Task-oriented hybrid conversational systems: towards flexible, domain-aware and reliable conversational interaction combining LLMs and rule-based approaches
Year: 2026
Type: dissertation
Abstract: This PhD thesis investigates how to design reliable, expressive, and data-driven conversational AI systems for multilingual settings. It addresses three overarching objectives: (O1) automatically extracting dialog structure and using it to control interaction with large language model (LLM)-based techniques; (O2) developing LLM-based systems that provide reliable, explainable, and non-hallucinated answers via evidence grounding; and (O3) creating specialized conversational resources to train and fine-tune dialog models on multi-turn interactions rather than generic plaintext corpora. Rather than treating these objectives as isolated challenges, the thesis frames them as mutually reinforcing requirements: structure is needed to keep interaction coherent and controllable, grounding is needed to make responses auditable and trustworthy, and dialog-native data is needed to learn and evaluate these properties in realistic multi-turn settings. For objective O1, the dissertation explores hybrid, LLM-assisted task-oriented dialog methods in both open and proprietary settings. One line of work proposes a two-phase pipeline that converts raw task-oriented dialog logs into operational Rasa assistants using LLM-generated artifacts, validator-driven structural bootstrapping, and harness-based interaction testing. The work completed during the international research stay at AppTek - RWTH Aachen University extends the methodology to AppTek’s proprietary Cheshire framework, showing that LLM-based authoring remains feasible under strict domain-specific language (DSL) and runtime constraints when combined with validator-driven micro-repairs and KPI-based stabilization, and delivering a reusable taxonomy of framework-specific failure modes, repair templates, and human-in-the-loop design recommendations. A complementary line introduces SGD+, an extension of schema-guided dialog state tracking (DST) that treats slot values as sets with Boolean semantics, temporal intervals and periodicity, and spatial hierarchies, together with Spanish template-generated datasets and set-based DST metrics. For objective O2, the thesis develops guideline-bounded, Retrieval-Augmented Generation (RAG) systems for Spanish clinical question answering. Using national Clinical Practice Guidelines as the only knowledge source, the proposed architectures combine dense retrieval, reranking, and LLM-based answer synthesis with explicit citation and refusal behavior. Expert evaluations by medical professionals show that advanced RAG pipelines substantially improve reliability, clarity, and traceability compared to naïve retrieval, and that well-configured open-source models can rival proprietary systems under strict evidence constraints. For objective O3, the PhD thesis introduces large-scale multilingual conversational and web resources and complements them with controlled synthetic benchmarks. esCorpiusDialog aggregates 30M dialogs (122.8M turns) from subtitles, forums, newsgroups, and books, while esCorpius-m provides a massive, cleaned multilingual web corpus with a strong Spanish focus. Additional resources include Chat- Subs (which also serves as part of esCorpiusDialog) and esCorpiusBIAS, SGD+ provides Spanish template-guided datasets for expressive dialog state tracking and negotiation, supporting set-valued supervision over complex temporal, spatial, and multi-objective constraints, with all instances verified by a native Spanish linguist. Across these resources, the pipelines are fully documented and reproducible, and validated through downstream experiments on multi-turn dialog modeling and safety classification. The three objectives tie together into a single design stance for multilingual conversational AI. O3 supplies the dialog-native data and controlled benchmarks needed to learn and stress-test multi-turn behavior; O1 provides the explicit structure - schemas, states, and flow constraints - that keeps interaction coherent and executable; and O2 supplies the evidence governance that makes answers traceable, defensible, and safe in high-stakes use. Their combination yields systems that are not only fluent, but also structurally controlled, evidence-based, and adapted to real-world multi-turn interaction across languages and domains. The results demonstrate that robust multilingual conversational AI requires (and benefits from) a combination of high-quality multi-turn data, including controlled synthetic benchmarks for underrepresented phenomena, evidence-grounded LLM architectures, and hybrid development workflows that integrate LLMs with interpretable, schema- and rule-based control. These results have also led to the publication of eight research papers as part of the thesis.
Author: Kharitonova Yakusheva, Ksenia
Institution: Universidad de Granada
Language: English
Topic: Speech and dialogue systems
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: University of Cambridge
OpenAlex ID: W7171482074
