explicitly about user preferences.

Applicability of the proposed recommendation setting can go beyond CS scenarios. In a daily life, people communicate with others to exchange information and improve social bonding. To exchange and share ideas, people frequently communicate with others following certain dialog flow patterns. Usually, humans do not rigidly answer questions of others and instead wait for the next question. Conversation participants may first respond to the previous context and only then propose their own questions and suggestions. In this way, people show their attention to the words of others and confirm that they are willing to continue the conversation. Daily conversations are rich in emotion. By expressing emotions, people show their mutual respect, empathy and understanding to each other, and thus improve the relationships. Daily dialogues (Li et al 2017) are also a good source for recommendation and advertisement experiments (Fig. 2), as well as dialogues mined from the web randomly (Fig. 3).

Conventional recommender systems bring together ideas from information retrieval and filtering, user profiling, adaptive interfaces and machine learning in an attempt to offer users more personalized and responsive active search systems (Galitsky 2016). Conversational recommenders guide a user through a sequence of iterations, suggesting specific items, and using feedback from users to refine their suggestions in subsequent iterations (Smyth et al 2004). The goal of recommendation systems is to help users find items of interest, and thus they must eliminate some choices from consideration while keeping others. To achieve this, recommenders must determine the interests of the current user, then present items that satisfy those interests. One of the most widely used approaches for information acquisition in recommendation systems is the use of a simple form that the user fills in. Then the recommender system presents a list of results, ordered by their predicted relevance to the user. The combined technique of information acquisition through a form and presentation in a list of results comes from in information retrieval.

In contrast, a conversational recommendation system supports a dialogue between a user and advisor as its mode of interaction (Thompson el al. 2004). For example, if the user is deciding on a restaurant for lunch, the dialogue may be as follows. Notice that it is fairly long, navigating through user preferences and particular instance. Towards the end of conversation, CRS intervenes with a recommendation and its back up (Fig. 4).

A conventional recommendation approach is good if a user is in a mood for recommendation and is ready to spend time for slot-filling. However, in many cases a user neither seeks nor expects a recommendation. In the course of a customer support session, an explicit product recommendation by CRS might be a last thing this user is expecting. That is why an unsolicited recommendation during such session must be acceptable as is, without additional clarification steps and/or slot filling for the desired parameters of the product being recommended. Recommendation during a CS session has a single chance to trigger user interest and therefore should not only be relevant but also timely.

Below (Fig. 5) is a sample session of CRS-agent in a car repair domain. In these dialogues scraped from the auto repair recommendation portal, utterances are concise and detailed, with proper explanation and argumentation patterns (Galitsky et al 2018). This is unusual for the most available dialogue training datasets nowadays. CRS agent does not have to ask the user additional questions for parameters of the entity (spark plugs) being recommended. Two utterances of the user are sufficient to determine the point of user pain and propose a business or service that might cure it.

In a regular recommendation / advertisement scenario, any popular product or the one meeting the user preferences is thought to be appropriate. Conversely, in the CRS scenario, a recommended product or service must be related to the product which is the man entity of the problem being resolved. Moreover, a feature of the product being recommended must address the problem caused by the respective feature of the problematic product being communicated with a CSA or other people.

Table 1. Examples of seed products and the ones being recommended, with features

Product (seed) that is the subject of the problem being resolved	Seed’s problematic attribute or feature that is a focus of a conversation	Product to be recommended	Recommended attribute/feature to solve problem	Search query
Checking account	No overdraft protection	Saving account	Linked with checking for overdraft protection	X for checking account with overdraft protection
Hotel @ <location>	No dogs allowed	Apartment	Dog friendly	Dog friendly apartment X @ <location>
Flight to <destination>	Ticket is not changeable	Flight insured for change of plans	Coverage for change of plans / air ticket change	Travel insurance for flight by X to <destination>
Camping tent of <brand>	Hard to pitch	Self-pitching tent	Tube frames allowing for self-pitching	Camping tent of <brand> X with self-pitching
Auto insurance from X	Does not cover roadside assistance	Additional coverage	Covering roadside assistance	Additional coverage X with roadside assistance

We show the cases of typical customer problems in varios domains:

A customer does not maintain a positive balance carefully and now wants to avoid NSF in the future;
A traveler with a pet finds himself in a hotel that does not allow dogs;
A traveler got a non-chengeable air ticket and now wants to change the flight.

In most of these cases (Table 1) the features of products and services were disclosed to customers but they did not pay enough attention. Once these customers find themselves in a situation where the limited product features presented a significant inconvenience, they contact CS and complain. This is a good time to recommend an alternative product or an addition to a service lacking the limitation that was raised as an issute in a CS session. A similar recommendation should have occurred when the user acquired this initial, seed product and service, but it is not plausible in most situations because it is hard to assess all limitations and future usability cases in advance.

The queries have a placeholder X for product/service name such as account type, accommodation name, air travel company etc. The role of this placeholder in a query is to assure the respective entity type does occur in an acceptable search result.

An ontology like the one shown in Fig. 6 is required to identify a parameter/feature of the seed entity that is the focus of a conversation with a CS. Relations in ontology are Part-of, Type-of, Same-as, Instance-of, Defines, Defined-by and others (Hoffman 2015). A feature of a product is connected with this product by Part-of, Type-of or Instance-of. Once an entity is identified in a noun phrase, the ontology is used to find an occurrence of its attribute or feature.

Fig. 7 show processing steps in the CRS component that finds a product to be recommended given an utterance with the problematic seed product. Ontology used, such as the one in Fig. 6, is shown in top-right corner.

A number of studies including (Berkovsky et al. 2012) demonstrated that explanation and persuasion are two important characteristics for convincing users to follow the recommendations.

