--- title: "PARTNER: A Persuasive Mental Health and Legal Counselling Dialogue System for Women and Children Crime Victims" authors: - "Priyanshu Priya" - "Kshitij Mishra" - "Palak Totala" - "Asif Ekbal" doi: "10.24963/ijcai.2023/686" url: "https://ijcai.org/proceedings/2023/686" pdf: "https://ijcai.org/proceedings/2023/0686.pdf" venue: "Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence" year: 2023 pages: "6183-6191" track: "AI for Good" keywords: - "AI for Good: Humans and AI" - "AI for Good: Natural Language Processing" --- # PARTNER: A Persuasive Mental Health and Legal Counselling Dialogue System for Women and Children Crime Victims Priyanshu Priya, Kshitij Mishra, Palak Totala, Asif Ekbal *Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence*, pages 6183-6191. DOI: [10.24963/ijcai.2023/686](https://doi.org/10.24963/ijcai.2023/686) ## Abstract The World Health Organization has underlined the significance of expediting the preventive measures for crime against women and children to attain the United Nations Sustainable Development Goals 2030 (promoting well-being, gender equality, and equal access to justice). The crime victims typically need mental health and legal counselling support for their ultimate well-being and sometimes they need to be persuaded to seek desired support. Further, counselling interactions should adopt correct politeness and empathy strategies so that a warm, amicable, and respectful environment can be built to better understand the victims’ situations. To this end, we propose PARTNER, a Politeness and empAthy strategies-adaptive peRsuasive dialogue sysTem for meNtal health and LEgal counselling of cRime victims. For this, first, we create a novel mental HEalth and legAl counseLling conversational dataset HEAL, annotated with three distinct aspects, viz. counselling act, politeness strategy, and empathy strategy. Then, by formulating a novel reward function, we train a counselling dialogue system in a reinforcement learning setting to ensure correct counselling act, politeness strategy, and empathy strategy in the generated responses. Extensive empirical analysis and experimental results show that the proposed reward function ensures persuasive counselling responses with correct polite and empathetic tone in the generated responses. Further, PARTNER proves its efficacy to engage the victim by generating diverse and natural responses. ## Full text ## 1 Introduction According to the World Health Organization (WHO), the increase in the number of crimes against women 1 and children 2 1 https://www.who.int/news-room/fact-sheets/detail/violenceagainst-women 2 https://www.who.int/news-room/fact-sheets/detail/violenceagainst-children (CAW&C) has become a major public concern with repercussions on various societal levels. WHO highlights the social and economic costs of violence and its negative impact on women's and children's health and well-being. Thus, there is a growing need to combat crime against women and children for accomplishing the United Nations Sustainable Development Goals (SDGs) 2030. The prevention/reduction of CAW&C is central to ensuring healthy lives and promoting well-being, achieving gender equality and women empowerment, fostering child development, ensuring equal access to justice and promoting peaceful and inclusive societies for all. Women and children who have been victimized or are at risk of any violence need to have access to quality services, including medical care or mental health support and legal assistance. However, there is a global scarcity of mental health professionals 3 .Also, the victims often fear stigmatization and societal condemnation and thus, hesitate to seek psychological support and/or are reluctant to report crimes. The issue is compounded by the fact that access to legal aid is constrained 4 . An artificial intelligence (AI)-assisted mental health and legal counselling dialogue system that can effectively interact with the victims and understand them could bridge the gap by serving as the first point of contact for them. Such an automated system can facilitate counselling sessions or assist the counsellors, thereby enhancing the accessibility to required counselling services for the victims. It is imperative for the counselling bot to understand the victims' intended requirements and perform the action (counselling act) accordingly in order to provide adequate and pertinent assistance during counselling. One of the important actions required during counselling is persuasion. A persuasive attitude is necessary to facilitate therapeutic change [Abroms, 1968]. Persuasion is shown to be beneficial in improving the users' health-related self-management competencies [Orji, 2014]. Through persuasion, a counselling conversational bot can convince the users to follow specific advice, change their attitudes and beliefs in their own best interest and assist them in developing a sense of readiness to seek expert assistance. Besides, the counselling bot should behave politely and empathetically with the victims to elicit positive outcomes in supportive conversations [Lucas et al., 2014; Norcross, 2002; Robert et al., 2011]. While counselling the victims, the bot must assess their mental and emotional state and tailor the conversation accordingly. Further, the bot needs to establish trust and rapport with the victims to provide effective counselling support. The use of polite and empathetic language displays a credible and cordial impression of the bot. Therefore, the bot should employ different politeness and empathy strategies as per conversation context to ensure a warm, amicable and respectful atmosphere for the victims and to build an emotional connection with them. The example given in Table 1 depicts how the incorporation of politeness and empathy strategies together with persuasive behaviour eventually contributes to a better personalized experience and motivates user engagement. 