Linguistic, Coherence, and Semantic Persona Validator Research Report

Full report

On this page

The proliferation of large-scale fictional-character populations generated by advanced computational models frequently yields structural correctness without semantic integrity. While generated outputs may perfectly adhere to predefined JSON schema constraints, they often exhibit profound grammatical degradation, narrative incoherence, chronological impossibilities, and unnatural dialogue patterns. The analysis indicates that existing syntactic validators fail to capture the complex, multi-dimensional nature of character coherence. A large fictional-character population recently exhibited repeated grammar and naturalness failures, including elementary linguistic errors (e.g., “They values,” “They is,” “a insurance”), extreme dialogue redundancy (e.g., “I am Aaliyah James Smith. My name is Aaliyah James Smith.”), formulaic biographies, contradictory relationship matrices, and a distinct lack of behavioral variation. Structural schema validation proved entirely blind to these semantic failures. To address systemic failures across generative pipelines, a multi-layered, language-independent semantic persona validator must be designed. This architecture synthesizes advancements in narrative coherence evaluation1, dependency parsing3, temporal constraint satisfaction via Allen’s Interval Algebra5, and knowledge graph verification8. The resulting semantic persona validator evaluates generated fictional characters beyond structural JSON correctness, providing robust mechanisms for English and subsequent multilingual expansion, alongside deterministic human-review escalation pathways.

semantic-persona-validator-specification.md

The Semantic Persona Validator operates as an independent, post-generation verification pipeline, entirely decoupled from the underlying generative model and structural JSON schema validation. The primary objective is to evaluate the semantic, chronological, and narrative properties of a generated persona. The architecture requires that schema validity and semantic validity exist as entirely separate validation stages. The validator assesses data across multiple intersecting fields, specifically Biography, Dossier, Relationships, Memory, Occupation, Residence, Languages, Goals, Dialogue, Emotional behavior, Scene-state facts, Provider activation, and Image prompts. The system evaluates linguistic constraints using the Universal Dependencies (UD) framework3, solves topological chronologies using a constraint satisfaction problem (CSP) solver grounded in Allen's Interval Algebra6, and extracts entities to ensure bidirectional relational symmetry and prevent unsupported world-fact hallucinations8. Furthermore, it computes lexical diversity, syntactic repetition entropy, and conversational entrainment to flag unnatural dialogue pragmatics11. The activation criteria for a persona are absolute. A character may pass semantic validation only when no blocker remains, no grammar error remains, identity is completely preserved across all fields, no hard cross-field contradiction remains, and relationship chronology is proven coherent. Furthermore, unsupported world facts must be removed or properly labeled as fictional inventions, dialogue must not overuse full-name introductions, naturalness warnings must remain below a documented threshold, and all required human-review findings must be fully resolved.

validation-finding.schema.json

The foundational data structure for the validator relies on a highly specific, machine-readable error model. Every identified defect is mapped to a localized JSON path with a defined severity, supporting human-in-the-loop review mechanisms and automated remediation pathways. The schema ensures that every error possesses a stable code and path.

JSON { "$schema": "http://json-schema.org/draft-07/schema\#", "title": "Validation Finding", "description": "A single semantic, grammatical, or coherence defect detected in a persona.", "type": "object", "properties": { "code": { "type": "string", "description": "Standardized error code from the error-code-catalog." }, "severity": { "type": "string", "enum": \["BLOCKER", "ERROR", "WARNING", "ADVISORY", "REQUIRES\_HUMAN\_REVIEW", "CANNOT\_DETERMINE"\], "description": "Impact level of the finding on persona activation." }, "field": { "type": "string", "description": "The high-level category of the persona structure containing the defect." }, "path": { "type": "string", "description": "JSONPath expression pointing to the exact location of the defect." }, "excerpt": { "type": "string", "description": "The specific substring or data point triggering the violation." }, "explanation": { "type": "string", "description": "Human-readable justification for the finding, referencing linguistic or narrative rules." }, "suggestedRepair": { "type": "string", "description": "Machine-generated correction or remediation step." }, "confidence": { "type": "number", "minimum": 0.0, "maximum": 1.0, "description": "Validator's confidence score in the finding." }, "blocksActivation": { "type": "boolean", "description": "Boolean flag indicating if this specific finding prevents pipeline progression." } }, "required": \["code", "severity", "field", "path", "excerpt", "explanation", "confidence", "blocksActivation"\] }

