💬 NLP

CARE: Counselor-Aligned Response Engine for Online Mental-Health Support

The paper introduces CARE, a counselor-aligned response engine that fine-tunes open-source large language models on expert-validated crisis conversations in Hebrew and Arabic to generate real-time, psychologically supportive responses that enhance counselor workflows and care quality in low-resource language contexts.

Original authors: Hagai Astrin, Ayal Swaid, Avi Segal, Kobi Gal

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hagai Astrin, Ayal Swaid, Avi Segal, Kobi Gal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🚨 The Problem: The Overwhelmed Lifeguard

Imagine a busy beach where thousands of people are in distress. The lifeguards (counselors) are volunteers who have been trained to save lives. They are kind, patient, and know exactly what to say to calm someone down.

But lately, the beach is too crowded. The lifeguards are drowning in requests. They have to talk to multiple people at once, often while the situation is getting dangerous (like someone thinking about jumping off a cliff). When they are stressed and tired, they might:

Take too long to reply.
Say something generic like, "Don't worry, it'll be fine," which can actually make the person feel unheard.
Forget the specific, gentle techniques they were trained to use.

🤖 The Old AI: The "Generic" Robot

Scientists tried to fix this by bringing in a robot assistant (a standard Large Language Model or LLM). But this robot was like a smart tourist who knows a lot of facts but has never been a lifeguard.

It speaks fluently, but it sounds like a textbook.
It might give medical advice it shouldn't give.
It doesn't "get" the emotional nuance. If a person is crying, the robot might say, "Here are 5 steps to solve your problem," instead of saying, "I hear how much pain you are in."

💡 The Solution: CARE (The "Apprentice" Lifeguard)

The researchers built a new system called CARE (Counselor-Aligned Response Engine). Think of CARE not as a replacement for the lifeguard, but as a super-smart apprentice sitting right next to them.

Here is how they built this apprentice:

1. The Training Camp (Data)

Instead of teaching the robot with random internet chats, the researchers gave it a secret library of 30,000+ real conversations from "Sahar," a real crisis hotline in Israel.

The Secret Sauce: They didn't just give the robot any chat. They only gave it the best chats—the ones where the counselor successfully calmed the person down.
The Language: They taught it specifically in Hebrew and Arabic, languages that are often ignored by big tech companies.

2. The "Full Story" Method (Full-History Training)

This is the most important part.

Old AI: Usually, AI looks at the last sentence you said and guesses the next one. It's like reading a book one page at a time without remembering the plot.
CARE: CARE reads the entire book before guessing the next page. It sees the whole conversation history. It knows if the person was angry 10 minutes ago, or if they just started to feel a little hopeful. It understands the story, not just the sentence.

3. The Result: The "Shadow"

When a real counselor is typing a message, CARE whispers a suggestion in their ear.

The Old Robot: "You should call a doctor." (Too clinical, too cold).
CARE: "It sounds like you feel completely alone right now. I'm here with you. Can you tell me one small thing that might make you feel safer tonight?" (Warm, validating, and strategic).

🏆 The Scoreboard: Did It Work?

The researchers put CARE to the test against the "generic" robots. The results were like a David vs. Goliath story, but with the underdog winning big:

Strategic Alignment (SIM): CARE learned when to use specific techniques (like "Reflecting feelings" or "Asking gentle questions") just by watching the best counselors. It matched the pros' style significantly better than the generic models.
Language Flow: In Arabic, CARE's responses were 20 times more accurate and natural-sounding than the baseline models. In Hebrew, it was 6 times better.
The "Vibe" Check: They measured how "human" the text sounded. CARE sounded much more like a compassionate counselor and much less like a computer.

🛡️ The Safety Rules

The paper is very careful to say: CARE does not replace the human.

The Pilot and the Co-Pilot: The human counselor is the pilot. CARE is the co-pilot suggesting the best route. The human always presses the "send" button.
Privacy: They scrubbed the data so no real names or phone numbers were ever used.
Ethics: The system is programmed to never give medical diagnoses, only emotional support.

🌟 The Big Picture

This paper shows that if you take a smart AI and train it specifically on the best examples of human kindness (in specific languages), it can become a powerful tool to help overworked counselors save lives. It bridges the gap between "smart computer" and "compassionate human," ensuring that even in a crisis, the person in distress feels truly heard.

1. Problem Statement

The paper addresses the critical shortage of mental health support resources, particularly in low-resource languages (Hebrew and Arabic) and high-stakes crisis scenarios (e.g., suicide prevention).

Counselor Overload: Online crisis hotlines face increasing traffic, leading to delayed responses and cognitive burnout for volunteer counselors. Timely intervention is vital for safety.
Limitations of General LLMs: While Large Language Models (LLMs) can generate empathetic text, general-purpose models often fail in crisis contexts. They tend to produce generic reassurance ("Don't worry, it will be fine") or directive medical advice, which can invalidate a user's experience or violate the specific supportive language and de-escalation strategies used by professional counselors.
Data Scarcity: Existing mental health AI research is heavily skewed toward English. There is a lack of large-scale, fine-tuned generative models trained on real-world crisis dialogues in Hebrew and Arabic that adhere to specific organizational protocols.

2. Methodology: The CARE Framework

The authors propose CARE (Counselor-Aligned Response Engine), a GenAI framework designed to assist counselors by generating real-time, psychologically aligned response recommendations.