3 https://www.who.int/publications/i/item/9789240036703 4 https://www.unodc.org/documents/justice-and-prisonreform/LegalAid/Global-Study-on-Legal-Aid Report01.pdf Table 1: Example showcasing the use of politeness (negative politeness) and empathy (emotional support) strategies with the appropriate counselling act (persuasion) in counselling. | Speaker | Utterance | Response Type | |-----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Victim | I am getting anxiety attacks after that incident. I often panic and feel restless. Help me. | Response Type | | Bot | Don't worry and tell us which incident you are talking about? Don't worry, we will surely help. Could you please tell us which incident you are talking about? We understand that anxiety attacks are scary and painful. Don't worry, we will surely help. Can you please let us know which incident you are talking about? We understand that anxiety attacks are scary and painful. Don't worry, we will surely help. A small step can relieve you from all pains, trust us and please let us know which incident you are talking about? | Generic Response Politeness strategy- adaptive Response Politeness and Empathy- strategies-adaptive Response Politeness and Empathy strategies-adaptive Persuasive Response | Considering the importance of mental health and legal support for the victims and the relevance of politeness, empathy and persuasion in counselling conversations, we propose PARTNER, a Politeness and empAthetic strategies-adaptive peRsuasive dialogue sysTem for meNtal health and lEgal counselling of cRime victims. However, developing such a dialogue system is challenging due to the subjective nature of politeness, sophisticated and nuanced ways of expressing empathy, and the intricate techniques of counselling. Further, scarcity of a counselling dialogue data poses a restriction in the development of such systems that can generalize to diverse users in different scenarios. Therefore, first, we create HEAL, a novel mental HEalth and legAl counseLling dataset consisting of counselling conversations between the bot and a crime victim. Second, inspired by the recent success of reinforcement learning (RL) [Casanueva et al., 2018; Mesgar et al., 2020; Mishra et al., 2022a] in fine-tuning a Cross-Entropy loss based model, when trying to ensure some defined aspects in generation, we develop our proposed dialogue system in an RL framework. Our proposed system PARTNER learns from user interactions and improves depending on the users' feedback in the form of rewards. Specifically, a novel reward function is designed consisting of five rewards to ensure counselling act consistency, correct use of different politeness and empathy strategies, fluency and diversity in generated responses. The key contributions of our current work can be summarized as follows: (i) Created a novel large-scale dataset, HEAL annotated with counselling act, politeness, and empathy strategies for mental health and legal counselling of crime victims; (ii) Designed robust transformer-based counselling act, politeness strategy, and empathy strategy classifiers; (iii) Designed a novel reward function to assess generated responses in terms of naturalness, user retainment, and correctness of counselling act, politeness and empathy strategies; (iv) Proposed PARTNER, RL-based politeness and empathy strategies-adaptive persuasive dialogue system for mental health and legal counselling of crime victims; (v) Performed extensive automatic and human evaluation to demonstrate the effectiveness of our proposed system 5 . Societal Impact. Psychological and/or legal counselling are the key elements for strengthening the social, emotional and mental well-being of the victims and for securing their access to justice. These are also fundamental for reducing the inequalities and vulnerabilities that leave people behind and undermine the potential of individuals and of humanity as a whole. Our research which aims at the dialogue generation module in mental health and legal conversational system can be viewed as a step towards achieving the overarching objective of the SDGs 2030 agenda to "leave no one behind". We believe that our current work would be beneficial to the victims as well as the counsellors, law enforcement agencies, NGOs, and other support groups by ensuring 24x7 assistance for the victims. ## 2 RelatedWork The crime victims often suffer from short-term and/or longterm physical and/or mental issues, hence they require both mental health as well as legal assistance for their overall physical, mental, and emotional recovery after victimization [Roberts, 2002; Williamson et al., 2010]. Though there have been attempts to build chatbots for mental health support [Fitzpatrick et al., 2017; Fulmer et al., 2018] or legal assistance [John et al., 2017], research on dialogue systems in the mental health and legal domain for victims is still in the nascent stage. A few rule-based chatbots have been reported in the literature that aims to help the victims of a particular type of crime like You are not alone (domestic violence), Hello Cass (sexual harassment), Law, etc. Due to dependency on rules, these systems are neither scalable nor generalizable. Further, some of these systems emphasize solely on solving mental health issues while others are designated to help victims in legal cases. 5 Codes and data can be accessed at PARTNER or https://www.iitp.ac.in/∼ai-nlp-ml/resources.html#PARTNER Studies have reported that dialogue systems to support victims should be 'high performing' (effective with information), 'smart' (accurate and correct) and 'personable' (personalizing information) 6 . Thus, depending upon the victim's need, the bot should carry out relevant actions, like persuading the victims, offering counselling support/legal aid, or performing casual actions like information seeking/delivery. Persuasion has been regarded as a crucial aspect in the counselling process [Strong, 1968; Heppner and Claiborn, 1989]; counsellors adhering to persuasive techniques are perceived to play an active role during counselling [Dorn et al., 1986]. Recently, there has been an increased interest in making the agent persuasive that can bring positive change and/or improve knowledge, awareness, or understanding to help people achieve better health [Fogg, 2002; Althoff et al., 2016; Demasi et al., 2019; Liang et al., 2021]. Persuasive dialogue systems have been applied in the legal arena as well [Gordon, 1993]. While a few prior works attempted to build counselling