We used a traditional advertisement format for the irrelevant and unpersuasive examples. A good example is a free-format text that includes a recommendation as well as its argumentative back up, an explanation why this product would solve a customer problem, as described in dialogue (Fig. 9). Negative examples, in particular, rely on imperative form of verbs that is heavily used in conventional advertisement.

To be a good recommendation, it needs to relate to the seed product and to its features and attributes which are the subjects of the CSA conversation. In addition, discourse structure of the recommendation text matters (Fig. 10)

Discourse tree representation for a recommendation allows judging on its quality. If rhetorical relations of Explanation, Cause, Enablement are recognized in recommendation text (Galitsky and Ilvovsky 2019a, Fig. 9, shown in bold), then there is a higher chance that this recommendation is reasonable, persuasive and well-argued. Recommendation with a discourse tree that contains only default rhetorical relations such as Elaboration and Join would not be as good. Moreover, discourse representation of the recommendation must match in terms of argumentation the discourse representation of the problem description of the product by customer. In other words, these two discourse representations must be in a rhetorical agreement. A generalized example of a proper correlation between the previous utterances about the seed product P and recommendation R is shown in Fig. 10.

To assure a recommendation makes sense to a user, it needs to be backed up by an argument. To find a textual recommendation that will be well perceived by the user, this recommendation should forms a well backed up claim where the utterances in the dialogue are premises. This is a special case of argument mining task that can be deﬁned as analyzing a discourse on the pragmatics level and employing a certain argumentation theory (Tuolmin 1958) to model available textual data. An argument mining framework includes two steps: argument extraction and argumentation relation prediction. The latter step may be based on learning or reasoning, or both.

Arguments’ extraction is the identiﬁcation of arguments within the input natural language text. This step may be further split in two different stages such as the detection of argument components (e.g., claim, premises) and the further identiﬁcation of their textual boundaries. Many approaches have recently been proposed to address such task, that adopt different statistical learning methods (Mochales and Moens, 2011; Lippi and Torroni, 2016; Bar-Haim et al., 2017)

Relations’ prediction determines what are the relations holding between the arguments identiﬁed in the ﬁrst stage. This is an extremely complex task, as it involves high-level knowledge representation and reasoning issues. The relations between the arguments may be of heterogeneous nature, like attacks and supports. They are used to build the argument graphs, in which the relations connecting the retrieved arguments (i.e., the nodes in the graph) correspond to the edges. Different learning methods have been employed to address this task, from standard SVMs to Textual Entailment (Cabrio and Villata, 2013). This stage is also in charge of predicting, in structured argumentation, the internal relations of the argument’s components, such as the connection between the premises and the claim.

A modiﬁed Toulmin’s model contains ﬁve argument components, namely:

claim, premise, backing, rebuttal, and refutation.

When annotating a document, any arbitrary token span can be labeled with an argument component; the components do not overlap. The spans are not known in advance and the annotator thus chooses the span and the component type at the same time. All components are optional (they do not have to be present in the argument) except the claim, which is either explicit or implicit. If a token span is not labeled by any argument component, it is not considered as a part of the argument and is later denoted as none (this category is not assigned by the annotators).

Modiﬁed Toulmin’s model used for annotation of arguments with an instantiated example of a CS dialogue on NSF (Fig. 11). The arrows show relations between argument components; the relations are implicit and inherent in the model. By contrast to the example of original Toulmin’s model, no connective phrases are attached to the relations (such as so, unless, etc.).

We show how to analyze argumentation relations in text relying on the discourse representation, so we will not repeat it in this paper. We refer that reader to that section for the algorithm for how to verify that one text defeats the claims from another text. On the contrary, in case of the pair of products P and RP, a recommendation for RP must be supported by the customers’ expression of her needs and problems in P.

Once a recommendation utterance is delivered, the user may choose to continue conversation with Ad-agent. Then the following algorithm is applied (Fig. 12).

To build a conversational grammar for dialogue management, we introduce the notion of adjacency-pair, sequences of two utterances that are:

adjacent (unless separated by an insertion sequence);
produced by different speakers;
ordered as a first part (which we will refer to also as the initiative) and a second part (which we will refer to also as the response);
typed, so that a particular initiative requires a certain type or range of types of response.

Adjacency-pairs are question-answer, greeting-greeting, or offer-acceptance/decline. Where there is a range of potential responses to an initiative (as with offer-acceptance/decline), a ranking operates over the options setting one response as most preferred (in the sense of normal, more usual) and others as less preferred (Bridge 2002). Other than preferred responses are usually lengthy and syntactically more sophisticated. Having produced a first part (utterance) of some pair, the current utterance stops and it is expected that the next speaker will produce one of the allowable second utterance of the same pair. The second part will often proceed right away. However, there frequently appear sequences of turns that intervene between the first and second utterances of a pair; the second utterance is in a holding pattern during the insertion sequence. An insertion sequence will be topically related to the pair of utterances it interrupts and may be used to determine the desired features for providing the second utterance of the original pair. Insertion sequences typically contain further adjacency-pairs, which may themselves be interrupted by further insertion sequences.

We use Prolog notations for the dialogue grammar: variables are capitalized (Fig. 13).

A number of decisions have been made in writing the grammar that constrains the contributions that the system can make to the dialogue. In particular, the system cannot abort pairs: rules 5 and 6 apply only to the user.

Recently, the emergence of ChatGPT has significantly advanced NLP tasks by enhancing the capabilities of conversational models. Nonetheless, the application of ChatGPT in the the CRS domain has not been investigated.

