cs.CLSep 3, 2026
SaveAdapting from Downturns: Prediction of Long-Term Conversational-Skill Development in Mental-Health Crisis Counselors
Organizations: Cornell University · Crisis Text Line
Abstract
How do people learn to become better conversationalists? This question is especially important in the context of mental-health counseling, where conversational skills are essential, yet volunteer counselors often have limited access to supervision and structured feedback. Understanding how counselors develop their ability to steer conversations toward positive outcomes -- and identifying early which counselors are (not) on track to improve -- can help prioritize support for the counselors who need it most. In this work, we introduce the task of predicting, early in a conversationalist's career, whether they will eventually improve at steering conversations toward positive outcomes, and demonstrate the feasibility of this task in the case of volunteer mental-health crisis counselors. Our central insight is that people may struggle with particular kinds of moments in a conversation, and that what is especially revealing of their likelihood of future improvement is how they learn to handle those moments over time. We operationalize this insight by designing a method that identifies the types of moments a counselor initially struggles with, captures how they adapt their response when they re-encounter similar moments in subsequent conversations, and learns which early adaptations predict improvement months or even years later. While this future-prediction task is challenging, our counselor-adaptation approach yields better results than baselines that learn directly from the conversation transcript.
Figures & tables
Figure 1: High-level overview of our four-step approach.
| Statistic | Value |
| Average conversation length (messages) | 28 |
| Average words per message | 22 |
| Average conversation duration (min) | 45 |
| Number of counselors | 1488 |
| Number of conversations | 148800 |
| % conversations with scores 1, 2, 3, 4, 5 | 8, 20, 20, 29, 23 |
Table 1: Descriptive statistics of the dataset used for our improvement-prediction task.
| Method | Val | Test |
| Downturn-based moments | ||
| Downturn (unsuccessful) paired similar moment | 59.7 | 59.2 |
| Downturn (all) paired similar moment | 59.3 | 56.9 |
| Downturn (unsuccessful) paired random moment | 59.4 | 58.9 |
| Alternative moments | ||
Table 2: Accuracy in detecting whether a counselor improved across different options for moment selections (early late). The option that performed best on validation was used for reinforcement learning (bottom).
| Repres. | Train | RL | Acc | P | R | F1 |
| msg. | 56.9 | 57.6 | 53.2 | 54.8 | ||
| msg. | 57.5 | 58.5 | 52.3 | 54.5 | ||
| diversity | 54.6 | 54.4 | 57.0 | 55.7 | ||
| time | 47.0 | 46.7 | 43.6 | 45.1 | ||
| residual | 56.7 | 55.4 | 68.4 | 61.2 | ||
| adapt. | ✓ | 61.9 | 61.5 | 66.0 | 63.1 |
Table 3: Results for the task of predicting future improvement from early-career conversations. We compare baselines (top) against our full system (middle), and ablations of the full system (bottom).
| Early-career response | Late-career response | Adaptation summary (excerpt) |
| T: I’m feeling suicidal C: Hi, I’m redacted . I’m here for you, tell me more. T: I’ve been wanting to die for a week and idk what to do C: You’re brave for being so open with me. Would you be comfortable telling me your first name? | T: I am suicidal C: Hi, I hear you’re feeling overwhelmed tonight. Can you tell me more? T: I feel alone and wanna die C: I’m here today to support you. What’s happened recently that made you feel this way? | […] demonstrates a consistent pattern of offering support and expressing concern for the individual’s well-being […] leans more heavily on direct statements of care […] the approach is generally more proactive in seeking clarification about the individual’s situation and needs […] |
| T: I want to die. C: Has anything happened to make you want to self-harm? T: I felt terrible and getting blamed everything C: I am sorry for your loss hon, it can be rather nasty when people unjustly blame you. Have you felt like this for a while? | T: I’ll be dead soon. I wish I had spared myself years of pain. C: You’re upset you are going to die soon? T: I am upset that my life ended. I am essentially already dead. C: I am sorry to hear that you are feeling this way, though do you still want to continue talking? | While [in early early-career responses the counselor] frequently used terms of endearment like “hon” and offered reassurance, [in late-career responses the counselor] opts for more direct and questioning statements […] consistently inquisitive tone […] more distanced interaction style […] move[s] towards a more objective communication. |
Table 4: Paraphrased downturn and later analogous moments with adaptation summary excerpts.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Distribution of counselor improvement.
