Evaluating Prompt-Based and Fine-Tuned Approaches to Czech
Anaphora Resolution

D 2026

Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution

STANO, Patrik a Aleš HORÁK

Základní údaje

Originální název

Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution

Název česky

Evaluace Prompt-Based a Fine-Tuned metod řešení anafor na českém textu

Autoři

STANO, Patrik a Aleš HORÁK

Vydání

Cham, Text, Speech, and Dialogue, TSD 2025, od s. 190-202, 13 s. 2026

Nakladatel

Springer

Další údaje

Jazyk

angličtina

Typ výsledku

Stať ve sborníku

Obor

10200 1.2 Computer and information sciences

Utajení

není předmětem státního či obchodního tajemství

Forma vydání

elektronická verze "online"

Impakt faktor

Impact factor: 0.402 v roce 2005

Označené pro přenos do RIV

Ano

Organizační jednotka

Fakulta informatiky

ISBN

978-3-032-02550-0

ISSN

Klíčová slova anglicky

anaphora resolution; sequence-to-sequence models; fine-tuning; prompt engineering

Štítky

firank_B

Změněno: 14. 4. 2026 16:37, Mgr. Petra Trembecká, Ph.D.

Anotace

V originále

Anaphora resolution plays a critical role in natural language understanding, especially in morphologically rich languages like Czech. This paper presents a comparative evaluation of two modern approaches to anaphora resolution on Czech text: prompt engineering with large language models (LLMs) and fine-tuning compact generative models. Using a dataset derived from the Prague Dependency Treebank, we evaluate several instruction-tuned LLMs, including Mistral Large 2 and Llama 3, using a series of prompt templates. We compare them against fine-tuned variants of the mT5 and Mistral models that we trained specifically for Czech anaphora resolution. Our experiments demonstrate that while prompting yields promising few-shot results (up to 74.5\% accuracy), the fine-tuned models, particularly mT5-large, outperform them significantly, achieving up to 88\% accuracy while requiring fewer computational resources. We analyze performance across different anaphora types, antecedent distances, and source corpora, highlighting key strengths and trade-offs of each approach.

Návaznosti

LM2023062, projekt VaV

Název: Digitální výzkumná infrastruktura pro jazykové technologie, umění a humanitní vědy

Investor: Ministerstvo školství, mládeže a tělovýchovy ČR, LINDAT/CLARIAH-CZ - Digitální výzkumná infrastruktura pro jazykové technologie, umění a humanitní vědy

90254, velká výzkumná infrastruktura

Název: e-INFRA CZ II

Citovat

STANO, Patrik a Aleš HORÁK. Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution. Online. In Text, Speech, and Dialogue, TSD 2025. Cham: Springer, 2026, s. 190-202. ISBN 978-3-032-02550-0. Dostupné z: https://doi.org/10.1007/978-3-032-02551-7_17.

@inproceedings{2505377,
   author = {Stano, Patrik and Horák, Aleš},
   address = {Cham},
   booktitle = {Text, Speech, and Dialogue, TSD 2025},
   doi = {https://doi.org/10.1007/978-3-032-02551-7_17},
   keywords = {anaphora resolution; sequence-to-sequence models; fine-tuning; prompt engineering},
   howpublished = {elektronická verze "online"},
   language = {eng},
   location = {Cham},
   isbn = {978-3-032-02550-0},
   pages = {190-202},
   publisher = {Springer},
   title = {Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution},
   year = {2026}
}

TY  - CONF
ID  - 2505377
AU  - Stano, Patrik - Horák, Aleš
PY  - 2026
TI  - Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution
PB  - Springer
CY  - Cham
SN  - 9783032025500
KW  - anaphora resolution
KW  - sequence-to-sequence models
KW  - fine-tuning
KW  - prompt engineering
N2  - Anaphora resolution plays a critical role in natural language understanding, especially in morphologically rich languages like Czech. This paper presents a comparative evaluation of two modern approaches to anaphora resolution on Czech text: prompt engineering with large language models (LLMs) and fine-tuning compact generative models. Using a dataset derived from the Prague Dependency Treebank, we evaluate several instruction-tuned LLMs, including Mistral Large 2 and Llama 3, using a series of prompt templates. We compare them against fine-tuned variants of the mT5 and Mistral models that we trained specifically for Czech anaphora resolution. Our experiments demonstrate that while prompting yields promising few-shot results (up to 74.5\% accuracy), the fine-tuned models, particularly mT5-large, outperform them significantly, achieving up to 88\% accuracy while requiring fewer computational resources. We analyze performance across different anaphora types, antecedent distances, and source corpora, highlighting key strengths and trade-offs of each approach.
ER  -

STANO, Patrik a Aleš HORÁK. Evaluating Prompt-Based and Fine-Tuned Approaches to Czech Anaphora Resolution. Online. In \textit{Text, Speech, and Dialogue, TSD 2025}. Cham: Springer, 2026, s.~190-202. ISBN~978-3-032-02550-0. Dostupné z: https://doi.org/10.1007/978-3-032-02551-7\_{}17.

Přehled o publikaci