ResPer: Computationally Modelling Resisting Strategies in Persuasive Conversations

Modelling persuasion strategies as predictors of task outcome has several real-world applications and has received considerable attention from the computational linguistics community. However, previous research has failed to account for the resisting strategies employed by an individual to foil such persuasion attempts. Grounded in prior literature in cognitive and social psychology, we propose a generalised framework for identifying resisting strategies in persuasive conversations. We instantiate our framework on two distinct datasets comprising persuasion and negotiation conversations. We also leverage a hierarchical sequence-labelling neural architecture to infer the aforementioned resisting strategies automatically. Our experiments reveal the asymmetry of power roles in non-collaborative goal-directed conversations and the benefits accrued from incorporating resisting strategies on the final conversation outcome. We also investigate the role of different resisting strategies on the conversation outcome and glean insights that corroborate with past findings. We also make the code and the dataset of this work publicly available at https://github.com/americast/resper.


Introduction
Persuasion is pervasive in everyday human interactions. People are often exposed to scenarios that challenge their existing beliefs and opinions, such as medical advice, election campaigns, and advertisements (Knobloch-Westerwick and Meng, 2009;Bartels, 2006;Speck and Elliott, 1997). Of late, huge strides have been taken by the Computational Linguistics community to advance research in persuasion. Some seminal works include identifying persuasive strategies in text  and conversations (Wang et al., 2019), investigating the * denotes equal contribution interplay of language and prior beliefs on successful persuasion attempts (Durmus and Cardie, 2018;Longpre et al., 2019), and generating persuasive dialogues (Munigala et al., 2018).
However, a relatively unexplored domain by the community is the investigation of resisting strategies employed to foil persuasion attempts. As succinctly observed by Miller (1965): "In our daily lives we are struck not by the ease of producing attitude change but by the rarity of it." Several works in cognitive and social psychology (Fransen et al., 2015a;Zuwerink Jacks and Cameron, 2003) have put forward different resisting strategies and the motivations for the same. However, so far, there has not been any attempt to operationalise these strategies from a computational standpoint. We attempt to bridge this gap in our work.
We propose a generalised framework, grounded in cognitive psychology literature, for automatically identifying resisting strategies in persuasion oriented discussions. We instantiate our framework on two publicly available datasets comprising persuasion and negotiation conversations to create an annotated corpus of resisting strategies.
Furthermore, we design a hierarchical sequence modelling framework, that leverages the conversational context to identify resisting strategies automatically. Our model significantly outperforms several neural baselines, achieving a competitive macro-F1 score of 0.56 and 0.66 on the persuasion and negotiation dataset, respectively.
We refer to our model as RESPER, which is not only an acronym for Resisting Persuasion, but also a play on the word ESPer: a person with extrasensory abilities. The name is apt since we observe that incorporating such resisting strategies could provide additional insight on the outcome of the conversation. In fact, our experiments reveal that the resisting strategies are better predictors of conversation success for the persuasion dataset than the strategies employed by the persuader. We also observe that the buyer's strategies are more influential in negotiating the final price. Our findings highlight the asymmetric nature of power roles arising in non-collaborative dialogue scenarios and form motivation for this work.