dialogue systems for providing personalized support (asking personalized questions and delivering personalized information) to the victims [Kim et al., 2022], incorporating more humanly quotient like politeness and empathy into these systems remains unexplored. Studies have proven that politeness and empathy are crucial for providing effective counselling support [Lucas et al., 2014; Norcross, 2002; Robert et al., 2011]. Politeness contributes to a sense of compassion and aids the disclosure of sensitive information [Lucas et al., 2014; Kim et al., 2018]. Empathetic interactions are crucial in establishing a therapeutic bond and rapport [Norcross, 2002; Robert et al., 2011] and there is substantial evidence that demonstrates that empathy is helpful in reducing symptom severity in mental health support seekers [Elliott et al., 2018]. The authors in [Saha et al., 2022a; Saha et al., 2022b] explored the ways to induce empathy in the mental health dialogue system to reflect an emotional attachment with the users. The existing research on mental health or legal intervention systems for victims primarily focused on delivering generic support, which is perhaps more akin to a self-help book and thus, often fails to cater to the urgent needs of the victims. In comparison, our proposed system enquires about the incident, assesses the immediate psychological needs, and suggests local or national-level resource channels or coping methods for certain forms of victimization. In addition, instilling socially desirable behaviour like politeness, empathy and/or persuasiveness was totally unexplored in earlier works. Our work further differentiates in the sense that we focus on a new task of developing an end-to-end counselling dialogue system for crime victims where five novel rewards force a dialogue agent to adopt different counselling acts (including but not limited to persuasion), politeness and empathy strategies based on the conversational context. The study close to our work is [Mishra et al., 2022b], which attempts to incorporate politeness strategy and emotion in the agent's responses to enhance its persuasiveness for social good. Following [Mishra et al., 2022b], we train our system with standard Proximal Policy Optimization (PPO) [Schulman et al., 2017] designing five different novel rewards pertaining to our task. To the best of our knowledge, this is the first attempt to build such a counselling dialogue system for crime victims. 6 https://www.filesforprogress.org/memos/sexual-assaultvictims-want-services-tailored-to-needs.pdf ## 3 Dataset Our proposed dataset HEAL comprises conversations between a victim of crime and a conversational bot. The purpose of these interactions is to inspire and uplift the victims, ultimately assisting them in restoring their physical and emotional well-being by offering relevant mental health and/or legal support to them. ## 3.1 Data Preparation The primary focus of HEAL is to address the mental health and legal needs of women and children who have experienced various crimes such as domestic violence, acid attacks, cyberstalking, online harassment, impersonation, etc. In order to offer genuine support and counselling services to the victims, various sources such as National Cybercrime Reporting Portal, National Commission for Women, Ministry of Women and Child Development, Criminal Law Amendment Act 2013 and Information Technology (Amendment) Act 2008 are referred during data preparation. Additionally, real-life stories of crimes against women and children are gathered from multiple websites. These stories, along with the information gathered from the various sources, are then given to four annotators who have post-graduate qualifications in English linguistics and substantial experience in related tasks. Based on this information, the annotators are instructed to create dialogues between a counselling bot and a victim using a well-known Wizard-of-Oz approach [Kelley, 1984], where one annotator assumes the role of the counselling bot and the other acts as the victim. To eliminate any correlation between the counselling acts or politeness strategies or empathy strategies of the counselling bot and the targeted victim's characteristics, one of the two annotators is randomly assigned the role of either the counselling bot or the victim. In order to prepare the counselling dataset and annotate it with suitable counselling acts, politeness, and empathy strategies, the annotators are provided with specific guidelines to follow. These guidelines are mentioned below: Data Preparation Guidelines. To develop the guidelines, we collaborated with mental health and legal experts from government institutions. The guidelines are outlined as follows: 1. Build relationship: The counselling bot should strive to establish a harmonious relationship with the victims by demonstrating a caring and amicable attitude towards them; 2. Identify the problem: The counselling bot should aim to recognize the difficulties faced by the victim and understand the root causes of these difficulties; 3. Facilitate change: The counselling bot should utilize suitable tactics and interventions to foster a constructive shift in the victim's circumstances and emotional state: (i) Imbibe politeness in the bot's responses to uncover the victim's innate qualities and maintain a seamless flow of conversation during counselling; (ii) Demonstrate care, warmth, and compassion to nurture a supportive environment and cultivate an empathetic bond with the victim.; (iii) Demonstrate comprehension of the victim's experiences and emotional state inferred from their responses, in order to foster a connection between the bot and the victim, and eventually facilitate the disclosure of concealed emotions and experiences; (iv) Maintain a tone of affirmation and positivity in the bot's responses to uplift the victim's self-esteem.; 4. Evaluation and Termination: Additionally, the bot should offer essential mental health and legal information/services (such as contact details of experts, medical care facilities, legal guidance, information about support groups, assistance with filing complaints, etc.) to the victims, when necessary. Subsequently, the bot should provide a set of safety tips to help prevent similar problems in the future. Counselling Act. Counselling conversations demand carefully devising counselling acts capable of catering to the needs of users. Thus, we design a set of eight counselling acts to meet the adequate requirement of the victims and also be readily understandable in order to aid in the development of a counselling conversational system for the victims. These counselling acts are formulated under the supervision of domain experts. They are described below: 1. Counselling support: Provide various support services like medical help, mental health-related aid, NGOs information etc. during counselling based on the victim's need. 2. Legal assistance: Ensures legal assistance to the victims. 