ChatGPT can be used as a general-purpose recommendation model to explore its potential for transferring extensive linguistic and world knowledge acquired from large-scale corpora to recommendation scenarios. Following Liu et al (2023), we form a set of prompts and did not fine-tune ChatGPT, relying only on the prompts themselves to convert recommendation tasks into natural language tasks. A few-shot prompting to incorporate interaction information that contains user potential interest to assist ChatGPT to better handle user needs and preferences

The architecture of embedding ChatGPT into conversational recommendation is shown in Fig. 14:

A number of prompts are constructed based on the previous utterances of conversing users, based on their recognized epistemic and emotional states;
These prompts are used as inputs for ChatGPT that produces the recommendation results according to the requirements specified in the prompts.
The output from ChatGPT is checked and refined by the refinement module, and the refined results are returned to the user as the final recommendation results

For example, let us consider a dialogue between two humans and CRS in auto repair domain (Fig. 15)

Each prompt includes three components: task description, behavior injection, and format indicator.

The task description is utilized to adapt recommendation tasks to natural language processing tasks.
The behavior, epistemic and emotional state injection is designed to assess the impact of few-shot prompting, which incorporates user-item interaction to aid ChatGPT in capturing user preferences and needs more effectively.
The format indicator serves to constrain the output format, making the recommendation results more comprehensible and assessable.

Comprehensive experimental results on Amazon Beauty dataset show that ChatGPT has achieved promising results in certain tasks and is capable of reaching the baseline level in others. We conduct human evaluations on two explainability-oriented tasks to more accurately evaluate the quality of contents generated by different models. And the human evaluations show ChatGPT can truly understand the provided information and generate clearer and more reasonable results. We hope that our study can inspire researchers to further explore the potential of language models like ChatGPT to improve recommendation performance and contribute to the advancement of the recommendation systems field.

Figure 16 illustrates the overall architecture of the Customer Resolution System (CRS). The primary objective of CRS is to monitor the conversation and detect instances where the customer is likely to abandon the problem resolution process or remains dissatisfied even after the issue is resolved. This monitoring relies on analyzing the emotional and sentiment profiles of the dialogue. When such an utterance is identified, CRS extracts a noun phrase from it and proceeds to identify the corresponding product name and its associated features.

To perform entity extraction, CRS utilizes Stanford NLP, which is enhanced by incorporating product-specific entity rules and leveraging product-specific lookups, such as the eBay product catalogue. Additionally, the system can cross-verify product-related named entities by consulting the eBay product search API.

Next, a search query is generated using the product name and its associated features. This query is used to initiate a search process. The search results produce a list of potential candidates, which undergo filtering based on the requirements of proper argumentation and discourse coordination.

To implement this filtering, argument mining and reasoning techniques are employed. These techniques verify whether the recommendation, presented as a claim, is logically supported by the preceding customer utterance. This logical support ensures that the recommendation is convincing for the customer.

Once we know the product to recommend, user intention, mood and emotional state, as well as an optimal recommendation format, a prompt is formulated. Once the prompt and the previous utterance of the users are passed to ChatGPT, it formulates the recommendation utterance (Fig. )

Furthermore, rhetorical agreement is validated by examining the coordination between the discourse trees of the previous customer utterances and the discourse tree of the candidate recommendation text. This analysis helps to ensure that the recommendation aligns with the overall conversation and maintains coherence.

Recognizing not only user’s preferences using entities but also the intent (Allen and Perrault 1980) is essential to determine a right time to invoke the recommendation engine. While current major chatbots and virtual assistant platforms vary, most use the same general principle, what we might call the intent-entity-context-response principle. In this principle:

The user’s intent, or conversational action, is determined largely through NL classification;
The entities in the user’s utterances are extracted through keyword matching;
The context of the current and prior turns in the conversation are preserved in variables and all of these are used by a dialogue manager to determine the type of the next utterance;
The system’s response back to the user, in text or voice.

In addition to physically hosting modules in the cloud environment, the role of Service Engine is to manage the functions between the conversation services, the recommendation engine and the web-conversation user interface.

We obtain human-human dialogues from the Customer Complaints and Car Repair datasets. For the first dataset, we obtain recommendations online from websites like www.bankrate.com and www.bloomberg.com. The Car Repair dataset is scraped from www.2carpros.com. This dataset contains conversations between people repairing their cars and auto repair professionals.

We also acquire recommendation sources from Yelp on restaurants and services such as repair and tuition. For book recommendations we used Amazon/LibraryThing (A/LT) dataset available at www.librarything.com/groups. For blogs and forums which can potentially be subject to CRS we relied on www.2carpros.com, www.immihelp.com, www.talkhealthpartnership.com and blog.feedspot.com.

To get closer to the CSA conversation setting, we selected Relational Strategies in Customer Service Dataset that is a collection of travel-related customer service data from four sources. The conversation logs three commercial customer services and the Airline forums on TripAdvisor.com. For a special case of conversations related to overall product opinion, we employ the Customer Support on Twitter dataset. It includes over 3 million tweets and replies from the biggest brands on Twitter. The datasets to evaluate CRS are enumerated in Table 2.

For scraped and indexed data, we use our own search for products, and for web data we either use APIs of a particular source or search this source via Bing API.

We use various source of dialogues:

Conversational data sets;
Scraped from online forums;
Cached search results from specific APIs.
A tool for automated construction of dialogues from the text (Galitsky 2019)

Table 2. Characteristics for each data source in evaluation of relevant recommendation

Source name	# in the dataset	Problem resolution in a dialogue	Origin of data	Source of recommendations
Finance	2200	Yes	my3cents.com bankrate.com	Web search of Bloomberg, Fidelity, Bankrate for financial products
Auto repair	9300	Yes	2carpros.com	Web search for services
Sports shopping	2740	No	REI and L.L.Bean data from RichRelevance.com	Internal API for product search
Home products shopping	3100	No	Walmart, HD Supply, OfficeDepot data from RichRelevance.com	eBay product search API
Home-related services			Yelp reviews	Yelp API
Travel	2430	No	zicasso.com/travel-reviews, tripadvisor.com reviews Airline forums on TripAdvisor.com	Tripadvisor.com
Daily Dialogues	2000	Sometimes	(Li et al 2018)	Yelp API
Genuine human dialogues	2000	Sometimes	(Li et al 2018) ENRON email thread Reddit discourse dataset (Logacheva et al. 2018)	Yelp API, eBay product search API, Amazon product search, TripAdvisor.com, Bing Azure Forum search, Bing AzureWeb search, Yandex XML web search, GPT4
Dialogues constructed from blogs, forums and other sources	2200	No	2carpros.com, immihelp.com, talkhealthpartnership.com, blog.feedspot.com librarything.com/groups. Book recommendations
Dialogues generated from text using doc2dialogue tool	3300	No	Web mining of documents on the web in Engineering, Finance, Legal, Health, Travel

