iSelf-Discover: Self-Aspect Analysis Framework

• iSelf-Discover is a computational framework that systematically identifies and categorizes fine-grained Self aspects using a structured ontology with clear dimensions like Minimal, Narrative, Bodily, Agentive, and Social Self.
• It integrates formal ontology, detailed annotation protocols, and a range of models—from conventional classifiers to generative LLMs—to ensure precise, interpretable mapping of Self phenomena in language.
• Its applications span mental health assessments, cognitive science research, and narrative identity studies, offering scalable, empirically grounded insights into subjective experiences.

iSelf-Discover is a computational framework for systematically analyzing and identifying fine-grained aspects of Self as expressed in natural language. Drawing together advances in formal ontology, annotation, and hybrid neural-symbolic modeling, iSelf-Discover operationalizes concepts from phenomenology, psychology, and cognitive science to enable scalable, interpretable detection of Self-related phenomena across diverse text corpora. While conceptually rooted in theoretical accounts of the Self, the iSelf-Discover architecture is deeply empirical, providing a robust foundation for research on subjectivity, mental health indicators, and narrative identity in language (Caporusso et al., 17 Jul 2025).

1. Ontology of Self-Aspects

The iSelf-Discover framework is built upon a structured ontology that formalizes the Self as a multidimensional construct, reflected in discrete, linguistically tractable aspects. The top-level categories are:

• Minimal Self (MS): The pre-reflective, first-personal givenness of immediate experience (“this experience is mine”).
• Narrative Self (NS): The diachronic, autobiographical sense of identity realized through narrative and memory.
• Bodily Self (BS): The felt sense of body ownership, location, and interoception.
• Agentive Self (AS): The experience of agency and causality behind one’s own actions.
• Social Self (SS): The aspect of Self as constructed through relations to others and the attribution of social agency.

Each aspect is further articulated into elements (e.g., for BS: sense of ownership, location, interoceptive awareness) and modes (e.g., present, weak, absent). The ontology is formalized in OWL with hierarchical and property relations, facilitating both automated reasoning and annotation consistency. For example, BodilySelf has the element SenseOfOwnership, which in turn is annotated with present, weak, or absent modes. This schema enables precise and interoperable mapping between theoretical constructs and observable linguistic cues (Caporusso et al., 17 Jul 2025).

2. Gold-Standard Dataset and Annotation Protocols

Robust model development and evaluation within iSelf-Discover rely on a gold-standard, multi-label annotated corpus. Annotation instances are typically sentences or short paragraphs labeled for relevant Self-aspects, elements, and modes according to detailed guidelines. The annotation protocol involves:

• Multi-label tagging: An instance may simultaneously express multiple Self-aspects and elements, each with a mode (e.g., “I remember feeling out of my body” → BS: present, NS: present).
• Tiered definitions and examples: Annotators follow formal definitions for each aspect and element, with positive/negative examples and decision rules for ambiguous cases.
• Reliability checks: Three annotators label each instance, with calibration meetings and majority-vote or adjudication for disagreements. Initial pilot studies report high Cohen’s κ (0.75–0.81) across aspects.

Corpus composition targets heterogeneity: phenomenological interview transcripts, clinical interview material (e.g., for cognitive decline), autobiographical writings, and social media posts. The projected scale is ~10,000 instances per language, supporting cross-lingual benchmarks and broad domain coverage. This dataset is foundational for reproducible, interpretable modeling (Caporusso et al., 17 Jul 2025).

