Speaker
Description
This poster presents the research on the application of different approaches to the information extraction process and encoding of the 18th century Larramendi’s Trilingual Dictionary. The digitisation and computational analysis of historical dictionaries present significant challenges due to archaic typefaces, complex layouts, and inconsistent lexicographic structures. Manuel de Larramendi's Diccionario trilingüe castellano, bascuence y latin (1745, hereafter LAR), a foundational text for modern Basque lexicography, remained largely inaccessible in a structured digital format, available only as scanned facsimile images.
Initial digitisation efforts employed a semi-automatic workflow combining machine learning (ML) for Optical Character Recognition (OCR) using Kraken and rule-based information extraction. The OCR for LAR was completed, and the resulting transcription was uploaded to Wikisource, a crowdsourcing platform that provides an interface for collaborative facsimile transcriptions. That process is ongoing and will very soon lead to an error-free digital version of the original text. Wikisource transcription quality is ensured by its rigorous, community-driven validation process, as completed works usually go through a two-step human review. The rule-based information extraction process facilitated the identification of Spanish headword candidates using positional data (line indentation) and orthographic features (capitalisation corresponding to alphabet sections). The extracted data was further enriched by linking Spanish lemmata to a contemporary reference, the Diccionario de Autoridades (1726-1739), and mapping historical Basque forms to modern standard lexicons. This work demonstrated the feasibility of converting a historical printed dictionary into structured datasets and linked data, though it highlighted limitations in fully capturing the complex microstructure of entries through rule-based methods alone (Lindemann & Alonso, 2021).
The first approach to information extraction from LAR involved a hybrid methodology. For a more nuanced structural analysis, it was then employed the Elexifier toolchain, a ML-based application. A sample of dictionary columns was manually annotated with tags from the TEI-Lex0 schema and used for training a model to predict and segment lexicographic components—such as headwords, Basque and Latin equivalents—across the entire dictionary text. This ML approach proved complementary to the rule-based method, showing high precision for Spanish headwords and Basque translations but lower accuracy for Latin equivalents and cross-references, revealing the complexity of the dictionary’s microstructure.
Advancing beyond specialised ML tools like Elexifier, contemporary research explores the application of general-purpose Large Language Models (LLMs) with vision capabilities for similar tasks. Inspired by methodologies applied to Estonian-German dictionaries (Jürviste & Jakobson, 2025) or to Swedish historical patent cards using the GPT-4o model (Xie et al., 2025), a third approach involves using DeepSeek, an advanced LLM, for the information extraction and encoding of LAR. The extraction of key microstructural elements like headwords (Spanish), translation equivalents (Basque, Latin), part-of-speech indicators, and example phrases and sentences is demanded from standard files, such as XML or JSON, or from the wikitext format text from Wikisource, with the few-shot prompting strategy, which basically consists of showing the model a few examples of inputs and desired outputs before asking it to solve the task. A fourth approach involves employing Microsoft Copilot, another multimodal LLM, for the same extraction task. The process consists of feeding Copilot with the same input and few-shot prompting, requesting structured TEI-Lex0 output that represents LAR’s tripartite entries.
Following an experimental framework used for historical dictionaries (Jürviste & Jakobson, 2025), this allows for an evaluation of different model performances on identical material. Key evaluation metrics include the structural correctness of the generated markup (adherence to TEI-Lex0) and the model’s ability to correctly classify and relate the Spanish, Basque, and Latin lexical units within each entry. Preliminary analyses of the results suggest that DeepSeek encodes TEI-Lex0 schema with almost no syntax errors and can correct the dictionary’s microstructural inconsistencies in the encoded output file, whereas Microsoft Copilot tends to fail to follow consistently the encoding schema showed in the given examples.
This comparative analysis extends the results of the mentioned studies about the strengths and weaknesses of different LLMs in handling the specific challenges posed by early modern bilingual/multilingual dictionary digitisation, particularly for languages with less digital representation. Nevertheless, it would be interesting for further digitisation processes to conduct experiments providing the tools with scanned images, because unlike the earlier workflow requiring separate OCR and segmentation stages, a vision-enabled LLM could potentially perform text recognition and semantic structuring in a single, integrated step. This approach promises a more direct path from facsimile to structured data, potentially reducing pipeline complexity and adapting more flexibly to layout irregularities.
In conclusion, the initial work on LAR established a viable pipeline and valuable datasets from rule-based and specialised ML tools, although the evolution to general-purpose (vision-enabled) LLMs represents a significant shift in historical dictionary digitisation methodology. Approaches leveraging LLMs aim to integrate OCR, segmentation, and semantic encoding into a more efficient workflow. This research not only advances the digitisation of a key Basque lexicographic resource but also can contribute to computational historical lexicography by evaluating and refining modern AI-driven extraction techniques.