Documents in the domain of Finance, Travel, Sports are converted into FAQ-type question answering pairs. The Reddit discourse dataset (Zhang et al., 2017) is manually annotated with dialog-acts via crowd sourcing. The dialog-acts comprise of an answer, a question, humor, an agreement, a disagreement, an appreciation, a negative reaction, an elaboration and an announcement. It comprises conversations from around 9000 randomly sampled Reddit threads with over 100000 comments and an average of 12 turns per thread.

The most frequently communicative actions (CAs) in the corpus are Statement-non-opinion, Acknowledge, Statement-opinion, Agree/Accept, Abandoned or Turn-Exit, Appreciation, Yes-No-Question, Nonverbal, Yes answers, Conventional-closing. Utterances with underscored CAs are primary candidates for CRS to act.

Overall, it is hard to find a dataset for CRS because most of the conversational datasets are very specific to movies, transportation, restaurants, and points of interest. These datasets are extremely limited to domain and conversation style. We do not use dialogue data from public task-oriented dialogues since the conversations are usually very primitive and are confined to a very narrow domain such as hotel and transportation reservation in a given metropolitan area (Budzianowski et al 2019, Schulz et al 2017, Facebook Babi 2019). It is quite difficult to recommend any product or service for such dialogue datasets developed to train chit-chat or task-oriented conversations. However, it is not the case for the datasets (Li et al 2018, Logacheva et al. 2018).

That is why the dialogue generation tool from an arbitrary document (Galitsky et al 2019) is required to generate dialogues in multiple domains to test CRS extensively. As a result of using this tool, our dialogue datasets are more natural and authentic human conversations than some popular datasets like Twitter Dialog Corpus (Ritter et al., 2011) and Chinese Weibo datasets, constructed by posts and replies on social networks, which are noisy, short and strongly deviate from real conversations.

To verify each component automatically, we develop a pair of implementations for each verification component and assess one against another for the full-size dataset. For each component, the first integration architecture in this pair has this component and the second architecture does not. Each implementation pair is spot-tested individually manually, on a significantly reduced dataset. This approach is similar to a keyword-based assessment of chatbot performance such as BLUE which is not always accurate but scales well.

To achieve automated evaluation it, we try to design an automated assessment algorithm and evaluate the quality of assessment manually. Assessment algorithm should be different from the recognition algorithm for a given component. We separately estimate the accuracy of the assessment algorithm and of the component itself.

The assessment algorithm is estimated manually with respect to how accurately it is measuring the performance of a system component. The assessment measurement error e = 1 – F1(assessment algorithm₎. We consider the assessment algorithm satisfactory for most evaluations if e <20%. Hence e is estimated manually but allows for an automated assessment of a system component with certain measurement error. Once this manual assessment is done, we can apply the assessment algorithm to the system component, measuring its accuracy as F1(component)±e.

Table 3 Accuracies of each component as well as overall recommendation relevance assessment, %

Source name	Correct dialogue turn for recommendation	Correctness of entity extraction from dialogue	Extracted product entity is properly matched in recommendation	Acceptable argumentation by the recommendation	Correct prompt is constructed for ChatGPT	ChatGPT generates the correct recommendation utterance	Proper recommendation discourse	Overall meaningful recommendation
Finance	91.3	94.5	91.2	73.2	87.7	86.8	75.3	78.9
Auto repair	88.4	96.0	92.6	78.1	84.7	85.8	74.8	78.3
Sports shopping	89.6	92.9	90.4	83.0	88.0	86.4	73.9	77.4
Home products shopping	90.3	92.1	94.7	80.7	87.6	86.7	75.4	78.7
Home-related services	89.3	93.7	91.7	77.7	86.5	83.3	76.5	77.3
Travel	90.8	92.7	93.6	77.1	85.4	84.2	72.4	78.2
Daily Dialogues	88.4	89.3	92.0	78.2	86.7	82.0	71.7	76.0
Genuine human dialogues	89.3	91.6	88.3	77.3	84.2	78.2	74.2	78.2
Dialogues constructed from blogs, forums and other sources	90.4	92.7	90.7	75.6	83.7	81.4	70.7	76.4
Dialogues generated from text using doc2dialogue tool	87.2	90.5	88.3	73.4	82.8	81.0	75.8	75.1

Table 3 shows the results of evaluation of the CRS pipeline. The CRS components correspond to the table columns from left to right. We start with assessment of the best utterance to join the dialogue, proceed to product entity extraction from this utterance to its matching with available product along with the winning attribute (the one which is going to persuade the customer that the product being recommended would solve his problem). We then assess argument extraction and reasoning related to this argument. After that, we measure the performance of prompt generation for ChatGPT and the recommendation utterance generation itself before the overall recommendation results.

We do not estimate recommendation recall since we assume there is an unlimited spectrum of products and services to recommend. If no utterance yields a recommendation we exclude the conversation from our evaluation. Once a product entity is correctly extracted from a user dialogue, there should be respective product on the web or in the available index that solves the customer problem. We show the precision for each processing step as a percentage of the correct cases.