3. Model Architectures and Training Paradigms

iSelf-Discover supports a modular suite of model architectures tailored to address the analytic requirements of Self-aspect identification:

• Conventional Discriminative Models: Classical classifiers (linear SVM, random forests, MLPs, BiLSTM) operating on feature representations such as TF-IDF, linguistic categories (LIWC), and pretrained embeddings. These models are trained with multi-label binary cross-entropy losses and can be interpreted via coefficients or feature importances.
• Generative LLMs: Leveraging pretrained models (e.g., GPT-4, T5), iSelf-Discover applies prompt-tuning or fine-tunes classification heads for multi-label output. Few-shot in-context learning enables high accuracy without labeled pretraining for every target domain.
• Embedding-Based Retrieval Approaches: Sentence-level embeddings (e.g., SBERT, Universal Sentence Encoder) enable non-parametric labeling via nearest-neighbor retrieval from the annotated corpus, with label aggregation via weighted voting.
• Mixture-of-Experts (MoE) Hybrid: Routing mechanisms dispatch instances to expert classifiers (e.g., aspect-specialized neural or symbolic models), balancing between interpretability and flexibility; a router entropy penalty encourages sparse activation.

All approaches are validated via precision, recall, F₁-score (macro/micro), and efficiency metrics (latency, FLOPS, parameter count). The architecture is explicitly designed to accommodate trade-offs between accuracy, interpretability, and computational demands, depending on the use case (Caporusso et al., 17 Jul 2025).

4. Evaluation: Criteria, Metrics, and Statistical Rigor

Evaluation in iSelf-Discover spans four principal dimensions:

1. Interpretability: The extent to which model predictions are traceable to features, exemplars, or clearly expressed model weights.
2. Ground-Truth Adherence: The tight linkage of predictions to annotated data facilitates trust and external validation, minimizing reliance on opaque pretraining artifacts.
3. Accuracy: Measured via per-label precision, recall, F₁-score, and overall label recovery on held-out sets. Macroscopic evaluation takes into account micro- and macro-averaging strategies to account for label imbalance.
4. Computational Efficiency: Assessed through latency per query, memory and compute requirements, and parameter footprint.

Statistical significance is established using the Friedman test with Nemenyi post-hoc comparisons for multi-classifier benchmarking, and the Wilcoxon signed-rank test for paired model comparisons (Caporusso et al., 17 Jul 2025).

Table: Illustration of model performance (Social Self, English pilot)

Model	Features	F₁	Latency	Interpretability
SVM	LIWC	.83	5 ms	High (coefficients)
GPT-4 few-shot	Prompting	.79	500 ms	Low
SBERT retrieval	k=5 voting	.80	10 ms	Medium (examples)

5. Case Studies and Applications

iSelf-Discover has been applied in multiple research domains, illustrating its breadth and capacity to deliver interpretable findings:

• Mental Health / Cognitive Decline: Aspect prevalence and coherence (e.g., Narrative Self) are compared across control and patient groups. Preliminary findings (illustrative) show decreased NS coherence in Alzheimer’s, reduced AS for future-oriented statements, heightened MS pronoun usage in depression, and diminished SS references. These differences provide early digital markers for clinical assessment (p<.01 in preliminary analyses).
• Empirical Phenomenology / Dissolution Experiences: In analyses of micro-phenomenological interviews, iSelf-Discover enables phase-wise tracking of Self-aspects during experiences such as ego-dissolution. Timeline plots reveal drops in Bodily Self ownership and agentive fluctuations at experiential peaks, with gradual restoration of the Narrative Self during recovery (Caporusso et al., 17 Jul 2025).

This supports hypothesis generation in cognitive science and informs therapeutic interventions targeting the structure of self-experience.

6. Synthesis and Limitations

iSelf-Discover combines:

• A rigorously instantiated ontology of Self,
• A gold-standard, annotated, multilingual corpus,
• Interoperable discriminative, generative, and retrieval-based modeling strategies,
• A multidimensional, statistically robust evaluation toolkit.

The result is a computational engine suitable for experimental psychology, digital mental health monitoring, and phenomenological research. Key limitations include dependence on the completeness and consistency of ontology design, annotation quality, and linguistic expressiveness of Self-aspects. This suggests future work should focus on cross-lingual expansion, richer multimodal datasets, and integration of physiological or behavioral signals as complementary markers.

iSelf-Discover, upon full realization, constitutes an ethical, empirically grounded, scientifically interoperable platform for mapping the manifold dimensions of Self as emergent in language, supporting both research and clinical translation (Caporusso et al., 17 Jul 2025).