A. Data Source and Curation

Dataset: The system is trained on data from Sahar, a national suicide-prevention chatline serving Arabic and Hebrew communities.
Selection Criteria: To ensure high quality, the authors curated a subset of 1,000 sessions (800 for training, 200 for testing per language) that received high clinical effectiveness scores (VED score of 4 or 5) from professional counselors.
Anonymization: A rigorous two-tier anonymization process was applied:
- Hebrew: Used the HebSafeHarbor package (developed by Clalit Health Services) for PHI removal.
- Arabic: Used a fine-tuned CAMeLBERT NER model, followed by regex filtering.
Structure: Unlike standard turn-by-turn training, the data was structured as full conversation histories. Each training sample consists of the entire dialogue history up to the point of the counselor's intervention, allowing the model to learn longitudinal emotional trajectories.

B. Model Architecture and Training

Base Model: The authors fine-tuned the open-source Gemma-3-12B-it model.
Training Technique:
- Supervised Fine-Tuning (SFT): The model was trained on the curated corpus using LoRA (Low-Rank Adaptation) to optimize memory usage (trained on a single RTX 4090).
- Masked Loss Objective: A critical technical innovation. The loss function was masked to penalize the model only on the counselor's response tokens, ignoring the help-seeker's input. This forces the model to focus exclusively on learning the specific supportive language patterns and intervention strategies of the counselors, rather than predicting generic conversational flow.
Inference: The model receives the full dialogue history and a language-specific system prompt enforcing empathy, non-judgment, and safety constraints (e.g., avoiding clinical diagnoses).

C. Evaluation Metrics

The study introduces a novel metric alongside standard NLP metrics:

Support Intent Match (SIM): A novel metric designed to measure strategic alignment. A Strategy Classifier (trained on expert-annotated data) categorizes responses into four strategies: Reflection, Prompting, Suggestion, and Neutral. SIM calculates the proportion of times the model's predicted strategy matches the gold-standard counselor's strategy.
Semantic/Lexical Similarity: BERTScore, ROUGE (1, 2, L), and BLEU scores to measure semantic fidelity and lexical overlap with human responses.
Supportive Language Perplexity (PPL): A GPT-2 model trained exclusively on counselor utterances was used to measure how well the generated text conforms to the specific stylistic and tonal patterns of professional counselors.

3. Key Contributions

First Domain-Specific Generative Model for Hebrew/Arabic Crisis Support: CARE is the first framework to fine-tune a large-scale LLM on professional suicide-prevention chat data in these languages, moving beyond risk classification (discriminative) to supportive generation.
Full-History Fine-Tuning Strategy: Demonstrates that training on complete conversation trajectories (rather than isolated turns) enables LLMs to implicitly learn complex, longitudinal counseling strategies without explicit strategy labels.
Novel Evaluation Metric (SIM): Introduces a method to quantify the alignment of an AI's intervention strategy with human experts, not just its linguistic similarity.
Privacy-Preserving Pipeline: Establishes a robust workflow for anonymizing sensitive mental health data in morphologically rich languages (Hebrew/Arabic) for AI training.

4. Results

The CARE model significantly outperformed unadapted baseline models (Gemma-3-12B-it, Llama-3.1-8B-it, and Qwen3-14B) across all dimensions:

Strategic Alignment (SIM): CARE achieved a SIM score of 0.53 (Hebrew) and 0.74 (Arabic), significantly higher than baselines. This proves the model implicitly learned when to use specific strategies (e.g., Reflection vs. Suggestion) based on context.
Semantic and Lexical Performance:
- ROUGE-2: Showed a ~9x increase in Hebrew and a massive improvement in Arabic, indicating better capture of bigram signatures and professional phrasing.
- BLEU: CARE scored 20.4 in Arabic (20x higher than baselines) and 6.0 in Hebrew (6x higher), demonstrating superior precision and fluency in professional terminology.
Stylistic Conformity (Perplexity): CARE achieved the lowest perplexity scores (3.33 in Hebrew, 2.63 in Arabic), indicating its outputs closely mimic the tone and style of trained counselors, whereas baselines often sounded generic or overly clinical.
Statistical Significance: Paired Wilcoxon signed-rank tests with Holm correction confirmed that CARE's performance advantages were highly significant ( $p \approx 0$ ) across all metrics and languages.
Robustness: Sensitivity analysis showed CARE maintained high performance even with long input histories (80–100% percentile), whereas baselines degraded.

5. Significance and Implications

Bridging the "Cultural and Tactical Gap": The study demonstrates that the failure of general LLMs in mental health is not just a language issue but a lack of alignment with specific clinical protocols. Domain-specific fine-tuning effectively closes this gap.
Human-in-the-Loop Design: CARE is explicitly designed as a decision-support tool, not an autonomous agent. It provides suggestions to human counselors, preserving human agency and accountability while reducing cognitive load and response times.
Scalability for Low-Resource Languages: The success of this approach in Hebrew and Arabic suggests a viable pathway for deploying high-quality, culturally adapted mental health AI in other under-resourced linguistic contexts.
Future Directions: The authors suggest integrating Retrieval-Augmented Generation (RAG) for organizational guidelines and Reinforcement Learning from Human Feedback (RLHF) to further refine alignment.

In summary, CARE proves that full-history supervised fine-tuning on expert-validated data allows LLMs to internalize the complex, empathetic, and strategic nuances of professional crisis counseling, offering a scalable solution to support overwhelmed mental health workers.