Recommendation by joining a conversation turns out to have a high overall relevance and appropriateness to needs of customers (right column in Table 3). The accuracy range of 66-73% shows that almost three quarters of recommendations should not cause user irritation and instead encourage a user to buy a product that would address a problem raised in the conversation. Although we do not assess an actual conversion rate of CRS, one can see that this form of recommendation and advertisement is least intrusive and has the highest instant relevance in comparison with other conversational recommendation means. Three greyed bottom rows in Table 3 show the datasets where we access the applicability of dialogue generation in comparison with genuine dialogues.

Accuracies of each component vary from domain to domain by less than 12% due to different linguistic and logical complexity of dialogues, product searches and argumentation analysis. Bottom greyed three rows show that genuine human dialogues are a bit more complex than the artificial ones obtained from documents (although the latter has more formal, professional language). One of the conclusions here is that dialogue generation by a doc2dialogue tool is an acceptable way to train and evaluate conversational systems. Three greyed bottom rows in Table 3 show the datasets where we access the applicability of dialogue generation in comparison with genuine dialogues.

9.1 Related Systems

In this section, we enumerate major vendors of conversational platforms that can potentially be used for recommendations

Amazon Personalize helps businesses create personalized experiences for their customers. It can be used to recommend products, services, and content to customers based on their past behavior, interests, and demographics (Fig. 17).

Google Cloud Dialogflow is a natural language processing platform that can be used to build conversational interfaces. It can be used to create chatbots, virtual assistants, and other conversational applications including recommendation engines

IBM Watson Assistant is a natural language processing platform that can be used to build conversational interfaces. It can be used to create chatbots, virtual assistants, and other conversational applications (Fig. 18).

For comparison with our proposed architecture, we show a conventional CRS architecture from IMB Watson. Figure 19 shows a main platform and components in the middle box, and input data sources and the front end components, such as user interface design, in bottom and top boxes respectively. Conversation service is a main framework for developing chatbots by understanding the user’s utterance. For the development of a recommendation system on top of a dialogue system, we learned three important aspects that must be considered thoroughly. Context variables must be set by the dialogue service to accommodate variations of entities. Once context variables are set up, then entities or key words can be detected from the conversation. For the development of destination recommendations, the preferences of users like nightlife, restaurants, beaches, are created and captured as entities in the dialogue system based on context variations.

Microsoft Bot Framework is a development platform that can be used to build conversational interfaces. It can be used to create chatbots, virtual assistants, and other conversational applications such as recommendation (Fig. 20).

Rasa is an open-source framework that can be used to build conversational interfaces. It can be used to create chatbots, virtual assistants, and other conversational applications.

These are just a few of the many conversational recommendation systems available. The best system for a particular business will depend on the specific needs of the business.

9.2 Related approaches and components of recommendation system

In their work, Sun and Zhang (2018) introduced a method for representing a user's conversation history as a semi-structured user query consisting of facet-value pairs. This query is dynamically generated and updated by a belief tracker, which analyzes the user's natural language utterances at each step of the conversation. To facilitate recommendation tasks, a set of machine actions specifically designed for recommendation agents is selected. A deep policy network is then trained to determine the most appropriate action to be taken by the agent at each step. These actions can involve requesting the value of a particular facet or providing a recommendation.

The authors developed a personalized recommendation model that leverages both the user's past ratings and the user query collected during the ongoing conversational session. This model is used to make rating predictions and generate personalized recommendations for the user. The conversational system actively gathers user preferences by asking relevant questions until a sufficient amount of user preference information is obtained. Once enough data is collected, personalized recommendations are provided to the user. To evaluate their framework, the authors conducted both simulation experiments and real online user studies, demonstrating the effectiveness of their proposed approach.

Argument mining techniques enable the extraction of underlying motivations expressed by consumers in reviews, offering more insightful information beyond simple statements like "I like" or "I don't like product A." The process of argument detection on the TextCoop platform was explained by Villalba and Saint-Dizier (2012). In a study by Ibeke et al. (2017) on the El Capitan dataset, the task of mining contrastive opinions was formulated using a unified latent variable model, with manual annotations of topic and sentiment labels on the reviews.

However, analyzing arguments within user reviews can be challenging due to the ambiguous relationship between argument mining and sentiment analysis. This challenge arises because sentiments expressed about specific aspects of a claim (in favor or against the product) sometimes encompass the reasons behind considering the product as either good or bad. Thus, sentiments can intertwine with the underlying argumentation in user reviews.

Argument Mining is intricately linked to prominent topics in the field of AI, including deep learning, fact-checking, misinformation detection, explanations of machine decisions, and more. Argument mining plays a pivotal role in various applications such as fake news detection, as predicting the interactions and attacks between arguments serves as a fundamental component. Furthermore, argument mining can provide insights into how machines gather and utilize information to make decisions, contributing to the field of explainable AI.

Argument mining also holds potential in domains such as medicine, where it can aid in identifying the relevant information necessary for reasoning about randomized clinical trials. In the realm of politics, argument mining can automate the identification of fallacies and unfair propaganda, enhancing the ability to analyze and evaluate political discourse. Additionally, argument mining can contribute to the prevention of cyber-bullying by supporting the detection of recurring attacks targeted at individuals or entities. These diverse scenarios highlight the broad applicability and significance of Argument Mining across various domains.

Approaching from a dialogical perspective, Cabrio and Villata (2012) expanded upon an argumentation framework initially proposed by Dung (1995). This framework represents arguments using a graph structure and offers a reasoning mechanism for resolving accepted arguments. To identify support and attack relations, they leveraged prior research on textual entailment, specifically utilizing the off-the-shelf EDITS system.

The authors collected test data from a debate portal called Debatepedia, covering nineteen different topics. The evaluation focused on measuring the acceptance of the "main argument" using automatically recognized entailments. The results showed an F1 score of approximately 0.75, indicating the effectiveness of their approach. It is worth noting that their work focuses on micro-level argumentation, in contrast to Dung's model, which provides an abstract framework designed for modeling dialogical argumentation.

