PsychoLexDialogue: Persian Therapy Dataset
- PsychoLexDialogue is a multi-turn therapy dialogue dataset in Persian that models structured therapeutic reasoning using hybrid simulated sessions and detailed client profiles.
- It uses a staged narrative based on Person-Centered Therapy to drive coherent, culturally grounded interactions with integrated long-term memory modules.
- The dataset supports evaluation of empathy, content coherence, and personalization, highlighting improvements in therapeutic alignment across various system configurations.
PsychoLexDialogue is a multi-turn, therapy-like dialogue dataset in Persian introduced in "PsycholexTherapy: Simulating Reasoning in Psychotherapy with Small LLMs in Persian" (Abbasi et al., 4 Oct 2025). It was built as “hybrid simulated sessions” to test whether structured therapeutic reasoning and long-term memory can sustain coherent, empathetic, and culturally grounded psychotherapy simulation across multiple turns. In the broader PsychoLex ecosystem, it functions as the dialogue-layer counterpart to PsychoLexEval and PsychoLexQuery, extending psychology-specific knowledge resources toward interactive, multi-turn evaluation (Abbasi et al., 2024).
1. Origin within the PsychoLex ecosystem
PsychoLexDialogue emerged from a three-stage development process in which the authors first assessed small LLMs’ psychological knowledge with PsychoLexEval, then designed the PsychoLexTherapy framework, and finally constructed two evaluation datasets: PsychoLexQuery for real Persian single-turn questions and PsychoLexDialogue for multi-turn hybrid simulated sessions (Abbasi et al., 4 Oct 2025). Within that design, PsychoLexDialogue is the primary benchmark for evaluating whether a system can preserve therapeutic coherence and personalization beyond single-turn answering.
Its intended functions are explicitly threefold. First, it targets psychological depth by modeling therapeutic trajectories rather than isolated question–answer pairs. Second, it targets cultural authenticity by reflecting Persian language, Iranian social context, and locally salient family, work, and relational dynamics. Third, it targets evaluative utility by providing a controlled multi-turn environment in which naive prompting, prompting with memory, and structured therapeutic reasoning can be compared under the same task conditions (Abbasi et al., 4 Oct 2025).
The dataset also occupies a natural position within the broader PsychoLex framework. PsychoLex itself provides a psychology-specific ecosystem consisting of PsychoLexQA, PsychoLexEval, and PsychoLexLLaMA, designed to enhance LLM proficiency in psychological tasks in Persian and English (Abbasi et al., 2024). PsychoLexDialogue extends that ecosystem from instructional and multiple-choice resources to dialogue-centered testing of long-horizon interaction.
2. Dataset construction and corpus structure
PsychoLexDialogue is built directly on top of PsychoLexQuery, which contains ~4,000 real questions from Persian mental-health forums and consultation platforms, carefully anonymized (Abbasi et al., 4 Oct 2025). These source questions cover common psychotherapy and counseling topics such as relationship issues, anxiety, self-esteem, depressive and hopeless affect, family conflict, academic stress, and workplace stress.
For each PsychoLexQuery item, the authors construct a structured client profile. The profile includes emotional themes, core psychological issues, past experiences, cognitive/behavioral patterns, therapeutic goals / desired outcomes, and contextual factors. The paper provides an explicit example in which a question about assertiveness versus stubbornness at work is mapped to a JSON-like client profile with fields such as "emotional_themes", "key_psychological_issues", "past_experiences", "patterns_and_behaviors", "desired_outcome", and "contextual_factors" (Abbasi et al., 4 Oct 2025). These profiles serve as the backbone for dialogue construction and as the substrate for later memory-based evaluation.
The emotional landscape of the source material is also quantified. Table 1 reports counts such as Frustration: 2,914, Sadness: 2,124, and Anxiety: 1,306, and the dataset is described as reflecting cases in which negative affect dominates and each case typically has 3–5 concurrent emotions (Abbasi et al., 4 Oct 2025). PsychoLexDialogue inherits that emotional complexity rather than flattening cases into single-label affective states.