validation-report.schema.json

The aggregation of all findings for a single persona or an entire character population is encapsulated in the validation report. This structure provides the summary statistics required for population-level findings, enabling machine-readable errors and machine-readable warnings to be processed by downstream observability tools.

JSON { "$schema": "http://json-schema.org/draft-07/schema\#", "title": "Persona Validation Report", "type": "object", "properties": { "personaId": { "type": "string" }, "timestamp": { "type": "string", "format": "date-time" }, "status": { "type": "string", "enum": \["PASSED", "BLOCKED", "PENDING\_REVIEW"\] }, "metrics": { "type": "object", "properties": { "totalFindings": { "type": "integer" }, "blockers": { "type": "integer" }, "errors": { "type": "integer" }, "warnings": { "type": "integer" }, "humanReviewRequired": { "type": "integer" }, "naturalnessScore": { "type": "number" } } }, "findings": { "type": "array", "items": { "$ref": "validation-finding.schema.json" } } }, "required": \["personaId", "timestamp", "status", "metrics", "findings"\] }

error-code-catalog.json

The error code catalog formalizes the taxonomy of potential failures across the system. Severity levels dictate the automated routing of the persona, acting as strict gates for the activation criteria. The severity levels are defined as follows:

• Blocker: Represents fundamental structural or semantic collapse. Examples include identity substitution, severe cross-field contradictions, and grammar failures. A blocker explicitly prevents activation.
• Error: Significant inconsistencies that degrade narrative quality but may be automatically repaired via the suggestedRepair parameter. Errors block activation until resolved or repaired.
• Warning: Improbable but not theoretically impossible traits, such as highly unusual career trajectories. Naturalness warnings that do not exceed a designated threshold permit activation.
• Advisory: Stylistic recommendations for dialogue or phrasing that do not impede the baseline functionality of the persona.
• Requires Human Review: Ambiguous relationship behaviors, cultural plausibility questions, dialect representations, and potential stereotyping that deterministic systems cannot reliably judge safely.
• Cannot Determine: A fallback severity utilized when text parsing fails or when multiple rule sets provide contradictory evaluations, requiring manual intervention.