Several studies have explored the concept of persuasiveness, particularly in the context of advertising. Schlosser (2011) examined the persuasiveness of online reviews and discovered that presenting both sides of an argument may not always be more helpful and can sometimes be less persuasive than presenting only one side. In a computational model described by Miceli et al. (2006), an attempt was made to integrate both emotional and non-emotional aspects of persuasion.

In a specific study, around twenty texts were assigned a level of persuasiveness (out of 100 texts that were manually preselected), and later, four of these texts were analyzed in detail by (2001) to compare how experts and students perceive persuasion. Bernard et al. (2012) investigated children's perception of discourse connectives, specifically focusing on the use of "because" to link statements in arguments. The study revealed that both 4- and 5-year-olds and adults showed sensitivity to the presence of connectives in arguments.

In their study, Bracewell et al. (2013) explored a broad perspective on social media dialogs. They put forth a collection of fifteen social acts, including actions like agreement, disagreement, and supportive behavior. These social acts aimed to infer the social goals of participants engaged in the dialogue. The authors also presented a semi-supervised model designed for the classification of these social acts.

Inspired by research in psychology and organizational behavior, the proposed social act types sought to capture various aspects of dialog understanding. By incorporating insights from these fields, the study aimed to enhance the comprehension of social interactions within social media dialog

The inclusion of advertising within a dialogue is closely associated with the concept of dialogue marketing, which originated around two decades ago. Dialogue marketing involves companies actively engaging with willing consumers in an ongoing conversation to foster long-term relationships. It encompasses various marketing activities that utilize media to establish interactive connections with individuals. The primary objective is to elicit a personalized and measurable response from the recipient, aiming for a meaningful and individualized interaction (Jaffe, 2008).

Using the existing data at their disposal, marketing personnel within companies extend invitations to groups of potential consumers, encouraging them to establish connections with the company. This engagement process serves as a mutually beneficial interaction for both consumers and the company. Marketers seize these opportunities to gather valuable data, which is then utilized to enhance the customization of marketing messages and personalize the overall consumer experience, catering to specific market segments.

In return for sharing their opinions, purchasing habits, product preferences, and other relevant information, consumers are offered incentives such as discounts, tips, free trials, and tailored messaging from the company. This exchange allows consumers to enjoy perks while enabling the company to better understand and meet their needs and preferences.

In order to achieve success, dialogue marketing necessitates that businesses possess a clear understanding of their unique value proposition and how it resonates with consumers. They should be able to identify their primary customers as well as potential customers. Moreover, businesses need to develop appropriate messaging and engagement methods to effectively connect with their target audience. It is crucial to implement a well-defined plan that enables reaching out and establishing connections with the right consumers while fostering long-term relationships.

Measurement plays a vital role in dialogue marketing as it allows businesses to track and assess the effectiveness of their marketing and sales efforts. This feedback-driven approach enables them to refine their strategies based on the received insights. Dialogue marketing consists of four essential stages that seamlessly integrate advertising, public relations, and marketing into a cohesive strategy. Various vendors, including advertising agencies, marketing and branding companies, digital printers, data specialists, social media experts, and loyalty and referral program designers, contribute to the successful execution of dialogue marketing initiatives.

Marketers have recognized the effectiveness and efficiency of dialogue marketing when combining Web 2.0, social media, personalized microsites, variable data printing, and blogs. By directing their marketing efforts towards individuals who are already receptive to engagement and providing them with opportunities to connect on their own terms, businesses can achieve higher brand loyalty, generate more referrals, boost cross-sales, and encourage repeat business.

The integration of Web 2.0 technologies and social media platforms allows marketers to engage with their target audience directly, fostering meaningful interactions. Personalized microsites, variable data printing, and blogs further enhance this personalized approach, tailoring messages and experiences to individual preferences and needs. By focusing on individuals who are already open to engagement, businesses can build stronger connections, leading to increased brand loyalty.

Moreover, dialogue marketing creates opportunities for customers to become brand advocates, referring others to the business and driving additional sales. This approach also facilitates cross-sales by identifying and offering complementary products or services to existing customers. The result is a positive feedback loop of engagement, loyalty, and repeat business.

In summary, dialogue marketing, empowered by Web 2.0 technologies and social media, along with personalized microsites, variable data printing, and blogs, proves to be a valuable investment for marketers, enabling them to maximize their marketing dollars while cultivating brand loyalty, generating referrals, increasing cross-sales, and fostering repeat business.

A relationship dialogue is a process of reasoning together in order for two or more parties to develop a common knowledge platform (Grönroos 2000). Relationship marketing is facilitated provided that this knowledge platform enables a supplier to create additional value for its customers on top of the value of the goods and services which are exchanged in the relationship. For a relationship dialogue to emerge, in an on-going process the communication effects of planned communication efforts and of product and service-based interactions between a supplier and its customers have to support each other. Then the required extra value of the relationship is created and favorable word of mouth follows.

When focused on customer - company relationships, conversational system design needs to be closely aligned with the discipline of CRM: Sales, Service and Marketing. A strong focus should be placed on outcomes, not on records or process management. To make user experience better, implementation of conversational systems must take into account the following (Greenberg 2018):

1) Conversations are multifaceted. Getting to the right conversation is about accuracy; having that conversation is about precision. Thus, determining which conversation to have and how to best have it are distinct exercises. While equally important, both are complex. Therefore, practitioners must be clear about which problem they are trying to solve.

2) Conversations reduce friction. They are familiar and easy to describe. Active dialog allows for clear communication and the ability to course correct. Conversations can take place in-person or over video, voice, web chat, and SMS/messaging. Conversations are synchronous, asynchronous, and may come in bursts. Technological definitions may conflict with human nature.