3. Persuasion: Assists the victims in developing a readiness to seek professional assistance by compelling them to adhere to specific recommendations, modifying their attitudes and beliefs in their own best interests, and fostering a sense of readiness to do so. 4. Seek information: Request for a few basic information in order to comprehend the problem and provide relevant assistance. 5. Deliver information: Provide information pertaining to the problem being discussed. 6. Re-check assistance: Inquires for further help or clarification about the problem under discussion. 7. Greet: Typically, each conversation begins with a greeting from one speaker and an appropriate response from the other. 8. Closing remark: Marks the end of the conversation. Politeness and Empathy Strategies. In order to establish a personal, friendly, and empathetic connection with the victim during counselling, a counselling bot can employ various politeness and empathy strategies. The use of appropriate politeness strategies and empathy strategies by the counsellor can help minimize threats to the victim's self-esteem and demonstrate emotional and cognitive understanding of the victim's situation, respectively. We consider three wellknown politeness strategies, namely positive politeness, negative politeness, and bald on-record, as outlined in Brown and Levinson's work [Brown et al., 1987]. However, since there is a lack of well-defined empathy strategies in mental health and legal counselling, based on the drafted guidelines, we formulated seven empathy strategies, considering the significance of empathy in counselling [Norcross, 2002]. To ensure consistency and make necessary adjustments, the four aforementioned annotators are paired with each other, resulting in six unique pairs, and they generate a total of 48 conversations. Discrepancies in the strategies are analyzed, and modifications are made under the guidance of domain experts. Once all seven empathy strategies are approved by the domain experts, they are applied to all conversations. The empathy strategies are as follows: 1. Reflective listening: Demonstrates a genuine curiosity to learn and delve deeper into the details shared by the victims, creating a sense that the bot is genuinely interested in listening to them. 2. Confidential comforting: Displays genuine interest and concern for the privacy of the victims, providing assurance that any information shared will be treated with the utmost confidentiality. 3. Evoke motivation: Encourages the victims to embrace a forward-looking perspective and participate in activities that promote feelings of positivity and optimism. 4. Express emotional support: Provides emotional solace or words of encouragement to entirely comprehend the problems faced by the victims and the intensity of their emotions. 5. Offer counselling: Provides essential mental health and legal counselling advice, along with contact information of experts, whom the victim can reach out to for further guidance and assistance. 6. Escalate assurance: Reassures the victims that they are never to blame for any form of assault and firmly conveys the message that they are not alone, emphasizing that they can always seek help and support. 7. No strategy: Assigned to the utterances which do not utilize any empathy strategy. ## 3.2 Data Annotation For our work, the counselling bot's utterances in each dialogue are annotated with one of the eight counselling acts, three politeness strategies, and seven empathy strategies. For annotating the bot's utterances, we instruct the same four annotators to label every utterance of the bot with the given set of counselling acts, politeness strategy and empathy strategy labels. To strengthen the annotations, annotators are clearly briefed about the annotation guidelines along with illustrative examples for each of the counselling acts, politeness and empathy strategies before beginning the annotation procedure. A reliable multi-rater Kappa [McHugh, 2012] agreement ratio of 79.4%, 83.7%, and 80.1% is observed for the counselling act, politeness strategy and empathy strategy, respectively, among all the four annotators. The dataset statistics are shown in Table 2. Table 2: Dataset statistics. | Metrics | Train | Dev | Test | |------------------------------|---------|-------|--------| | No. of Dialogues | 162 | 22 | 32 | | No. of Utterances | 4,133 | 407 | 664 | | Avg. Utterances per Dialogue | 25.51 | 18.50 | 20.75 | ## 4 Proposed System: PARTNER Our proposed system comprises of two components viz. Cross-Entropy Loss based Dialogue Model (CELDM) - to foster natural language interaction between the counselling bot and the crime victim, and Proximal Policy Optimization (PPO) Loss based Dialogue Model (PLDM) - to generate politeness and empathy strategies-adaptive counselling responses. PLDM further can be divided into two subcomponents viz. Reward - uses counselling dialogue act, politeness and empathy strategies classifiers to calculate the respective rewards, and Policy Optimization - updates policy parameters to attain optimal policy. Each preceding component or sub-component is used by the succeeding component or sub-component, i.e. first a dialogue model is trained using cross-entropy loss, then to obtain the final PARTNER, it is fine-tuned with PPO loss. ## 4.1 Cross-Entropy Loss-based Dialogue Model We represent the exchanges between the counselling bot and the crime victim as a multi-turn counselling dialogue, denoted as d = {cb0, cv0, .., cbi, cvi, .., cbT-1, cvT-1}. Here, cbi and cvi represent the ith response from the counselling bot and the crime victim, respectively. Inspired by the approach in [Wu et al., 2021], we employ two GPT2-small [Radford et al., 2019] models, namely ρcb and ρcv, which are trained alternatively to approximate the distribution of utterances from the counselling bot (cb) and the crime victim (cv). Given the preceding context, ρcb and ρcv predict the next token grj to generate a response gr = {gr1, gr2, .., grj , .., grm} comprising of m tokens. We can express the joint probability of both the victim's and the counselling bot's utterances as follows: Finally, the dialogue model ρθ(d) is trained by minimising the cross-entropy loss between the predicted and ground truth probability distribution of utterances. Mathematically, ρθ(d) can be expressed as follows: ## 4.2 PPOLoss based Dialogue Model In order to strengthen appropriate counselling dialogue acts, empathy, and politeness strategies with the generation of diverse yet natural responses, we begin by developing a novel reward function encompassing five distinct rewards tailored to specific aspects. Subsequently, the value of this reward function is utilized in the PPO loss to fine-tune the dialogue model ρθ(d) generating a set of n-candidate responses based on a given context. These set of responses are quality checked in terms of different aspects using five rewards. ## Rewards Our proposed reward function R incorporates two types of rewards. First, we have task-specific rewards, namely R1 - ensuring appropriate counselling dialogue acts, R2 - incorporating correct empathy strategies, and R3 - enforcing politeness strategies in the generated responses. Secondly, we include generic rewards, specifically R4 to measure user retainment and R5 to assess the naturalness. The formulation of the reward function R is as follows: Counselling Dialogue Act Consistency Reward. To ensure consistency in the generated responses, we fine-tuned a RoBERTa-large [Liu et al., 2019] based counselling dialogue act classifier. Each generated candidate is passed through this classifier to predict the counselling dialogue act and then it is compared against the gold response labels. Deviations from the expected act are penalized. where, CCDAL = {0, 1, 2, 3, 4, 5, 6, 7} represents the set of counselling dialogue act labels, Pcda(cbT ), and Pcdai (grT ) gives the predicted counselling dialogue act probabilities for ground truth and generated utterance, respectively. To control the penalization, we introduce the factor γ, where higher values of γ result in more severe penalties 7 . Empathy and Politeness Strategy Rewards. To incorporate empathy and politeness strategy rewards, we utilize dedicated classifiers. For this purpose, we fine-tune a RoBERTalarge model [Liu et al., 2019] using empathy strategy labels CESTL = 0, 1, 2, 3, 4, 5, 6 and politeness strategy labels CPSTL = 0, 1, 2. In order to enforce correct empathy and politeness strategies, at each turn T, we compare the predicted labels for empathy and politeness strategies for the generated response grT with their respective ground truth response labels. Any responses that deviate from the ground truth labels are penalized accordingly. The rewards for empathy strategy (R2) and politeness strategy (R3) can be formulated as follows: where, at a turn T, Pest(cbT ), Ppst(cbT ) corresponds to the ground truth empathy and politeness strategy predicted probabilities, and Pesti (grT ), Ppsti (grT ) gives the predicted probabilities for empathy and politeness strategies respectively. 7 γ is set to be greater than or equal to 1 Retainment Reward. A counselling dialogue should be interactive and engaging, hence, should be able to retain the victim by generating diverse responses. Therefore, retainment reward R4 is designed considering two dialogue quality factors viz. lexical level and sentence level similarities, calculated here as the Jaccard similarity [Jaccard, 1912] and cosine similarity between grT and grT-1 at turns T and T - 1, respectively. Naturalness Reward. To reinforce that a counselling dialogue system generates natural responses of high linguistic quality, we first compute the Negative Log-Likelihood (NLL) of the generated responses. Then, to scale down the values within the range of 0 and 1, we apply the tanh () function 8 to the NLL as formulated below: ## Policy The action selection of a conversational bot can be described by a policy, which is a probability mapping function denoted as Pθ. This function generates a feasible response gr comprising of m tokens, which can be considered as an action, given a particular context representing the current state. Proximal Policy Optimization (PPO). Policy updates are performed at each step using PPO-loss, which effectively ensures low variance compared to the old policy. The optimization of the policy can be formulated in three steps. Initially, gradient ascent is applied to the loss function J(θ) in order to maximize the expected reward: Secondly, in order to restrict large deviations, the log term in the optimization process is replaced with an importance sampling term. Additionally, clipping is employed to prevent catastrophic forgetting, ensuring that the updates do not have excessively large magnitudes: Here, the probability ratio between the new and old policies is calculated as prr(θ) = P new θ /P old θ . ε and ˆ Ay gives the clipping range and estimated advantage (normalized rewards) respectively. Finally, the parameters are updated according to the following equation: 8 The tanh () function maps values between -1 and 1, but in this case, the loss cannot be negative, so the resulting values will always be between 0 and 1. Moreover, sigmoid is not chosen as it would map values between 0.5 and 1. Table 3: Evaluation results of the Empathy and Politeness strategy classifiers. | Classifier | BERT-large - W-ACC | BERT-large - Macro-F1 | RoBERTa-large - W-ACC | RoBERTa-large - Macro-F1 | |----------------------|----------------------|-------------------------|-------------------------|----------------------------| | Counselling strategy | 0.881 | 0.849 | 0.904 | 0.891 | | Empathy strategy | 0.912 | 0.864 | 0.940 | 0.909 | | Politeness strategy | 0.926 | 0.918 | 0.952 | 0.952 | ## 5 Experiments Implementation Details. The pre-trained GPT2-small [Radford et al., 2019] is utilized to train CELDM. The fine-tuning of the trained CELDM is conducted in an RL environment, with experiments carried out using different numbers of candidate responses (n = 2, 3, 4, 5, 10). Based on the obtained loss, the value of n = 3 is selected as the final choice. For decoding the generated candidates, nucleus sampling, following [Holtzman et al., 2019], is employed with a temperature of T = 0.8 and a probability threshold of p = 0.9. During training, several parameters are considered: the seed value is set to 10, the human reward is set to 10, the maximum candidate length is limited to 50, and the AdamW optimizer [Loshchilov and Hutter, 2018] is employed with a learning rate of α = 2e-05 , an epsilon value of ε = 0.2 and epochs = 20. The final weights chosen for the reward combinations are 0.3, 0.25, 0.2, 0.15, 0.1 for β1, β2, β3, β4 and β5 respectively. Lastly, a penalization factor of γ = 2 is set for rewards R1, R2 and R3. Evaluation Metrics. The performance of the proposed system, PARTNER, is evaluated through both automatic and human evaluations. The metrics employed to assess the performance of counselling dialogue act, empathy strategy, and politeness strategy classifiers are Weighted Accuracy (W-ACC) and Macro-F1. The efficacy of PARTNER is evaluated in terms of two aspects viz. task success (counselling, empathy, and politeness) and response quality (fluency, consistency, and response length). For automatic evaluation, task success is measured using metrics such as CoAct (consistency of predicted counselling acts with ground truth), EmpStr (consistency of predicted empathy strategies with ground truth label), and PolStr (consistency of predicted politeness strategies with ground truth) 9 . Response quality is evaluated using metrics such as Perplexity (PPL) and response length (R-LEN). Table 4: Results of automatic evaluation. PARTNER refers to our proposed system considering all rewards. PARTNER-R, PARTNER-GR and PARTNER-TR refers to PARTNER with no rewards, without generic reward and without task-specific reward respectively. | Model | CoAct | EmpStr | PolStr | PPL | R-LEN | |------------------------|---------|----------|----------|-------|---------| | ARDM [Wu et al., 2021] | 52.8% | 57.2% | 66.1% | 3.74 | 14.8 | | PARTNER-R | 51.9% | 57.3% | 66.3% | 3.68 | 14.4 | | PARTNER-GR | 55.3% | 59.1% | 69.3% | 3.17 | 15.7 | | PARTNER-TR | 53.8% | 58.2% | 67.8% | 3.31 | 15.4 | | PARTNER | 56.5% | 61.8% | 69.9% | 2.55 | 16.06 | To conduct the human evaluation, we recruited three evaluators holding postgraduate qualifications and having experience in the related works 10 . Initially, each evaluator engaged with the system ten times, with a fresh set of responses provided for each interaction. Subsequently, domain experts from government-run institutions are asked to cross-verify these 30 human-evaluated interactions to ensure the quality of the evaluation process. Upon passing the quality check, additional 45 interactions are evaluated, with each evaluator interacting and assessing 15 dialogues. As a result, we obtained a total of 75 human-evaluated dialogues. The human evaluation metrics encompass several aspects. For task success, we considered the correctness of the counselling act (CounC), empathy strategy (EmpC), and politeness strategy (PolC). To evaluate response quality, we focused on naturalness (Nats), correctness (Corr), and non-repetitiveness (NRep). All dialogue interactions were assessed on an integer Likert scale ranging from 1 to 5 11 . The evaluation process exhibited an inter-evaluator agreement ratio of 75.1%. 9 CoAct, EmpStr and PolStr is evaluated using counselling act, empathy and politeness strategy classifiers Table 5: Results of human evaluation. | Model | CounC | EmpC | PolC | Nats | Corr | NRep | |-----------|---------|--------|--------|--------|--------|--------| | ARDM | 2.44 | 2.60 | 2.97 | 4.10 | 3.84 | 3.71 | | PARTNER-R | 2.36 | 2.66 | 3.01 | 4.21 | 3.91 | 3.82 | | PARTNER | 3.06 | 3.11 | 3.42 | 4.52 | 4.14 | 4.07 | ## 6 Results and Analysis To begin, the first step involves analyzing the subcomponents, namely the counselling act, empathy, and politeness classifiers. Following that, the results of the proposed PARTNER system are presented and compared against four baselines: ARDM (CELDM) [Wu et al., 2021], PARTNERR (PARTNER with zero reward), PARTNER-GR (PARTNER without generic rewards), and PARTNER-TR (PARTNER without task-specific rewards). Table 3 provides an overview of the evaluation results for the classifiers. The results indicate that the classifiers perform well in terms of Weighted Accuracy (W-ACC) and Macro-F1. Moreover, it is observed that RoBERTa-large [Liu et al., 2019] outperforms BERTlarge [Devlin et al., 2018]. The good scores obtained from the classifiers demonstrate their effectiveness in designing R1, R2, and R3 rewards. Automatic evaluation. Table 4 illustrates that PARTNER outperforms the baselines: ARDM, PARTNER-R, PARTNER-GR, and PARTNER-TR across all metrics. The metrics CoAct, EmpStr, and PolStr obtain good scores of 56.5%, 61.8%, 69.9% with a significant difference of < 3.7, 4.6, 3.8 >, < 4.6, 4.5, 3.6 >, < 0.3, 2.7, 0.6 > and < 1.8, 3.6, 2.1 >, respectively in comparison to the baselines ARDM, PARTNER-R, PARTNER-GR, and PARTNERTR 12 . These results provide justification for the design of our reward function in PARTNER. By incorporating task-specific rewards, the counselling bot is incentivized to generate appropriate counselling act, empathy, and politeness strategyadaptive responses. These findings strengthen our hypothesis that rewards for counselling dialogue act consistency, empathy, and politeness strategies contribute to the development of a more engaging, persuasive, empathetic, and polite counselling dialogue system. Table 4 demonstrates that PARTNER achieves significantly high scores, with PPL and R-LEN = 2.55 and 16.06, respectively when compared to the baselines: ARDM, PARTNER-R, PARTNER-GR, and PARTNER-TR with a difference of < 1.19, 1.26 >,< 1.13, 1.66 >,< 0.62, 0.36 > and < 0.76, 0.66 >, respectively. This can be attributed to the utilization of generic rewards, which encourage the model to establish a natural and smooth connection with the victim, resulting in interactive and engaging conversations. Additionally, it is worth noting that PARTNERGR achieves lower PPL, while PARTNER-TR attains lower scores in CoAct, EmpStr, and PolStr, which highlights the importance of both generic rewards and task-specific rewards in the system, respectively. 