Related Works
The use of persuasion strategies to change a person's view or achieve a desired outcome finds several real-world applications, such as in election campaigns (Knobloch-Westerwick and Meng, 2009;Bartels, 2006), advertisements (Speck andElliott, 1997), and mediation (Cooley, 1993). Consequently, several seminal NLP research have focused on operationalising and automatically identifying persuasion strategies (Wang et al., 2019), propaganda techniques (Da San Martino et al., 2019), and negotiation tactics , as well as the impact of such strategies on the outcome of a task He et al., 2018;Joshi et al., 2021). However, there is still a dearth of research from a computational linguistic perspective investigating resisting strategies to foil persuasion.
Resisting strategies have been widely discussed in literature from various aspects such as marketing (Heath et al., 2017), cognitive psychology (Zuwerink Jacks and Cameron, 2003), and political communication (Fransen et al., 2015b) . Some notable works include the identification and motivation of commonly-used resisting strategies (Fransen et al., 2015a;Zuwerink Jacks and Cameron, 2003), the use of psychological metrics to predict resistance (San José, 2019;Ahluwalia, 2000), and the design of a framework to measure the impact of resistance (Tormala, 2008). However, these works have mostly relied on qualitative methods, unlike ours, which adopts a data-driven approach. We propose a generalised framework to characterise resisting strategies and employ state-of-the-art neural models to infer them automatically. Thus our work can be considered complementary to past research.
The closest semblance to our work in NLP literature ties in with argumentation, be it essays (Carlile et al., 2018), debates (Cano-Basave and He, 2016), or discussions on social media platforms (Al-Khatib et al., 2018;Zeng et al., 2020). Such works have revolved mostly on analysing argumentative strategies and their effect on others.
Recently, Al Khatib et al. (2020) demonstrated that incorporating the personality traits of the resis-tor was influential in determining their resistance to persuasion. Such an observation acknowledges the power vested in an individual to resist change to their existing beliefs. Our work exhibits significant departure from this because we explicitly characterise the resisting strategies employed by the user. Moreover, our work focuses on the general domain of non-collaborative task-oriented dialogues, where several non-factual resisting strategies are observed, making it distinctly different from argumentation (Galitsky et al., 2018). We assert that focusing on both parties is imperative to get a complete picture of persuasive conversations.

Framework
In this section, we describe the datasets, the resisting strategies employed, and the annotation framework to instantiate the strategies.

Dataset Employed
We choose persuasion-oriented conversations, rather than essays or advertisements , since we can observe how the participants respond to the persuasion attempts in real-time. To that end, we leverage two publicly available corpora on persuasion (Wang et al., 2019) and negotiation . We refer to these datasets as "Persuasion4Good" or P4G and "Craigslist Bargain" or CB hereafter.
P4G comprises conversational exchanges between two anonymous Amazon Mechanical Turk workers with designated roles of the persuader, ER and persuadee, EE. ER had to convince EE to donate a part of their task earnings to the charity Save the Children. We investigate the resisting strategies employed only by EE in response to the donation efforts. We emphasise that the conversational exchanges are not scripted, and the task is set up so that a part of EE's earnings is deducted if they agree to donate. Since there is a monetary loss at stake for EE, we expect them to resist.
CB consists of simulated conversations between a buyer (BU) and a seller (SE) over an online exchange platform. Both are given their respective target prices and employ resisting strategies to negotiate the offer.
We choose these datasets since they involve noncollaborative goal-oriented dialogues. As a result, we can definitively assess the impact of different resisting strategies on the goal.

Resisting Strategy
Persuasion (P4G) Negotiation (CB) Source Derogation Attacks/doubts the organisation's credibility. Attacks the other party or questions the item. My money probably won't go to the right place Was it new denim, or were they someone's funky old worn out jeans?

Counter Argument
Argues that the responsibility of donation is not on them or refutes a previous statement.
Provides a non-personal argument/factual response to refute a previous claim or to justify a new claim. There are other people who are richer It may be old, but it runs great. Has lower mileage and a clean title.

Personal Choice
Attempts to saves face by asserting their personal preference such as their choice of charity and their choice of donation.
Provides a personal reason for disagreeing with the current situation or chooses to agree with the situation provided some specific condition is met. I prefer to volunteer my time I will take it for $300 if you throw in that printer too. Information Inquiry Ask for factual information about the organisation for clarification or as an attempt to stall.
Requests for clarification or asks additional information about the item or situation. What percentage of the money goes to the children?
Can you still fit it in your pocket with the case on? Self Pity Provides a self-centred reason for not being able/willing to donate at the moment.
Provides a reason (meant to elicit sympathy) for disagreeing with the current terms.

Hesitance
Attempts to stall the conversation by either stating they would donate later or is currently unsure about donating.
Stalls for time and is hesitant to commit; specifically, they seek to further the conversation and provide a chance for the other party to make a better offer. Yes, I might have to wait until my check arrives. Ok, would you be willing to take $50 for it? Self-assertion Explicitly refuses to donate without even providing a factual/personal reason Asserts a new claim or refutes a previous claim with an air of finality/ confidence.