3) Conversations have a mission. They create optimal experiences and prove value to both participants of the conversation. This involves supporting and enhancing communication between two people, a person and a system/brand, or two systems. The objective is to have the best, informed, value-based and outcome-driven conversation possible. This is Conversational Experience

There have been many works emphasizing the importance of interactivity in recommenders so that the user has a more active role over the recommendations. It includes a critique-based recommendations (Chen and Pu 2012), constraint-based (Felfernig et al 2011), dialog, and utility-based recommenders). However, these studies employ a prior modeling of the items’ features, preventing the flexibility in adaptation to different recommendation domains.

Most conventional recommendation techniques are domain and task- specific. Therefore, specific content and user data is required to train a particular model for a given application scenario under a deep learning approach. This cannot leverage an efficient generalization ability. To tackle this problem, the deep learning recommendation community has evolved towards implementing Large Language Models in recommendation scenarios since they have demonstrated strong auto-adjustment capability to improve the performance of downstream NLP tasks significantly. To effectively convert user interaction data into text sequences, a variety of prompts (Zhang et al 2023) is developed to represent user interaction scenarios as sequences of text tokens.

We conclude that CRS delivering relevance of recommendations close to 75% is an efficient and effective way to provide organic recommendation in a less intrusive mode than other forms of conversational recommendation. CRS does not require a user spend time on sharing personal information which can potentially be wasted if a recommended product ends up not been needed. Being superior to the conversational recommendation mode with the focus on information gathering from a user, CRS also significantly exceeds a relevance of conventional recommendation which is usually not associated with a particular scenario of user decision making. Focusing recommendation on the current user point of pain dramatically increases recommendation relevance and meaningfulness as well.

We conclude the paper by enumerating the observed features of CRS:

• Recommendation by joining a conversation turns out to have a high overall relevance and appropriateness to the needs of customers;

• The accuracy range of 68–74% shows that at least ¾ of recommendations should not cause user irritation and instead encourage a user to buy a recommended product;

• In most cases the recommended products and services indeed address a customer problem raised in conversation;

• Explainable AI – compliant: it is clear why this product is needed;

• This form of recommendation and advertisement is least intrusive as the CRS utterance can be ignored.

One of the tasks of a future study is to evaluate an actual convergence rate of the CRS advertisement mode.