grammar-rule-catalog.json

Traditional syntactic validation relies on English-specific regular expressions, which fail spectacularly when applied multilingually or against diverse dialects. To ensure language-independent validation, the Semantic Persona Validator leverages the Universal Dependencies (UD) parsing framework3. UD provides cross-linguistically consistent treebank annotations, mapping words to universal part-of-speech (POS) tags and extracting morphological features such as Number, Person, Tense, and Case13. Grammar errors strictly block activation. By parsing the generated text into a dependency graph, linguistic rules are evaluated as graph constraints. Subject-verb agreement is detected by locating the root node (typically a verb) and traversing the nsubj (nominal subject) edge. The validator compares the morphological features of the subject node and the verb node. For instance, the phrase “They is” yields an nsubj mapping of "They" (Number=Plur, Person=3) connected to the root "is" (Number=Sing, Person=3). This morphological Number conflict triggers a SUBJECT\_VERB\_AGREEMENT error. Similarly, pronoun-verb agreement utilizes identical edge traversal logic to detect mismatches. Number agreement across modifiers and nouns operates by evaluating the amod and nummod dependencies to ensure plural markers align correctly within the language's specific constraints. Indefinite-article agreement is uniquely complex because it relies on phonology. The validator locates det (determiner) edges. If the determiner lemma represents an indefinite article, the system evaluates the phonological onset of the head noun or the intermediate adjectival modifier. A phrase like "a insurance" flags a mismatch between the consonant-expecting "a" and the vowel-onset "insurance", triggering an INDEFINITE\_ARTICLE\_AGREEMENT error. Furthermore, missing determiners are flagged when a singular countable noun lacks a required det dependency in languages where determiners are syntactically obligatory. Tense consistency is monitored by extracting the Tense and Aspect morphological features of all verbs within contiguous narrative blocks14. Unjustified temporal shifts that lack appropriate discourse markers trigger a TENSE\_CONSISTENCY warning. Malformed coordination is identified through UD's specific handling of conjunctions. In dependency syntax, coordination designates the first conjunct as the head, with subsequent conjuncts attached via conj and conjunctions via cc4. Broken coordination patterns arise when conj nodes possess fundamentally incompatible POS tags or case assignments that violate the specific language's coordination syntax. Structural sentence errors are identified via graph topologies. Sentence fragments are detected when the dependency graph entirely lacks a valid root predicate. Run-on sentences are identified when multiple finite verbs exist within a single boundary without proper coordination (cc or conj) or subordination marks. Duplicate clauses are flagged via structural isomorphism, where two parallel subtrees contain identical semantic lemmas and syntactic relations. Pronoun-level grammar rules are strictly enforced. Incorrect possessives are flagged when nmod:poss (possessive nominal modifier) tags are misaligned with their subjects. Repeated subjects manifest as redundant nsubj attachments to a single uncoordinated verb. Inconsistent person occurs when a text shifts from first-person to third-person pronouns referring to the same entity without narrative bridging. Incorrect reflexives (e.g., using "themselves" as a primary nominative subject) violate syntactic binding theories and are flagged by checking reflexive features against their structural position. Finally, pronoun-case errors (e.g., "Her went to the store") are detected when the nominative Case feature is required but an accusative or genitive form is supplied by the generative model.

Identity Preservation Rules

Identity preservation requires maintaining a mathematically consistent state across the persona's JSON fields. When a fictional-character population is generated, large language models frequently suffer from attention-mechanism drift, blending identities, hallucinating new names mid-biography, or reversing cultural naming conventions. Identity substitution strictly blocks activation. The validator matches the generated name against the initial prompt request. To account for varying native scripts, it utilizes phonetic and structural transliteration mapping. If the requested name differs from the resolved name, or if the requested age differs from the resolved age, the pipeline halts. The system checks for instances where preferred names and legal names are mixed incorrectly, preventing generative outputs from arbitrarily alternating between a chosen name and a discarded legal name in ways that violate the character's specified preferences. Cultural name preservation is highly prioritized. For specific linguistic backgrounds, the validator verifies that family-name-first names are not reversed by western-centric generation algorithms. The disappearance of diacritics (e.g., "Müller" degrading to "Muller" across fields) triggers a formatting error. Native script changes—such as a Japanese character's biography abruptly switching from Kanji to Romaji without justification—are strictly monitored. The validator traces coreference chains to prevent identity blending, wherein the traits of multiple different characters are merged into one file. If the biography refers to a different name, or if the dialogue uses a completely different identity (introducing themselves as an unrelated individual), a blocker is issued. Furthermore, relationship records that refer to the wrong person—linking an emotional state to a non-existent ID—are flagged. Finally, pronouns must not change arbitrarily across fields. Pronoun usage is tracked across the Biography, Dossier, and Relationships fields; if a character is strictly defined with a specific pronoun set, but alternating gendered pronouns appear without narrative justification, the persona fails validation.

chronology-rule-catalog.json

Extracting timelines from natural language and unstructured JSON fields to ensure chronological coherence requires transforming narrative events into topological sort problems. The Semantic Persona Validator leverages Allen's Interval Algebra6, which defines 13 mutually exclusive relations between time intervals (e.g., before, meets, overlaps, starts, during, finishes, and their inverses, plus equals). By mapping life events as Interval Event Units (IEUs)5, the validator constructs a temporal graph. An inconsistency creates an unsatisfiable cycle in the graph, which constraint satisfaction algorithms detect7.