Not today
That is way too little.

Framework Description
In this subsection, we briefly describe the resisting strategies commonly referenced in social and cognitive psychology literature. This enables us to design a unified framework for the two datasets, built upon common underlying semantic themes. Fransen et al. (2015a) identified 4 major clusters of resisting strategies, namely contesting (Wright, 1975;Zuwerink Jacks and Cameron, 2003;Abelson and Miller, 1967), empowerment (Zuwerink Jacks and Cameron, 2003;Sherman and Gorkin, 1980), biased processing (Ahluwalia, 2000), and avoid-ance (Speck and Elliott, 1997). Each individual category can be subdivided into finer categories showcased in italics henceforth.
Contesting refers to attacking either the source of the message (Source Derogation) or its content (Counter Argumentation). A milder form of contesting involves seeking clarification or information termed Information Inquiry. Prior work has shown a positive association between working knowledge and one's ability to resist persuasion (Wood and Kallgren, 1988;Luttrell and Sawicki, 2020). Therefore, Information Inquiry can be interpreted as a form of resistance where the resistor seeks to satisfy their doubts because they are sceptical of the persuader's intents or messages. This is prominent in certain conversations in P4G where a sceptical EE questions the charity's legitimacy.
Biased processing mitigates external persuasion by selectively processing information that conforms with one's opinion or beliefs (Fransen et al., 2015a). For simplicity, we subsume strategies that denote personal preference, namely Attitude Bolstering and Biased Processing, into a unified category Personal Choice. We refrain from incorporating Self Assertion into the Personal Choice category since it deals with bolstering one's confidence and not one's opinions or attitudes. The subtle difference is highlighted in Table 1.
Avoidance strategies distance the resistor from persuasion, either physically or mechanically, or refuse to engage in topics that induce cognitive dissonance (Fransen et al., 2015a). However, in the context of task-oriented conversations, wherein participants are expected to further a goal, avoidance often manifests as Hesitance to commit to the current situation.
We identify seven major resisting strategies across the datasets, namely Source Derogation, Counter Argumentation, Information Inquiry, Personal Choice, Self Pity, Hesitance, and Self Assertion. Since the datasets comprise two-party conversations between strangers, Social Validation, which requires garnering the support of others, was absent. We now describe how these resisting strategies were instantiated in the following section.

Instantiating the Resistance Framework
We emphasise that although the description and meaning of a strategy remain the same across the two datasets, their semantic interpretation depends on the context. For example, scepticism towards the charity in P4G and criticism of the product in CB are instances of Source Derogation. This is because ER represents the charity, whereas the seller is being accused of selling an inferior product. Likewise, we instantiate the predicates for the remaining six resisting strategies for the two datasets, with examples in Table 1.
We label the utterances of persuadee (EE) in P4G and the buyers (BU) and sellers (SE) in CB with at least one of the seven corresponding resisting strategies, or 'Not-A-Strategy' if none applies. The 'Not-A-Strategy' label includes greetings, offtask discussions, agreement, compliments, or other tokens of approval. We acknowledge that an utterance can have more than one resisting strategy embedded in it. For example, the utterance "The price is slightly high for used couches, would you come down to 240 if I also picked them up?", is an instance of both Personal Choice and Counter Argumentation.
We also note that Information-Inquiry is not a resisting strategy for CB since asking additional information/clarification is an expected behaviour before finalising a deal. We keep the label nevertheless to show comparison with P4G. We present the flowchart detailing the annotation framework in Figure 3 of Appendix.

Annotation Procedure and Validation
We describe the annotation procedure for both the CB and P4G dataset here and its subsequent validation. For CB, three authors independently annotated five random conversations adhering to the flowchart. If the conversations chosen were simple or had few labels, a new set of 5 conversations were taken up. This constitutes one round. After each round, the Fleiss Kappa score was computed, and the authors discussed to resolve the disagreements and revise the flowchart. Then began the next round on a new set of 5 random conversations. For CB, 5 rounds of revision were carried out over 24 conversations, until a high Fleiss kappa (0.790) (Fleiss, 1971) was obtained. Finally, the three authors independently went ahead and annotated approximately 250 distinct conversations, yielding a corpus of 800 CB conversations. Our annotation procedure requires a rigorous reliable refinement phase but a comparatively faster annotation phase by dividing the annotation between the authors. Thus the conversations annotated by each author were mutually exclusive. Similarly, for P4G dataset, four authors annotated 3 conversations per round, since a conversation in P4G was comparatively longer. 4 rounds of revision across 12 conversations was done to achieve the final kappa-score of 0.787. The four authors then went ahead and divided the task of annotating the 500 conversations amongst themselves. We show an annotated conversation snippet for the two datasets in Table 3.