Figure 3: Distribution of the change in variance of outcomes between late and early careers of improved and nonimproved counselors.
| Method | Acc | P | R | F1 |
| Downturn (unsuccessful) paired similar moment | 59.7 | 59.1 | 63.8 | 61.1 |
| Downturn (all) paired similar moment | 59.3 | 58.6 | 64.2 | 61.1 |
| Downturn (unsuccessful) paired random moment | 59.4 | 58.2 | 67.7 | 62.3 |
| High sentiment drop (unsuccessful) paired similar moment | 56.8 | 57.0 | 63.0 | 58.7 |
| Random moment (unsuccessful) paired similar moment | 57.9 | 58.4 | 59.6 | 58.1 |
| Random moment (all) paired similar moment | 55.2 | 57.7 | 37.3 | 45.0 |
Table 5: Full validation results for detecting whether a counselor has improved across different moment selections and pairings (early late). Averaged across multiple seeds.
| Method | Acc | P | R | F1 |
| Downturn (unsuccessful) paired similar moment | 59.2 | 58.5 | 65.5 | 61.4 |
| Downturn (all) paired similar moment | 56.9 | 56.0 | 65.6 | 60.1 |
| Downturn (unsuccessful) paired random moment | 58.9 | 58.1 | 63.3 | 60.4 |
| High sentiment drop (unsuccessful) paired similar moment | 54.6 | 54.4 | 62.2 | 56.8 |
| Random moment (unsuccessful) paired similar moment | 55.7 | 55.5 | 57.5 | 55.6 |
| Random moment (all) paired similar moment | 52.2 | 52.9 | 40.6 | 45.7 |
Table 6: Full test results for detecting whether a counselor has improved across different moment selections and pairings (early late). Averaged across multiple seeds.
Figure 6: Box and whiskers plot comparing the accuracy of our full system (right) and the best baseline: msg. (left). Whiskers not visible due to overlap with box edges.
| Not before downturn | Before downturn | ||
| n-gram | n-gram | ||
| have to go | my name is | ||
| do you think | name is scrubbed | ||
| what do you | want to die | ||
| going to go | and want to | ||
| no no no | to kill myself | ||
Table 7: Top 30 tri-grams associated with texter messages that appear right before a downturn (right) or not (left), ranked by the z-score of the Bayesian distinguishing-word analysis.
| Downturn | Distinguishing trigrams |
| Seeking support, clarification | “yes it is”, “is this real”, “this real person”, “okay thank you”, “are you real”, “do you mean”, “don think so”, “yes ma am”, “what do you”, “no not really”, “do have to”, “ok thank you”, “no don know”, “thank you for”, “how do you”, “would like to”, “you real person”, “at the moment”, “know my name”, “old are you” |
| Introduction, self-disclosure | “name is scrubbed”, “my name is”, “scrubbed my name”, “hi scrubbed scrubbed”, “my names scrubbed”, “call me scrubbed”, “is scrubbed and”, “hi scrubbed my”, “hi my name”, “scrubbed thank you”, “thank you scrubbed”, “can call me”, “yes my name”, “to do scrubbed”, “talk to scrubbed”, “you can call”, “hello my name”, “nice to meet”, “scrubbed scrubbed scrubbed”, “my name scrubbed” |
| Isolation, hopelessness | “have no one”, “just feel like”, “feel so alone”, “in my life”, “my life is”, “like have no”, “to be happy”, “have no friends”, “feel like have”, “like no one”, “and feel like”, “no one to”, “feel like not”, “feel like no”, “just feel so”, “to do anymore”, “don feel like”, “feel like giving”, “feel like my”, “and have no” |
| Relationships, conflict | “he told me”, “he said he”, “my best friend”, “me and my”, “mad at me”, “to me and”, “he says he”, “doesn want to”, “broke up with”, “to help him”, “talk to her”, “at me and”, “he wants to”, “up with me”, “and he just”, “we ve been”, “him and he”, “want him to”, “with him and”, “talk to him” |
| Suicidal ideation, self-harm | “want to die”, “to kill myself”, “want to kill”, “don want to”, “kill my self”, “going to kill”, “want to cut”, “end my life”, “to end my”, “home want to”, “to kill my”, “to be alive”, “help want to”, “to cut myself”, “start want to”, “want to end”, “hello want to”, “hate my life”, “to commit suicide”, “want to be” |
| Anxiety, depression, panic attacks | “having panic attack”, “an anxiety attack”, “anxiety and depression”, “my anxiety is”, “depression and anxiety”, “lot of anxiety”, “having an anxiety”, “go to sleep”, “having really bad”, “really bad anxiety”, “to calm down”, “my thoughts are”, “my anxiety and”, “ve been having”, “panic attack and”, “an eating disorder”, “my mind is”, “eating disorder and”, “during the day”, “bad anxiety and” |
Table 8: Distinguishing trigrams by downturn type.