The validator scrutinizes migration events to ensure they do not occur in an impossible order; geographic residency intervals separated by vast distances cannot hold an equals or overlaps relation, preventing simultaneous incompatible locations. It also verifies that the character's current age is logically consistent with any explicitly dated biography facts. Temporal linguistic parsing prevents a future event from being described as completed17. Furthermore, topological graph checking ensures that a dead person does not appear in current routines or daily schedules without explicit, provider-activated explanations (e.g., ghost/sci-fi attributes).

Occupation and life-course plausibility

Evaluating occupation and life-course plausibility requires distinguishing between logically impossible sequences and highly unique, non-traditional life paths. The validator explicitly does not automatically reject unusual careers; instead, it requires causal support rather than relying on demographic stereotypes. Rare identities and unusual trajectories are not treated as defects, preserving narrative diversity. The system detects an advanced career without training or experience by mapping occupational titles to required prerequisite nodes in the knowledge graph. A seniority level that is entirely inconsistent with the character's biological age is flagged for temporal review. Instances of multiple full-time occupations are flagged unless an explicit explanation (such as severe financial distress or extreme overwork) bridges the logical gap. Job responsibilities must remain consistent across fields; if a character's dossier lists them as a retail worker, but their dialogue references conducting open-heart surgery, a cross-field contradiction is generated. Similarly, an occupation must not be incompatible with established geographic locations or daily schedules. Skill claims that lack a corresponding acquisition history in the biography are marked as unsupported. Finally, an economic situation that is wildly inconsistent with both occupation and housing—without a contextual explanation such as inheritance or debt—triggers a plausibility warning.

relationship-rule-catalog.json

Relationships within a persona population must be validated as a directed, bidirectional Knowledge Graph (KG)8. Evaluating a single character in isolation is fundamentally insufficient; relationship consistency requires edge-checking between Character A and Character B to ensure symmetry. The validator detects different durations and different relationship labels between two connected entities. While asymmetrical emotional states (e.g., unrequited love) are valid, fundamentally different factual labels (e.g., A claims B is a sibling, B claims A is a stranger) trigger hard errors. One-sided shared events and one-sided shared memories are strictly prohibited; if an event is explicitly marked as a "shared memory" for one party, the corresponding memory edge must exist in the second party's records. Behavioral relationship metrics are monitored for internal logic. Contradictory trust levels and conflicting contact frequencies are flagged (e.g., a relationship described simultaneously as "current and deeply estranged" without any bridging context, or "complete lack of trust" paired with "shares all deepest secrets"). Impossible chronologies within relationships—such as a friendship lasting forty years between two thirty-year-olds—are caught by the Interval Algebra engine. The system prevents a person from being described as both living and dead simultaneously. Finally, the validator monitors for knowledge exceeding plausible access; a character cannot possess detailed knowledge of another character's private, non-shared events without an explicit information-transfer mechanism.