Dataset Statistics
The P4G and CB datasets comprise 530 and 800 labelled conversations, respectively, spanning an average of 37 and 12 utterances per conversation.  The datasets cover two distinct persuasion scenarios and also illustrate the rights and obligations shown by the participants. For example, in P4G, EE comes into the interaction blind and is unaware of the donation attempt. We encounter several conversations where EE is willing to donate since it resonates with their beliefs, and no resisting strategies are observed. However, for CB, the participants received prior instructions to negotiate a deal, and hence resisting strategies were more prominent. We present the frequency distribution of the seven strategies in Table 4. We observe that the distributions of strategies are skewed for both the datasets and is more pronounced for P4G, where 'Not-A-Strategy' accounts for the lion's share. We also see that the buyer exhibits more resisting strategies than the seller highlighting the asymmetric role of the two participants.
Nevertheless, we reiterate that the resisting strategies we propose are applicable for both the domains. In the next section, we propose the framework to infer such strategies automatically.

Methodology
In this section, we describe the methodology adopted for inferring the resisting strategies in persuasion dialogues and how they can be leveraged to determine the dialogue's outcome.

Resisting Strategy prediction
We model the task of identifying resisting strategies as a sequence labelling task. We assign each utterance in the dialogues with a label representing either one of the seven resisting strategies or Not-A-Strategy.
Since the resisting strategies, by definition, occur in response to the persuasion attempts, our model architecture needs to be cognizant of the conversational history. To that end, we adopt a hierarchical neural network architecture, similar to Jiao et al. (2019), to infer the corresponding resisting strategy. The architecture leverages the previous conversational context in addition to the current contextualised utterance embedding. Our choice is motivated by the recent successes of hierarchical sequence labelling frameworks in achieving state-of-the-art performance on several dialogueoriented tasks. Some myriad examples include emotion recognition (Majumder et al., 2019;Jiao et al., 2019), dialogue act classification Raheja and Tetreault, 2019), face act prediction (Dutt et al., 2020), open domain chit-chat (Zhang et al., 2018;Kumar et al., 2020) and the like. We hereby adopt this as the foundation architecture for our work and refer to our instantiation of the architecture as RESPER. Architecture of RESPER: An utterance u j We acknowledge that an utterance can have multiple labels. However, such utterances comprise only 1.2% and 3.85% of the P4G and the CB datasets, respectively. In such cases, the label is randomly selected.  Figure 1: A diagram illustrating how RESPER works. The encoder shown on the left takes the BERT representations of a token as input and passes it through a BiGRU layer followed by Self Attention. The outputs from BERT, BiGRU and self-attention are then concatenated to form the output. Max-pooling over this output yields the corresponding utterance embedding. This utterance representation is passed through a uni-directional GRU followed by Masked-Self-Attention and fusion to yield the contextualised utterance embedding.
is composed of tokens [w 0 , w 1 , ..., w K ] represented by their corresponding embeddings [e(w 0 ), e(w 1 ), ..., e(w K )]. In RESPER, we obtain these using a pre-trained BERT model (Devlin et al., 2019). We pass these contextualised word representations through a bidirectional GRU to obtain the forward − → h k and backward ← − h k hidden states of each word, before passing them into a Self-Attention layer. This gives us the corresponding attention outputs, − − → ah k and ← − − ah k as described below.
Finally, we concatenate the contextualised word embedding with the GRU hidden states and Attention outputs in the fusion layer to obtain the final representation of the word e c (w k ). We represent the bias as b w . Here, We perform max-pooling over the fused word embeddings to obtain the j th utterance embedding, e(u j ).
e(u j ) = max(e c (w 1 ), e c (w 2 ), ...e c (w K )) We use a unidirectional GRU and Masked Self-Attention to encode conversational context, to ensure that the prediction for the j th utterance is not influenced by future utterances. Similarly, we calculate the contextualized representation of an utterance e c (u j ) using the conversation context. We pass e(u j ) through a uni-directional GRU that yields the forward hidden state − → H j . Masked Self-Attention over the previous hidden states, yields −−→ AH j . We fuse e(u j ), − → H j and −−→ AH j before passing it through a linear layer with tanh activation to obtain e c (u j ).
We project the final contextualised utterance embedding e c (u j ) onto the state space of resisting strategies. We apply softmax to obtain a probability distribution over the strategies, with Negative Log-Likelihood (NLL) as the loss function to obtain the strategy loss.