| # | Downturn | Early career response | Late career response | Summary excerpt |
| 1 | Suicidal ideation, self-harm | T: I’m feeling suicidal C: Hi, I’m scrubbed. I’m here for you, tell me more. T: I’ve been wanting to die for a week and idk what to do C: You’re brave for being so open with me tonight. Would you be comfortable telling me your first name? | T: I am suicidal C: Hi, I hear you’re feeling overwhelmed tonight. Can you tell me more? T: I feel alone and wanna die C: I’m here today to support you. What’s happened recently that made you feel this way? | […] demonstrates a consistent pattern of offering support and expressing concern for the individual’s well-being […] leans more heavily on direct statements of care […] the approach is generally more proactive in seeking clarification about the individual’s situation and needs […] |
| 2 | Introduction, self-disclosure | T: I can’t meet my family’s expectations. It’s too much C: It makes sense to feel this way. How long has this been? T: A year, I just need someone to talk to C: I am here to help you through your crisis. You will not always feel this way, it’ll get better. Do you mind sharing your name with me? | T: My names scrubbed, my family is too much C: That must be really tough to have your family shut you down like that. How are you feeling? T: I feel frustrated and all alone C: Thank you for sharing that with me. You’re not alone anymore, I have your back and you can tell me anything if you want? | […] shift towards a more supportive and encouraging tone […] more focused on validating the speaker’s feelings and strengths, expressing gratitude for their sharing […] prioritizing reassurance and encouragement over detailed questioning […] overall interactional style is characterized by a greater degree of empathy |
| 3 | Relationships, conflict | T: Okay, I’m scrubbed. C: Thanks, scrubbed. Can you tell me more about the conflict going on at home? T: My dad is angry and rude towards all the time. I’m afraid to be myself around him because he’s so negative. C: That sounds like an awful situation to live with. How long has this been? | T: Yes it’s scrubbed. C: It sounds like you’re under much stress, tell me more? T: I’m going to attend dream music school. My dad can be abusive if he found out C: You sound like a really creative person, congrats! It must be super stressful w/your dad’s reaction, you don’t deserve that. | […] frequently starts with expressions of empathy and concern […] utilizes phrases of affirmation and encouragement […] overall approach is less exploratory and more immediately concerned with providing a supportive and protective environment […] |
| 4 | Isolation, hopelessness | T: I used to be really happy but this sadness just hit me C: I am glad you shared that with me. It seems that you are feeling sad but you don’t really know why. T: Exactly it’s hopeless C: That must be frustrating for you. | T: What’s wrong with me C: That’s not something I can answer, how are you feeling? T: I’ve been unusually sad for a long time I can’t remember when I was happy C: Have you considered journaling about your feelings to sort them out? | […] tone is considerably more focused on problem-solving […] conversational style is less tentative, with a move toward suggesting coping mechanisms like journaling […] move toward a more active and interventionist role |
| 5 | Suicidal ideation, self-harm | T: I want to die. C: Has anything happened to make you want to self-harm? T: I felt terrible and getting blamed everything C: I am sorry for your loss sweetheart, it can be rather nasty when people unjustly blame you for things. Have you felt like this for a while? | T: I’ll be dead soon. I wish I had spared myself years of pain. C: You’re upset you are going to die soon? T: I am upset that my life ended. I am essentially already dead. C: I am sorry to hear that you are feeling this way, though do you still want to continue talking? | While [early set] frequently used terms of endearment like “sweetheart” and offered reassurance, [late set] opts for more direct and questioning statements. […] more distanced interaction style […] indicates a move towards a more objective communication […] |
| 6 | Isolation, hopelessness | T: I’m lonely, I have no friends C: Can you tell me more? T: I just lost my best friend since she was making me try to change myself. I have nobody. C: That sounds devastating. You deserve to be treated with respect. Are you comfortable sharing your first name with me? | T: I’m kinda lonely C: What’s going on? T: My girlfriend is away, all my friends are gone so I’m alone. C: It can be hard when the people you rely on aren’t around. That can be tough to manage. | […] acknowledging the speaker’s experience and validating their feelings […] phrasing is slightly more conversational and less formal […] the communication style remains consistent in its empathy, but it’s subtly altered by a stronger desire to understand the speaker’s perspective […] |
Table 9: Paraphrased downturn and later analogous moments with adaptation summary excerpts, separated by improved counselors (top) and nonimproved counselors (bottom). Adaptation summary compares the entire collection of downturns and analogous moments from the same counselor, not just the one exemplified.
More in this topic
Human Behavior Prediction
Human behavior prediction models how people will act, decide, move, or feel from their histories, context, and observable signals.
See all 75 papers