dialogue-naturalness-rules.json

Generated characters frequently suffer from a severe lack of dialogue naturalness, manifesting as robotic repetition, extreme over-explanation, and identical behavioral responses12. To quantify naturalness, the validator avoids simple keyword matching and instead utilizes linguistic entrainment, lexical diversity, and discourse analysis metrics to ensure naturalness has measurable indicators11. The system tracks excessive full-name self-introductions, flagging instances where a character states their full name on every single dialogue line, as well as tracking the repeated phrase “My name is.” Biography recitation is penalized; language models often force characters to unnaturally over-explain their backstory in response to trivial prompts. By computing the semantic embedding similarity between the Biography field and Dialogue responses, the validator identifies when dialogue is merely a paraphrased dossier. Syntactic diversity is strictly enforced. If every line begins with the pronoun “I,” or if every line uses the exact same underlying syntax tree (e.g., Noun-Verb-Adjective), the syntactic repetition entropy score drops, triggering a warning. Identical responses to different intents are blocked. Excessive occupation references—where a character cannot answer a greeting without mentioning their job—are flagged as unnatural motifs. Human conversations feature fluid transitions, lexical entrainment, and spontaneous disfluencies18. The validator flags unnaturally formal ordinary dialogue and a distinct lack of contractions where expected by the character's dialect. Emotional expression is monitored to prevent repeated emotional labels and identical uncertainty language. Self-BLEU algorithms ensure that identical disagreement language, identical silence behavior, and identical interruption behavior are flagged, forcing the generative model to produce varied, context-aware pragmatic responses.

Cross-field contradictions

Generative models frequently fail to maintain state across disparate data structures. The Semantic Persona Validator executes a matrix comparison to detect contradictions among all major fields: Biography, Dossier, Relationships, Memory, Occupation, Residence, Languages, Goals, Dialogue, Emotional behavior, Scene-state facts, Provider activation, and the Image prompt. For example, if the Residence field dictates an impoverished rural setting, but the Image prompt requests a futuristic cybernetic cityscape, the contradiction is immediately logged. If the Languages field lists strictly monolingual Spanish, but the Dialogue field generates intricate colloquial English without translation artifacts, the discrepancy is flagged. Cross-field tracking utilizes the FACTTRACK validity interval logic to ensure that states changing over time do not trigger false contradictions, distinguishing legitimate chronological progression from actual factual conflicts17.

unsupported-world-fact-rules.md

Hallucinations in generative knowledge extraction represent a significant threat to narrative integrity10. When a persona claims participation in a historical event or affiliation with an external agency, the validator must distinguish between intentional fictional world-building and unsupported large language model hallucinations. Unsupported world facts must be removed or properly labeled to clear the activation criteria. The system utilizes an internal Knowledge Graph acting as the "World Bible," alongside external Retrieval-Augmented Generation (RAG) paradigms for fact validation8. The validator detects when a persona invents a government agency, invents a company, or invents a war or historical event that does not exist in the canonical database. Furthermore, it strictly blocks personas that invent family members not established elsewhere in the graph or invent a player relationship without explicit provider activation. Epistemic modality is deeply parsed. The validator detects when a persona invents current world-state facts or claims access or knowledge that has not been chronologically established. Furthermore, natural language inference (NLI) models monitor how characters process information, flagging instances where the generated text converts rumors into objective facts, or treats unverified allegations as confirmed historical truths.

Boilerplate and duplication

To maintain a vibrant, diverse population of characters, the system aggressively hunts for boilerplate generation. Generative models can become trapped in local minimums, resulting in widespread population-level duplication. The validator utilizes Locality-Sensitive Hashing (LSH) and MinHash algorithms to detect exact biography duplicates, normalized biography duplicates, and near-duplicate biographies across the entire character database. At the granular level, it scans for exact dialogue duplicates and normalized dialogue duplicates. Shared n-grams and shared sentence skeletons are tracked across the population to identify formulaic generation patterns. Semantic embeddings (e.g., sentence-transformers) cluster more abstract concepts. The system flags reused relationship combinations, reused goals, reused ordinary concerns, and reused metaphor families. If a disproportionate percentage of the population exhibits identical behavioral responses to specific stimuli, a population-level finding is escalated, indicating a fundamental failure in the generator's sampling temperature or prompt diversity.