Dialogue generation proceeds through a staged hybrid pipeline. Each profile is first expanded into a five-stage narrative aligned with Person-Centered Therapy (PCT): Establishing trust and rapport; Empathic listening and emotional reflection; Encouraging free exploration of thoughts and feelings; Supporting personal growth and cognitive restructuring; and Summarizing insights and planning future steps (Abbasi et al., 4 Oct 2025). The authors initially create a scripted backbone to guarantee thematic consistency and coverage of core issues and goals. They then refine it with a hybrid approach in which an LLM therapist agent and an LLM client agent play out the conversation while remaining grounded in the scripted structure and the client profile. The pipeline is summarized as Web Crawling → Client Profile → Structured Script → Hybrid Approach (LLM Therapist + LLM Client) → Generated Dialogue → Data Filtering (Abbasi et al., 4 Oct 2025).
The resulting dataset contains 3,400 dialogues, spans 16 thematic categories, and has an average of 10–14 turns per dialogue (Abbasi et al., 4 Oct 2025). The most common thematic categories are relationship issues, anxiety, and self-esteem problems. Dialogues are role-structured with alternating client and therapist turns, and each dialogue typically contains 3–5 co-occurring emotions rather than a single affective state (Abbasi et al., 4 Oct 2025).
3. Representational design and therapeutic framing
PsychoLexDialogue is not merely a sequence of turns; it is a dataset whose multi-turn structure is anchored in therapeutic process design. Each conversation is implicitly staged by the five-step PCT-oriented narrative, and each session is grounded in a structured profile that encodes emotional themes, psychological issues, experiences, behavioral patterns, desired outcomes, and contextual factors (Abbasi et al., 4 Oct 2025). This gives the corpus a profile-to-session linkage that supports both personalization and evaluation of long-term memory.
At the representational level, the paper states that no explicit per-turn labels such as emotion annotations or therapy-stage labels per turn are described; instead, structure is primarily embedded through the narrative design and client profiles (Abbasi et al., 4 Oct 2025). This is consequential. PsychoLexDialogue is therefore less a turn-labeled corpus in the style of dialogue act datasets and more a controlled evaluation environment in which macro-structure, client state, and therapeutic direction are built into the generation process.
The therapeutic content is defined by both the dataset narrative and the reasoning framework used to test on it. The macro-narrative is PCT-oriented, emphasizing empathy, reflective listening, non-directive exploration, and gradual movement toward insight and future planning. At the same time, the associated PsychoLexTherapy framework implements three structured reasoning paths—CBT, RT, and PCT—that are exercised on PsychoLexDialogue during evaluation. The CBT reasoning path extracts automatic thoughts, infers emotional consequences and behavioral tendencies, generates balanced alternative thoughts, derives adaptive behaviors, and integrates them into a final response. The RT reasoning path identifies core needs and wants, analyzes current behaviors, evaluates their consequences, plans alternative behaviors, and integrates them into a final response. The PCT reasoning path performs empathic reflection and emotional understanding, exploratory questioning for self-awareness, and integration into a final supportive response, usually ending with an open question (Abbasi et al., 4 Oct 2025). As a result, PsychoLexDialogue implicitly encodes therapy-consistent reasoning trajectories rather than only surface empathy.
The dataset is also explicitly culturally grounded. Dialogues reflect Persian linguistic norms, Iranian social context, and recurring local patterns such as strong emphasis on relationship and family dynamics, work and academic stress, social expectations, and gendered roles (Abbasi et al., 4 Oct 2025). The corpus is therefore not presented as a culturally neutral psychotherapy dataset; its internal validity is tied to Persian and Iranian context.
