Evaluating Lexical Extraction in Hiberno-English: Rule-Based and LLM- Assisted Approaches

This paper presents and evaluates two computational approaches for parsing A Dictionary of  Hiberno-English (Terence Patrick Dolan)(Dolan, 2006) into a structured, machine-readable  lexical resource. Although the dictionary exhibits a high degree of editorial consistency, its  entries remain computationally challenging to structure (Atkins & Rundell, 2008) . Lexical  information is distributed across semi-regular templates that are frequently disrupted by  redirects, variant spellings, homonym numbering, multi-part grammatical labels, numbered  senses, phrase entries, and entries with missing or atypical pronunciation markers. Further  complexity arises because region codes, informant identifiers, editorial asides, etymologies,  examples, and cross-references are often embedded within running prose or parenthetical  material rather than isolated in stable positions. As a result, extraction requires field recognition  and careful boundary resolution, normalization, and disambiguation (Ide & Véronis, 1995). 

To address these challenges, we implement and compare two complementary approaches: a  rule-based parser and a large language model (LLM)-assisted extraction pipeline(Brown et al.,  2020; Jurafsky & Martin, 2023). The rule-based system uses deterministic segmentation and  pattern recognition to identify core lexical components, including headwords, pronunciation,  variants, part-of-speech labels, definitions, etymologies, examples, and cross-references, while  explicitly handling recurrent irregularities such as redirect entries, variant forms, numbered  senses, compound grammatical labels, and embedded metadata. The generative pipeline, by  contrast, employs schema-driven prompting, encoding the target data model within the prompt  itself targeted at field-consistent outputs and reducing ambiguity in the model’s interpretation  of semi-structured inputs and edge cases(Ji et al., 2023). 

To enable systematic comparison, we develop an evaluation framework that captures multiple  dimensions of extraction quality, including (1) field-level presence and absence, (2) exact string  agreement between methods, (3) token-level similarity using both ordered and unordered  matching, and (4) source coverage through token- and character-level containment checks targeted at detection of overgeneration and omissions. These measures function as lightweight proxies for identifying hallucination-like behaviour in generative outputs, without requiring  exhaustive manual annotation. 

Our evaluation reveals distinct performance profiles across the two approaches. The rule-based  system achieves high precision and strong source fidelity, producing outputs that remain  consistently traceable to the input text, but is sensitive to formatting variation and requires  substantial manual rule engineering. By contrast, the LLM-based system demonstrates greater  robustness to irregular structures and achieves higher recall on complex entries but exhibits  variability in output length and occasional divergence from the source text. These behaviours  are reflected in overgeneration patterns and reduced containment scores, highlighting trade offs between flexibility and control. 

By systematically comparing these approaches, we show that LLM-based extraction can  produce usable lexical data when guided by schema-driven prompting and evaluated through  explicit, field-aware metrics. Rather than treating generative outputs as opaque, we  demonstrate how they can be assessed through multi-dimensional measures, including field 

level validation, token similarity, and source coverage. In the wake of recent LLM-driven  approaches, our findings suggest that some aspects of lexical structuring can be achieved with  reduced reliance on extensive rule engineering, opening new possibilities for working with  semi-structured historical resources. 

References 

Atkins, B. T. S., & Rundell, M. (2008). The Oxford Guide to Practical Lexicography. Oxford  University Press. 

Brown, T. B., Mann, B., Ryder, N., & al, et. (2020). Language Models are Few-Shot  Learners. Advances in Neural Information Processing Systems (NeurIPS). Dolan, T. P. (2006). A Dictionary of Hiberno-English: The Irish Use of English (2nd ed.).  Gill & Macmillan. 

Ide, N., & Véronis, J. (1995). Encoding Dictionaries. In Text Encoding Initiative:  Background and Context. Kluwer Academic Publishers.

Ji, Z., Lee, N., Frieske, R., & others. (2023). Survey of Hallucination in Natural Language  Generation. ACM Computing Surveys

Jurafsky, D., & Martin, J. H. (2023). Speech and Language Processing.