Conversation Outcome prediction
We further investigate the impact of resisting strategies on the outcome of the conversation. We represent a strategy as a fixed dimensional embedding initialised at random. We subsequently encode a sequence of strategies by passing them through a unidirectional GRU to obtain a final representation for the sequence. We project the representation onto a binary vector which encodes for the conversation outcome. We apply softmax with NLL across all the conversations to obtain the outcome prediction loss.

Experiments
In this section, we describe the baselines and evaluation metrics. We present the experimental details of our model in Table 5.

Baselines
Resisting strategy prediction: We experiment with standard neural baselines for text classification, which have also been used in classifying persuasion strategies, namely CNN (Kim, 2014;Wang et al., 2019) and BiGRU . To ensure a fair comparison, we introduce pre-trained BERT-embeddings (Devlin et al., 2019) as input to the baselines, henceforth denoted as BERT-CNN and BERT-BiGRU. Furthermore, to inspect the impact of conversational history, we remove the conversational GRU from RESPER such that the utterance embedding e(u j ) is directly used for prediction. We refer to this architecture as BERT-BiGRUsf, since it employs self-attention(s) and fusion (f) on top of BERT-BiGRU. Finally, we experiment with the best performing HiGRU-sf model of Jiao et al. (2019) as another baseline. Conversation success prediction: The notion of conversation success depends on the choice of dataset. For P4G, we consider the resisting strategies to be successful if the persuadee (EE) refused to donate to charity. For CB, we adopt the same notion of success as , namely when the seller (SE) can sell at a price greater than the median sale-to-list ratio r. r = sale price − buyer target price listed price − buyer target price (1) To observe the effect of conversation success, we experiment with strategies of both the parties involved. For P4G, we encode separately (i) the persuasion strategies of ER as identified by Wang et al. (2019), (ii) the resisting strategies employed by EE and (iii) both the persuasion and resisting strategies. Likewise, for CB, we encode the resisting strategies of only (i) the buyer (BU) (ii) the seller (SE) (iii) both. These experiments would enable us to investigate which party has a greater influence on conversation success.  Table 5: Here we describe the search-space of all the hyper-parameters used in our experiments and describe the search space we used to find the hyper-parameters. d h1 , d h2 represents the hidden dimensions of the Utterance GRU and the Conversation GRU.

Evaluation metrics
We adopt the same evaluation procedure for both the resisting strategy and the conversation outcome prediction task across the datasets. In either case, we perform five-fold cross-validation due to paucity of annotated data. We report performance in terms of the weighted and macro F1-scores across the five folds. Our choice of the metric is motivated by the high label imbalance, as observed in Table 4.

Results
In this section, we answer the following : Q1. How well does RESPER identify resisting strategies for Persuasion and Negotiation?
Q2. Are resisting strategies good predictors of conversation success? What insights can one glean from the results?