4. Evaluation role in PsychoLexTherapy
PsychoLexDialogue is the benchmark used to evaluate multi-turn variants of PsychoLexTherapy, especially the interaction between structured reasoning and long-term memory (Abbasi et al., 4 Oct 2025). The paper compares four configurations:
| Configuration | Core components | Average |
|---|---|---|
| Multi-Turn w/o Memory | Simple multi-turn prompt | 5.43 |
| Multi-Turn + Memory | Multi-turn prompt with long-term memory | 6.34 |
| PsychoLexTherapy w/o Memory | Therapy approach selector + structured reasoning paths | 6.99 |
| PsychoLexTherapy + Memory | Full system with reasoning paths and MemoBase | 8.14 |
The baseline without memory uses naive history concatenation, meaning the entire raw transcript is appended to the prompt at each turn. The paper explicitly states that “naïve history concatenation caused incoherence and information loss” on PsychoLexDialogue (Abbasi et al., 4 Oct 2025). The contrast condition adds MemoBase, a long-term memory module that builds a structured user profile, buffers new information, and preserves basic information, preferences, personalization settings, and recent events. This structured memory is intended to prevent earlier disclosures from being lost as conversations lengthen.
The quantitative results are reported on a ten-point Likert scale via LLM-as-a-judge. For Multi-Turn w/o Memory, the scores include Empathy: 7.8, Cultural Fit: 6.8, Therapeutic Alignment: 7.2, Content Coherence: 6.4, Emotional Coherence: 3.1, and Personalization: 3.3, yielding Average: 5.43. For Multi-Turn + Memory, emotional coherence and personalization rise to 7.6 and 7.4, and the average rises to 6.34. For PsychoLexTherapy w/o Memory, structured therapeutic reasoning improves Empathy: 8.6, Cultural Fit: 8.6, Therapeutic Alignment: 8.2, Content Accuracy: 8.2, and Average: 6.99. The full PsychoLexTherapy + Memory system reaches Empathy: 9.2, Cultural Fit: 8.8, Therapeutic Alignment: 8.6, Content Accuracy: 8.4, Adaptability: 8.8, Content Coherence: 8.6, Emotional Coherence: 8.6, Personalization: 7.3, and Average: 8.14 (Abbasi et al., 4 Oct 2025).
These results are used to support two specific conclusions. First, long-term memory substantially improves emotional coherence, personalization, and continuity of content across turns. Second, structured therapeutic reasoning improves therapeutic alignment, content accuracy, and cultural fit even without memory. Their combination yields the highest ratings across empathy, coherence, cultural fit, and personalization (Abbasi et al., 4 Oct 2025).
5. Relation to broader psychologically structured dialogue research
Although PsychoLexDialogue is a dataset name in (Abbasi et al., 4 Oct 2025), later work also treats it as a broader design agenda for psychologically grounded dialogue via lexicons, ontologies, and structured state models. In "PsyProbe: Proactive and Interpretable Dialogue through User State Modeling for Exploratory Counseling," the system is described as directly relevant to a “PsychoLexDialogue” agenda because it operationalizes a psychologically grounded dialogue ontology through PPPPPI, cognitive errors, MI process labels, and MI behavioral codes (Park et al., 27 Jan 2026). That work frames PPPPPI slots as schema types, cognitive error categories as lexical-psychological features, and MI codes as behavioral labels, thereby offering a more explicit state ontology than PsychoLexDialogue itself.
A different but complementary line appears in "Psy-Copilot: Visual Chain of Thought for Counseling," where the authors describe Psy‑COT as the sort of resource that a “PsychoLexDialogue” system might want: a counseling dialogue resource with explanatory chain-of-thought content, a structured representation that maps raw dialogue language to explicit strategy and event nodes, and a retrieval backbone that makes reasoning inspectable (Chen et al., 5 Mar 2025). In that account, a PsychoLexDialogue-like resource would not stop at turns and themes; it would connect utterances to a lexicon of strategy and interpretation phrases, causal links, and graph-structured reasoning.
"Structured Dialogue System for Mental Health: An LLM Chatbot Leveraging the PM+ Guidelines" provides another point of comparison. SuDoSys does not define a lexicon-centered resource, but it does show how psychologically structured dialogue can be controlled through stage-aware instruction generation, a topic database, and a stage controller grounded in PM+’s seven-step framework for problem management (Chen et al., 2024). Relative to PsychoLexDialogue, SuDoSys is process-structured rather than psycholexically structured.