Mochales R and Moens M-F (2011) Argumentation mining. Artiﬁcial Intelligence and Law, 19(1):1–22.
Lippi M and Torroni P (2016) Argument mining from speech: Detecting claims in political debates. In AAAI, 2979–2985.
Bar-Haim R, Indrajit Bhattacharya, Francesco Dinuzzo, Amrita Saha, and Noam Slonim (2017) Stance classiﬁcation of context-dependent claims. In EACL, 251–261.
Jaffe J (2008) Join the Conversation: How to Engage Marketing-Weary Consumers with the Power of Community, Dialogue, and Partnership John Wiley & Sons. New Jersey US.
Thompson CA, Mehmet H. Göker, and Pat Langley (2004) A personalized system for conversational recommendations. J. Artif. Int. Res. 21-1 , 393-428.
Grönroos C (2000) Creating a Relationship Dialogue: Communication, Interaction and Value. The Marketing Review, V1, N1, pp. 5-14(10).
Logacheva V., Burtsev M., Malykh V., Polulyakh V., Seliverstov A. (2018) ConvAI Dataset of Topic-Oriented Human-to-Chatbot Dialogues. In: Escalera S., Weimer M. (eds) The NIPS '17 Competition: Building Intelligent Systems. The Springer Series on Challenges in Machine Learning. Springer, Cham
Cabrio E and Serena Villata (2013) A natural language bipolar argumentation approach to support users in online debate interactions. Argument & Computation, 4(3):209–230.
Bernard S, Hugo Mercier, and Fabrice Clément (2012) The power of well-connected arguments: early sensitivity to the connective because. Journal of experimental child psychology, 111(1):128–35.
Murphy PK (2001) What makes a text persuasive? Comparing students’ and experts’ conceptions of persuasiveness. International Journal of Educational Research, 35(7-8):675–698.
Dagan I, Bill Dolan, Bernardo Magnini, and Dan Roth. (2009) Recognizing textual entailment: Rational, evaluation and approaches. Natural Language Engineering, 15(Special Issue 04):i–xvii.
Dung P-M (1995) On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artiﬁcial Intelligence, 77(2):321 – 357.
Li Y, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu (2017) DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset. IJCNLP.
Schlosser A E. (2011) Can including pros and cons increase the helpfulness and persuasiveness of online reviews? the interactive effects of ratings and arguments. Journal of Consumer Psychology, 21(3):226–239.
Miceli M, Fiorella de Rosis, and Isabella Poggi. (2006) Emotional and non-emotional persuasion. Applied Artiﬁcial Intelligence, 20(10):849–879.
Greenberg P (2018) Conversational experiences: Building relationships one conversation at a time. ZDNet https://www.zdnet.com/article/conversational-experiences-building-relationships-one-conversation-at-a-time/
Sun Y and Zhang Y (2018) Conversational Recommender System. SIGIR '18 The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, 235-244 Ann Arbor, MI, USA.
Christakopoulou K, Filip Radlinski, and Katja Hofmann (2016) Towards Conversational Recommender Systems. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16). ACM, New York, NY, USA, 815-824.
Chen L and P. Pu (2012) Critiquing-based recommenders: survey and emerging trends. In User Modeling and User-Adapted Interaction, 22(1-2):125–150.
Felfernig A, G. Friedrich, D. Jannach and M. Zanker (2011) Developing Constraint-based Recommenders. In Recommender systems handbook, 187–212.
Villalba MPG and Patrick Saint-Dizier P (2012) A framework to extract arguments in opinion texts. IJCINI, 6(3):62–87.
Ibeke R, Chenghua Lin, Adam Z. Wyner, and Mohamad Hardyman Barawi (2017) Extracting and understanding contrastive opinion through topic relevant sentences. In IJCNLP, 395–400.
Anelli VW, Pierpaolo Basile, Derek Bridge, Tommaso Di Noia, Pasquale Lops, Cataldo Musto, Fedelucio Narducci, and Zanker M(2018) Knowledge-aware and conversational recommender systems. In Proceedings of the 12th ACM Conference on Recommender Systems (RecSys '18). ACM, New York, NY, USA, 521-522
Narducci F., de Gemmis M., Lops P., Semeraro G. (2018) Improving the User Experience with a Conversational Recommender System. In: Ghidini C., Magnini B., Passerini A., Traverso P. (eds) AI*IA 2018 – Advances in Artificial Intelligence. AI*IA 2018. Lecture Notes in Computer Science, vol 11298. Springer, Cham.
Budzianowski, P., Wen, T., & Gasic, M. (2018). Research data supporting "MultiWOZ - A Large-Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modelling" [Dataset]. https://doi.org/10.17863/CAM.27632
Hou Y(2019) Task Oriented Dialogue Dataset Survey. https://github.com/AtmaHou/Task-Oriented-Dialogue-Dataset-Survey
Hoffman C (2019) Financial Report Ontology. http://www.xbrlsite.com/2015/fro/.
Bridge D (2002) Towards Conversational Recommender Systems: A Dialogue Grammar Approach. Proceedings of the Workshop in Mixed-Initiative Case-Based Reasoning, Workshop Program at the Sixth European Conference in Case-Based Reasoning, 9-22.
Facebook Babi (2019) https://research.fb.com/downloads/babi/
Galitsky B (2016) Providing personalized recommendation for attending events based on individual interest profiles. Artif. Intell. Research 5 (1), 1-13
Galitsky B and Dmitry Ilvovsky (2017) Chatbot with a discourse structure-driven dialogue management. EACL System Demonstrations
Galitsky B (2018) Building dialogue structure from discourse tree of a question. Workshops at the Thirty-Second AAAI Conference on Artificial Intelligence
Galitsky B, D Ilvovsky, SO Kuznetsov (2018) Detecting logical argumentation in text via communicative discourse tree. Journal of Experimental & Theoretical Artificial Intelligence 30 (5), 637-663
Galitsky B, Dmitry Ilvovsky and Elizaveta Goncharova (2019) On a Chatbot Conducting Dialogue-in-Dialogue. SIGDIAL Demo paper. Stockholm, Sweden, 118-121.
Galitsky B (2019) Discourse-level Dialogue Management. In Developing Enterprise Chatbots, Springer, Cham, Switzerland.
Gao Y, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-REC: Towards Interactive and Explainable LLMs-Augmented Recommender System. arXiv preprint arXiv:2303.14524 (2023).
Makhalova T, Ilvovsky D, Galitsky B (2019) Information Retrieval Chatbots Based on Conceptual Models
International Conference on Conceptual Structures, 230-238
Tseng B-H, Budzianowski P,Yen-Chen Wu, Gasiˇc M (2019) Tree-Structured Semantic Encoder with Knowledge Sharing for Domain Adaptation in Natural Language Generation. SIGDIAL. Stockholm, Sweden, 155-164.
Schulz H, Jeremie Zumer, Layla El Asri, and Shikhar Sharma. (2017) A frame tracking model for memory-enhanced dialogue systems. CoRR, abs/1706.01690.
Berkovsky S, Jill Freyne, and Harri Oinas-Kukkonen (2012) Influencing individually: fusing personalization and persuasion. ACM Transactions on Interactive Intelligent Systems (TiiS), 2(2):9.
Ritter A, Colin Cherry, and William B Dolan (2011) Data-driven response generation in social media. In EMNLP, 583–593.
Toulmin SE (1958) The Uses of argument. Cambridge University Press, Cambridge, UK.
Smyth B, McGinty L, Reilly J, McCarthy K(2004) Compound Critiques for Conversational Recommender Systems. IEEE/WIC/ACM International Conf. on Web Intelligence. IEEE. Beijing, China.
Kostelník P, Pisařovic I, Muron M, Dařena F, Procházka D (2019). Chatbots for Enterprises: Outlook. Acta Universitatis Agriculturae et Silviculturae Mendelianae Brunensis. 67. 1541-1550.
eBay (2020) eBay product search API. https://developer.ebay.com/api-docs/buy/browse/resources/item_summary/methods/search
Conversational Recommender System
Pramod D, Bafna P (2022) Conversational recommender systems techniques, tools, acceptance, and adoption: A state of the art review, Expert Systems with Applications, V 203
Costello, E., Doody, J., McGinty, L., & Smyth, B. (2006). iCARE: Intelligent customer assistance for recommending eyewear. Paper presented at the International Conference on Intelligent User Interfaces, Proceedings IUI, 2006, 282–284.
Liu J, Chao Liu, Renjie Lv, Kang Zhou, Yan Zhang (2023) Is ChatGPT a Good Recommender? A Preliminary Study, https://arxiv.org/abs/2304.10149
Zhang Y, Ding H, Zeren Shui, Yifei Ma, James Zou, Anoop Deoras, and Hao Wang. (2021) Language Models as Recommender Systems: Evaluations and Limitations. In I (Still) Can’t Believe It’s Not Better! NeurIPS 2021 Workshop. https://openreview.net/forum?id=hFx3f Y7- m9b

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Recommendation via Entering a Human Conversation

Status:

Version 1

Abstract

Figures

1 Introduction

2 Slot-filling Conversational Recommendation Systems

3 Computing Recommendation for a Dialogue

4 Assuring the recommendation is persuasive and properly argued for

5 Continuing conversation with CRS Agent

6 CRS and ChatGPT

7 System Architecture

8 Evaluation

9 Related Work and Discussion

9.1 Related Systems

9.2 Related approaches and components of recommendation system

10 Conclusions

References

Additional Declarations

Status:

Version 1