multilingual-validation-plan.md

The validator is required to support English initially, followed by immediate multilingual expansion. A core completion criterion demands that multilingual rules do not assume English grammar. Attempting to apply English logic—such as evaluating gendered pronouns or indefinite articles—to disparate language families results in catastrophic validation failures. The architecture incorporates fastText-based language detection. The system seamlessly handles mixed-language biographies and code-switching by segmenting the text and applying language-specific dependency parsers sequentially. Crucially, the validator adjusts its pronoun systems based on the detected language. For languages lacking grammatical gender in the third person (e.g., Finnish, Turkish), gender agreement rules are entirely bypassed, relying instead on Named Entity Recognition (NER) coreference resolution. Similarly, for languages without articles (e.g., Russian, Japanese), the indefinite-article agreement rules are nullified. The system accounts for complex honorifics and politeness levels14. If a character utilizes extreme honorifics when speaking to a subordinate—without a narrative tag denoting sarcasm—a pragmatic mismatch is flagged. Name order conventions are strictly enforced based on cultural origins. Native script usage is monitored, and phonetic transliteration bridges comparisons between requested names and generated localized outputs. Finally, the validator includes specialized modules for evaluating dialect, creoles, sign-language representation (gloss transcription formats), and identifying translation artifacts caused by generative models internally translating from English rather than natively generating the target language.

human-review-escalation.md

Deterministic automation cannot safely evaluate every nuance of human culture, humor, and complex identity. Therefore, the validator features explicit escalation pathways where automation must defer to a human reviewer, assigning a REQUIRES\_HUMAN\_REVIEW severity. Required human-review findings must be resolved prior to activation. Escalation is triggered by assessments of cultural plausibility; if a generative model combines traits that are statistically rare but culturally significant, the system escalates rather than automatically deleting the character. Deliberate character voice can be distinguished from accidental error through this mechanism. If a character's dialogue features deliberate nonstandard grammar in dialogue (such as regional dialects or African American Vernacular English), grammatical rule failures are suppressed, and the persona is escalated to ensure the dialect representation is respectful and accurate. Humor naturalness, particularly sarcasm and dark humor, often registers as a cross-field contradiction to a deterministic system; these are flagged for human validation. Ambiguous relationship behavior and conflicting but plausible life histories (e.g., the "unreliable narrator" trope) are escalated to preserve artistic intent. Furthermore, any legal or historical claims generated by the persona, alongside complex mental-health characterizations, are mandatorily escalated. Semantic analysis actively screens for combinations of traits that align with potential stereotyping, blocking activation until a human sensitivity reviewer provides clearance.

validator-test-corpus.json

A comprehensive test corpus dictates the acceptance criteria for the validator's performance. The corpus is represented as specifications and examples, ensuring every defect possesses an expected code, expected severity, expected path, expected blocking state, and a corrected version. The minimum test corpus consists of:

• 100 valid character examples: Establishing the baseline for perfect coherence.
• 150 single-defect examples: Isolating individual grammar or logic failures.
• 100 multi-defect examples: Testing cascading failures across fields.
• 50 chronology contradictions: Validating Interval Algebra solvers.
• 50 relationship contradictions: Testing bidirectional graph symmetry.
• 50 identity-preservation failures: Ensuring pronoun and name consistency.
• 50 naturalness failures: Testing entropy and disfluency metrics.
• 50 unsupported-world-fact examples: Validating RAG and KG checks.
• 50 cross-language or Unicode cases: Testing script boundaries and honorifics.
• 50 false-positive challenge cases: Ensuring rare identities and deliberate dialects are not incorrectly penalized.

JSON { "testId": "TC-CHRONO-088", "category": "Chronology", "description": "Employment before plausible age.", "payload": { "character": { "birthDate": "2010-01-01", "occupationHistory": \[ { "title": "Senior Staff Architect", "startDate": "2018-06-01" } \] } }, "expectedFinding": { "code": "IMPOSSIBLE\_CAREER\_PROGRESSION", "severity": "BLOCKER", "path": "$.character.occupationHistory\[0\]", "blocksActivation": true }, "correctedPayload": { "birthDate": "1985-01-01" } }

source-register.csv

The underlying academic, temporal, and dependency parsing frameworks utilized to construct this specification are documented below, mapping fundamental research to specific validator applications.

