Abstract
This paper points out a linguistic phenomenon that, due to the current stage of research, can be analyzed only insufficiently with the help of electronic text corpora. It adds a new aspect to the discussion about historical corpora by tackling the question of how they should be designed to be useful for linguistic research on so-called formulaic patterns. In section 1, we define the term formulaic pattern as a prerequisite for collaborative research. Section 2 outlines the state of the art in modern formulaic language within corpus and computer linguistics. Section 3 demonstrates how problems in this area are exacerbated when applied to historical texts. Section 4 presents a solution implemented by the HiFoS Researchers' Group at the University of Trier, while section 5 discusses joint research with the UKP Lab at TU Darmstadt, showing that the restrictions posed by historical patterns are challenges to be overcome rather than insurmountable obstacles.
Introduction: The "Pain in the Neck" of Computational Linguistics
In modern linguistics, it is well-established that language is not merely a sequence of isolated words but is constructed from "formulaic patterns"—including idioms, proverbs, and collocations (e.g., "to spill the beans" or "Ladies and Gentlemen"). While these structures are essential for human communication, they represent a significant challenge for computational detection due to their often low frequency and high structural variation. This difficulty is magnified tenfold when examining historical German texts spanning from 700 to 1650 AD.
The Challenge of Historical Corpora
Researchers investigating historical formulaic language face a "perfect storm" of obstacles that render standard modern NLP tools ineffective:
- Lack of Orthographic Standards: Historical German lacks the spelling norms found in modern language, leading to massive variation in how the same phrase might be written.
- Data Scarcity: The surviving record is fragmentary, providing only a partial view of the linguistic landscape of the past millennium.
- Corpus Design Flaws: Many existing electronic corpora are either too small for statistical significance or consist of "normalized" snippets that strip away the very linguistic variations that are the object of study.
The authors argue that current digital tools are insufficient for analyzing how these linguistic building blocks evolved, necessitating a new framework for corpus design.
Methodology: Bridging Philology and NLP
To address these challenges, the paper describes a collaborative methodology between the HiFoS Research Group (University of Trier) and the UKP Lab (TU Darmstadt). This approach combines traditional philological expertise with modern computational power through several key stages:
Intellectual Extraction and Annotation
Unlike standard projects that rely on automated scripts from the outset, the HiFoS team manually identifies and annotates patterns from original manuscripts. This ensures that the "frozen" or "semi-frozen" nature of the language is captured accurately. Patterns are encoded using Unicode and TEI (Text Encoding Initiative) standards, incorporating metadata on morphology, syntax, pragmatics, and cultural-historical context.
The Normalization Pipeline
To bridge the gap between ancient text and modern tools, the UKP Lab utilizes the DKPro repository. They implement a "modernization" layer where archaic words are mapped to modern equivalents. This allows NLP tools trained on modern German to assist in the identification and processing of historical data.
Expert-in-the-Loop
The framework proposes a task-oriented user interface where domain experts inspect and correct automatic processing results. These corrections are fed back into the system to improve the accuracy of future automated runs, creating a virtuous cycle of refinement.
The HiFoS Belegkorpus
The primary output of this research is the HiFoS Belegkorpus, a specialized database of annotated formulaic patterns. The dataset comprises approximately 30,250 fully annotated entries, categorized by temporal period:
- Old High German (700–1050): ~9,494 entries.
- Middle High German (1050–1350): ~11,644 entries.
- Early Modern High German (1350–1650): ~8,973 entries.
Crucially, the source material extends beyond poetry and fiction to include legal texts, travel reports, and religious documents, providing a broader view of historical formulaic usage.
Key Insights: Variation as a Feature
A central finding of the project is that in historical German, variation is not an error but the most characteristic feature. A single formulaic pattern may appear in dozens of different syntactic or morphological forms across different centuries and regions.
Furthermore, the authors advocate for the term "formulaic pattern" over "phraseme." This broader terminology includes expressions that may have low idiomaticity or syntactic stability but serve a clear pragmatic function. The research also highlights a "Corpus Size Paradox": even a 100-million-word corpus may be insufficient to statistically detect certain patterns because, while the category of formulaic language is common, individual instances are often rare.
Results and Future Directions
This research marks a significant shift from purely manual philology to a hybrid "computational philology." By creating a 30,000+ entry "gold standard" dataset, the project provides a foundation for training future machine learning models to recognize historical multi-word expressions. It has successfully demonstrated that "formulaic traditions" (Formulierungstradition) can be tracked across centuries, showing how specific expressions were handed down and modified, often under the influence of Latin or Greek.
While fully automatic identification remains a future goal, the project has established that the restrictions of historical data are not insurmountable. Future work aims to scale the corpora and establish cross-lingual links with Latin and Greek databases, as well as historical art databases, to explore the intersection of verbal and visual formulaic expressions.
Summary updated