Predicting resisting strategies
We present the results for the automated identification of resisting strategies in Table 6. We observe that all the models achieve a comparatively lower performance on P4G, mainly due to the higher proportion of 'Not-a-Strategy' labels for the latter. We gauge the benefits of incorporating conversational context by the significant improvement of Macro F1 score by 0.036 and 0.011 for P4G and CB respectively. In fact, RESPER outperforms all the proposed baselines significantly. Error Analysis: We present the confusion matrix for predicting resisting strategies using RESPER on the Persuasion (P4G) and Negotiation (CB) Weighted F1 Scores are calculated by taking the average of the F1 scores for each label weighted by the number of true instances for each label.
We estimate the statistical significance using the paired bootstrapped test of Berg-Kirkpatrick et al. (2012), due to the small number of data (Dror et al., 2018).   datasets in Figures 2(a) and 2(b) respectively. We observe that most classification errors occur when a resisting strategy is incorrectly inferred as 'Not-A-Strategy'. The effect is more prevalent for P4G since 'Not-A-Strategy' comprises 80% of all annotated labels. Other notable instances of misclassification for P4G occurs when Self Assertion is predicted as Self Pity since both strategies refer to one's self. These strategies occur so infrequently (see Table 4) that the models lack sufficient information to distinguish between the two categories. Likewise, for the CB corpus, Hesitance utterances which constitute a price request, are often posed as questions. This causes the model to predict the strategy as Information Inquiry instead. Self Assertion is often incorrectly marked as Source Derogation possibly because it often takes a firm stance, and is likely to disparage the other party in the process, thereby confusing the model.

Conversation Outcome Prediction
We observe how the sequence of strategies adopted by the two participants have a disproportionate impact on the final conversation outcome in Table 7. It is interesting to note that the resisting strategies for the persuadee have a greater effect on the conversation outcome (macro-F1 score of 0.62) than the persuasion strategies themselves (macro-F1 score of 0.59). Moreover, incorporating both the persuasion and resisting strategies boosts the prediction performance even further to 0.65.
We also observe an asymmetry in the roles of the buyer (BU) and the seller (SE) for the CB dataset. We observe that BU's strategies are significantly more effective in deciding the conversation outcome, probably because buyers demonstrate a higher number of resisting strategies. These experiments highlight the importance of incorporating resisting strategies to gain a complete picture.

Comparative Analysis of Strategies
Emboldened by the success of resisting strategies to infer the conversational outcome, we probe deeper to investigate the impact of individual strategies. We apply logistic regression with the frequency of strategies, of either participant, as the features while the outcome variable denotes conversation success. We observe the coefficients of the strategies to infer their correlation with conversation success and their corresponding p-values to determine whether the correlation was indeed statistically significant. Our procedure follows previous work in identifying influential persuasion strategies Wang et al., 2019). We present the results of this analysis in Table 8.  For P4G, all the resisting strategies for persuasion apart from Counter-Argumentation are positively correlated with a refusal to donate. The highest impact stems from Self Assertion. Previous research (Fransen et al., 2015a;Zuwerink Jacks and Cameron, 2003) has noticed that Self Assertion is prominent amongst individuals with high self-esteem. Such individuals are confident about their beliefs and less likely to conform. Similarly, a high positive coefficient for Information Inquiry can be attributed as follows. EE inquires information about the charity not only as a means to verify their legitimacy, but also to gain the knowledge they can exploit to their advantage. An innocuous question like 'Where will my money go?' would enable EE to assert that they are keener to help children in their own country instead, thereby resisting the donation attempt and saving face.
The CB scenario setup ensures that the coefficients of the strategies set for BU and SE would be anti-correlated, which holds for the Table 8. Like P4G, a high negative coefficient of Self Assertion signifies that SE's price is disagreeable to BU -they would instead not buy. Moreover, the high coefficient of Counter Argumentation justifies that it is an effective tactic for both parties.

Conclusion
We present a generalised computational framework grounded in cognitive psychology to operationalise resisting strategies employed to counter persuasion. We identify seven distinct resisting strategies that we instantiate on two publicly available corpora comprising persuasion and negotiation conversations. We adopt a hierarchical sequence labelling architecture to infer the resisting strategies automatically and observe that our model achieves competitive performance for both datasets. Furthermore, we examine the interplay of resisting strategies in determining the final conversation outcome, which corroborates with previous findings. In the future, we would like to explore better models to encode the strategy information and apply our framework to improve personalised persuasion and negotiation dialogue systems. We would also like to study the influence of other confounding factors such as power dynamics on the outcomes of conversations featuring resisting strategies.