The proposed Semantic Persona Validator successfully decouples structural JSON schema validity from deeply integrated semantic, chronological, and pragmatic validity. By prioritizing cross-linguistic dependency parsing and topological interval algebra, the architecture enforces rigorous coherence across large populations without sacrificing narrative diversity or cultural nuance.

Works cited

1. Narrative Theory-Driven LLM Methods for Automatic Story Generation and Understanding: A Survey \- arXiv, https://arxiv.org/html/2602.15851v2

2. Amory: Building Coherent Narrative-Driven Agent Memory through Agentic Reasoning, https://arxiv.org/html/2601.06282v1

3. Introduction \- Universal Dependencies, https://universaldependencies.org/docsv1/introduction.html

4. Cross-linguistically consistent semantic and syntactic annotation of child-directed speech \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12086124/

5. IA-RAG: Interval-Algebra–Driven Temporal Reasoning for Dynamic Knowledge Retrieval, https://arxiv.org/html/2606.06044v1

6. Allen's interval algebra \- Grokipedia, https://grokipedia.com/page/Allen's\_interval\_algebra

7. TLEX: An Efficient Method for Extracting Exact Timelines from TimeML Temporal Graphs \- arXiv, https://arxiv.org/pdf/2406.05265?

8. Benchmarking Large Language Models for Knowledge Graph Validation \- dei.unipd.it, https://www.dei.unipd.it/\~silvello/papers/2026\_EDBT\_FactCheck.pdf

9. LLM-based Reranking and Validation of Knowledge Graph Completion \- CEUR-WS.org, https://ceur-ws.org/Vol-3999/paper6.pdf

10. LLMs as Knowledge Graph Refiners: Mitigating Factual Inconsistencies in Generative Knowledge Extraction \- ACL Anthology, https://aclanthology.org/2026.acl-long.1353.pdf

11. LEEETs-Dial: Linguistic Entrainment in End-to-End Task-oriented Dialogue systems \- arXiv, https://arxiv.org/html/2311.09390v2

12. LEEETs-Dial: Linguistic Entrainment in End-to-End Task-oriented Dialogue systems \- ACL Anthology, https://aclanthology.org/2024.findings-naacl.46.pdf

13. Marrying Universal Dependencies and Universal Morphology \- ACL Anthology, https://aclanthology.org/W18-6011.pdf

14. Enhancing Korean Dependency Parsing with Morphosyntactic Features \- arXiv, https://arxiv.org/html/2503.21029v1

15. TLEX: an efficient method for extracting exact timelines from TimeML temporal graphs, https://users.cs.fiu.edu/\~markaf/doc/j11.ocal.2026.lre.60.10\_archival.pdf

16. TLEX: An Efficient Method for Extracting Exact Timelines from TimeML Temporal Graphs, https://arxiv.org/html/2406.05265v1

17. arXiv:2407.16347v2 \[cs.CL\] 31 Jan 2025, https://arxiv.org/pdf/2407.16347?

18. (PDF) DRIVE: Disfluency-Rich Synthetic Dialog Data Generation Framework for Intelligent Vehicle Environments \- ResearchGate, https://www.researchgate.net/publication/394081616\_DRIVE\_Disfluency-Rich\_Synthetic\_Dialog\_Data\_Generation\_Framework\_for\_Intelligent\_Vehicle\_Environments

19. Beyond Task-Oriented and Chitchat Dialogues: Proactive and Transition-Aware Conversational Agents \- ACL Anthology, https://aclanthology.org/2025.emnlp-main.672.pdf

20. KGValidator: A Framework for Automatic Validation of Knowledge Graph Construction \- arXiv, https://arxiv.org/html/2404.15923v1