10 Evaluators are paid in accordance with the university norms. 11 Rating of 1 denoted low and 5 represented high performance in the respective metric 12 We performed a statistical significance test, Welch's t-test [Welch, 1947], and it is conducted at 5% (0.05) significance level. Human evaluation. The evaluation results from the human assessment are presented in Table 5. A comparison between the proposed system, PARTNER and the baselines, ARDM and PARTNER-R, reveals that PARTNER achieves higher scores in CounC, EmpC, PolC, Nats, Corr, and NRep with a with a difference of < 0.62, 0.70 >, < 0.51, 0.45 > ,< 0.45, 0.41 >,< 0.42, 0.31 >,< 0.3, 0.23 >, and < 0.36, 0.25 >, respectively. These results indicate that all four rewards have played a vital role in generating accurate, natural, and non-repetitive counselling utterances with politeness and empathy imbibed in them. Therefore, it can be concluded that incorporating rewards for empathy and politeness strategies enhances the ability of PARTNER to establish a connection with the victim by generating engaging and interactive responses. ## 7 Conclusion and FutureWork The prevention/reduction in CAW&C will foster a nurturing, equitable and secure environment for all, which consequently will improve mental health, ensure a safer and fairer world for all and thereby promote sustainable development. Taking into account the importance of preventing/reducing CAW&C in achieving the SDGs, in this paper, we proposed a politeness and empathy strategies-adaptive persuasive mental health and legal counselling dialogue system, PARTNER to help the crime victims by offering (or sometimes persuading them for seeking) pertinent counselling support in a cordial and compassionate setting. For this, we prepared a novel counselling conversational dataset, HEAL and utilized it to train PARTNER in an RL framework by designing novel rewards for facilitating correct counselling act, politeness and empathy strategies in the counselling bot's response. Detailed automatic and human evaluation results conclude that PARTNER is capable of performing adequate counselling act and adopt desired politeness and empathy strategies whilst ensuring engagingness and naturalness during the counselling session. A counselling system can cover different aspects of victim's personal traits to counsel each victim in a different way. This drives us towards future directions for our current research work. ## Contribution Statement Priyanshu Priya and Kshitij Mishra are jointly the first authors in this work. Subsequently, Palak Totala and Asif Ekbal are second and third authors respectively. ## Ethical Statement To avoid any possible harm or potential for harm, it is crucial for a counsellor (here, counselling dialogue system) to follow some ethics principles [Association and others, 2016]. All responses should be formed with upholding the victim's rights, maintaining appropriate professional competence, safeguarding the welfare of the victim, obtaining informed consent and delivering authentic information. Therefore, to create the dialogues in the dataset, first authentic websites and documents are referred to; then, the dataset is duly verified by experts from government-run institutions. A counselling dialogue system on the one hand can counsel and provide support, whereas, on the other hand, it can also be used to mould a person's beliefs, which poses a big challenge in the development of these systems. Due to the sensitivity of the data content and system, the codes and data would only be provided for research purposes through appropriate data agreement procedures. ## Acknowledgments Priyanshu Priya acknowledges the Innovation in Science Pursuit for Inspired Research (INSPIRE) Fellowship implemented by the Department of Science and Technology, Ministry of Science and Technology, Government of India for financial support. Kshtij Mishra gratefully acknowledges the assistance of the Indian government's "Prime Minister's Research Fellowship (PMRF) Program". Authors acknowledge the partial support from the project titled "An Empathetic Knowledge Grounded Conversational System for Mental Health Counseling and Legal Assistance", Sponsored by IHUB Anubhuti, TIH, IIIT Delhi. ## References 1. [Abroms, 1968] Gene M Abroms. Persuasion in psychother- apy. American Journal of Psychiatry, 124(9):1212-1219, 1968. 2. [Althoff et al., 2016] Tim Althoff, Kevin Clark, and Jure Leskovec. Large-scale analysis of counseling conversa- tions: An application of natural language processing to mental health. Transactions of the Association for Com- putational Linguistics, 4:463-476, 2016. 3. [Association and others, 2016] American Psychological As- sociation et al. Revision of ethical standard 3.04 of the" ethical principles of psychologists and code of con- duct"(2002, as amended 2010). The American Psycholo- gist, 71(9):900, 2016. 4. [Brown et al., 1987] Penelope Brown, Stephen C Levinson, and Stephen C Levinson. Politeness: Some universals in language usage, volume 4. Cambridge university press, 1987. [Casanueva et al., 2018] Inigo Casanueva, Paweł Budzianowski, Pei-Hao Su, Stefan Ultes, Lina Rojas- Barahona, Bo-Hsiang Tseng, and Milica Gašić. Feudal reinforcement learning for dialogue management in large domains. arXiv preprint arXiv:1803.03232, 2018. 5. [Demasi et al., 2019] Orianna Demasi, Marti A Hearst, and Benjamin Recht. Towards augmenting crisis counselor training by improving message retrieval. In Proceedings of the Sixth Workshop on Computational Linguistics and Clinical Psychology, pages 1-11, 2019. 6. [Devlin et al., 2018] Jacob Devlin, Ming-Wei Chang, Ken- ton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805, 2018. 