"Script-Strategy Aligned Generation" offers a further complementary perspective by formalizing therapeutic topics, therapeutic questions, and MI strategies as explicit alignment targets for LLMs (Sun et al., 2024). That work suggests a resource design in which dialogue data are linked to a strategy layer and a topic layer, an arrangement that is congenial to later “PsychoLexDialogue” formulations emphasizing lexicons of therapeutic acts and psychologically meaningful dialogue functions.
At the diagnostic end of the spectrum, "From Medical Records to Diagnostic Dialogues" introduces PsyCoTalk, described as the first large-scale dialogue dataset supporting comorbidity, with 3,000 multi-turn diagnostic dialogues generated through a hierarchical state machine and context tree (Wan et al., 29 Oct 2025). PsyCoTalk is not PsychoLexDialogue, but it shows how a clinically grounded, lexically rich diagnostic dialogue resource can be built around explicit symptom states and comorbidity structure.
Taken together, these related systems suggest that PsychoLexDialogue has become more than a dataset name. A plausible implication is that it now designates a family of research problems centered on psycholexically grounded dialogue, structured therapeutic reasoning, and inspectable multi-turn state representations.
6. Limitations and future directions
The most explicit limitations come from the originating PsychoLexDialogue paper itself. The dataset uses hybrid construction with LLM assistance, and the authors note that it may contain stylistic artifacts and synthetic biases diverging from real client speech or therapist styles (Abbasi et al., 4 Oct 2025). Even with human filtering, it is not fully naturalistic clinical data. This places PsychoLexDialogue in the category of controlled evaluation resource rather than direct clinical transcript substitute.
Its cultural specificity is equally explicit. All three PsychoLex datasets, including PsychoLexDialogue, are Persian and Iranian-centric (Abbasi et al., 4 Oct 2025). This yields strong internal coherence and local relevance, but limited generalizability to other languages, therapeutic cultures, and institutional settings. A direct implication is that cross-lingual transfer cannot be assumed.
Therapeutic coverage is also bounded. The associated framework models only three approaches—CBT, RT, and PCT—and does not cover ACT, EFT, schema therapy, psychodynamic therapy, or other modalities named elsewhere in the broader literature (Abbasi et al., 4 Oct 2025). The long-term memory module is similarly limited: the paper states that MemoBase profiles do not implement forgetting / decay, uncertainty modeling, or conflict resolution when later user messages contradict earlier disclosures (Abbasi et al., 4 Oct 2025).
Evaluation remains another constraint. For PsychoLexDialogue, the multi-turn side relies on LLM-as-a-judge; there is no separate human evaluation described for PsychoLexDialogue (Abbasi et al., 4 Oct 2025). The dataset is therefore central for reproducible benchmarking, but not yet tied to human-rated multi-turn clinical utility or to longitudinal patient outcomes. The paper also states that PsychoLexDialogue is an evaluation dataset for simulated therapy, with no clinical outcome validation using scales such as PHQ‑9, GAD‑7, or alliance measures (Abbasi et al., 4 Oct 2025).
Future directions are correspondingly clear in the source description. These include extending to more therapy modalities, refining memory mechanisms with better forgetting and conflict handling, creating more naturalistic multi-turn data, and porting the pipeline to other underrepresented languages and cultural contexts (Abbasi et al., 4 Oct 2025). In the broader literature, adjacent work suggests that such extensions may plausibly involve explicit state ontologies, graph-based strategy layers, or stage-aware controllers (Park et al., 27 Jan 2026, Chen et al., 5 Mar 2025, Chen et al., 2024).
PsychoLexDialogue thus occupies a distinctive position in current research: it is a culturally grounded Persian multi-turn psychotherapy dataset, a benchmark for testing structured reasoning and long-term memory, and a point of departure for a wider research program on psychologically structured dialogue resources (Abbasi et al., 4 Oct 2025).