7. [Dorn et al., 1986] Fred J Dorn, Barbara Kerr, Mark J Miller, and C Edward Watkins. Breaking into print: Guidelines for mental health counselors. American Mental Health Counselors Association Journal, 1986. 8. [Elliott et al., 2018] Robert Elliott, Arthur C Bohart, Jeanne C Watson, and David Murphy. Therapist em- pathy and client outcome: An updated meta-analysis. Psychotherapy, 55(4):399, 2018. 9. [Fitzpatrick et al., 2017] Kathleen Kara Fitzpatrick, Alison Darcy, and Molly Vierhile. Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (woe- bot): a randomized controlled trial. JMIR mental health, 4(2):e7785, 2017. 10. [Fogg, 2002] Brian J Fogg. Persuasive technology: using computers to change what we think and do. Ubiquity, 2002(December):2, 2002. 11. [Fulmer et al., 2018] Russell Fulmer, Angela Joerin, Bre- anna Gentile, Lysanne Lakerink, Michiel Rauws, et al. Using psychological artificial intelligence (tess) to relieve symptoms of depression and anxiety: randomized con- trolled trial. JMIR mental health, 5(4):e9782, 2018. 12. [Gordon, 1993] Thomas F Gordon. The pleadings game: for- malizing procedural justice. In Proceedings of the 4th in- ternational conference on artificial intelligence and law, pages 10-19, 1993. 13. [Heppner and Claiborn, 1989] P Paul Heppner and Charles D Claiborn. Social influence research in counseling: A review and critique. Journal of counseling psychology, 36(3):365, 1989. 14. [Holtzman et al., 2019] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neu- ral text degeneration. arXiv preprint arXiv:1904.09751, 2019. 15. [Jaccard, 1912] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 16. [John et al., 2017] Adebayo Kolawole John, Luigi Di Caro, Livio Robaldo, and Guido Boella. Legalbot: A deep learning-based conversational agent in the legal domain. In International Conference on Applications of Natural Lan- guage to Information Systems, pages 267-273. Springer, 2017. 17. [Kim et al., 2018] Junhan Kim, Yoojung Kim, Byungjoon Kim, Sukyung Yun, Minjoon Kim, and Joongseek Lee. Can a machine tend to teenagers' emotional needs? a study with conversational agents. In Extended Abstracts of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1-6, 2018. 18. [Kim et al., 2022] Hyeok Kim, Youjin Hwang, Jieun Lee, Youngjin Kwon, Yujin Park, and Joonhwan Lee. Per- sonalization trade-offs in designing a dialogue-based in- formation system for support-seeking of sexual violence survivors. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-15, 2022. 19. [Liang et al., 2021] Kai-Hui Liang, Patrick Lange, Yoo Jung Oh, Jingwen Zhang, Yoshimi Fukuoka, and Zhou Yu. Evaluation of in-person counseling strategies to develop physical activity chatbot for women. arXiv preprint arXiv:2107.10410, 2021. 20. [Liu et al., 2019] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 21. [Loshchilov and Hutter, 2018] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Inter- national Conference on Learning Representations, 2018. 22. [Lucas et al., 2014] Gale M Lucas, Jonathan Gratch, Aisha King, and Louis-Philippe Morency. It's only a computer: Virtual humans increase willingness to disclose. Comput- ers in Human Behavior, 37:94-100, 2014. 23. [McHugh, 2012] Mary L McHugh. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276-282, 2012. 24. [Mesgar et al., 2020] Mohsen Mesgar, Edwin Simpson, and Iryna Gurevych. Improving factual consistency be- tween a response and persona facts. arXiv preprint arXiv:2005.00036, 2020. 25. [Mishra et al., 2022a] Kshitij Mishra, Mauajama Firdaus, and Asif Ekbal. Please be polite: Towards building a po- liteness adaptive dialogue system for goal-oriented conver- sations. Neurocomputing, 494:242-254, 2022. 26. [Mishra et al., 2022b] Kshitij Mishra, Azlaan Mustafa Samad, Palak Totala, and Asif Ekbal. Pepds: A polite and empathetic persuasive dialogue system for charity dona- tion. In Proceedings of the 29th International Conference on Computational Linguistics, pages 424-440, 2022. 27. [Norcross, 2002] John C Norcross. Psychotherapy relation- ships that work: Therapist contributions and responsive- ness to patients. Oxford University Press, 2002. 28. [Orji, 2014] Rita Orji. Design for Behaviour Change: A Model-driven Approach for Tailoring Persuasive Tech- nologies. PhD thesis, University of Saskatchewan Saska- toon, 2014. 29. [Radford et al., 2019] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 30. [Robert et al., 2011] Elliot Robert, Arthur C Bohart, JC Wat- son, and LS Greenberg. Empathy. Psychotherapy, 48(1):43-49, 2011. 31. [Roberts, 2002] Albert R Roberts. Handbook of domestic vi- olence intervention strategies: Policies, programs, and le- gal remedies. Oxford University Press, 2002. 32. [Saha et al., 2022a] Tulika Saha, Vaibhav Gakhreja, Anindya Sundar Das, Souhitya Chakraborty, and Sriparna Saha. Towards motivational and empathetic response generation in online mental health support. In Proceed- ings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2650-2656, 2022. 33. [Saha et al., 2022b] Tulika Saha, Saichethan Reddy, Anindya Das, Sriparna Saha, and Pushpak Bhattacharyya. A shoulder to cry on: Towards a motivational virtual assistant for assuaging mental agony. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2436-2449, 2022. 34. [Schulman et al., 2017] John Schulman, Filip Wolski, Pra- fulla Dhariwal, Alec Radford, and Oleg Klimov. Prox- imal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 35. [Strong, 1968] Stanley R Strong. Counseling: An interper- sonal influence process. Journal of counseling Psychol- ogy, 15(3):215, 1968. 36. [Welch, 1947] Bernard L Welch. The generalization of 'student's'problem when several different population var- lances are involved. Biometrika, 34(1-2):28-35, 1947. 37. [Williamson et al., 2010] E Williamson, NM Dutch, and HJ Clawson. Evidence-based mental health treatment for victims of human trafficking. washington, dc: Office of the assistant secretary for planning and evaluation, us depart- ment of health and human services, 2010. 38. [Wu et al., 2021] Qingyang Wu, Yichi Zhang, Yu Li, and Zhou Yu. Alternating recurrent dialog model with large- scale pre-trained language models. In Proceedings of the 16th Conference of the European Chapter of the Associa- tion for Computational Linguistics: Main Volume, pages 1292-1301, 2021.