Parsing the Past - Identification of Verb Constructions in Historical Text

Eva Pettersson, Beáta Megyesi, Joakim NivreLaTeCH@EACL 2012 · OpenReview

Abstract

Researchers at Uppsala University developed a pipeline to automatically extract verb constructions from Early Modern Swedish (1550–1800) by "translating" historical spelling into modern forms before applying standard NLP tools. Despite the linguistic "noise" of the past, the system achieved a 77.2% F-score in identifying verbs and surprisingly outperformed modern text benchmarks in exact-match accuracy for complete verb-complement sets.

Introduction

Historians involved in the "Gender and Work" project face the monumental task of manually extracting work-related activities from centuries of court records and church documents. In these historical texts, labor is typically described through verb constructions—verbs paired with their respective objects or modifiers. However, applying modern Natural Language Processing (NLP) tools to these documents is problematic. Historical Swedish (1550–1800) lacks standardized spelling and exhibits significant grammatical shifts compared to modern Swedish, which modern parsers are trained on.

To address this, researchers developed a "normalize-then-parse" architecture. The study utilizes a corpus of 10 historical texts spanning 1527 to 1737, totaling approximately 444,000 tokens. A manually annotated "gold standard" of 400 sentences was created to evaluate the system's performance against contemporary benchmarks like the Stockholm-Umeå Corpus (SUC).

Methodology

The proposed pipeline allows the use of high-quality modern NLP tools without requiring extensive historical training data. It consists of five key stages:

  1. Tokenization: Segmenting text into individual words.
  2. Normalization: The core innovation, using 29 hand-crafted rules to map historical orthography to modern Swedish (e.g., changing "schall" to "skall").
  3. Part-of-Speech Tagging: Utilizing HunPOS, trained on modern Swedish, to identify verbs.
  4. Dependency Parsing: Using MaltParser to determine grammatical relationships and complements.
  5. Valency Filtering: Applying modern valency dictionaries to prune "hallucinated" complements that do not align with a verb's known linguistic properties.

Results and Discussion

The normalization step proved to be the most critical component of the pipeline. Without it, the system's recall for verb identification was only 60.0%. Implementing basic spelling rules increased recall to 76.0%, resulting in an overall F-score of 77.2%. Overall, the system successfully identified a partially or completely correct set of complements for 55.6% of the verbs in the historical corpus.

Performance of verb identification on raw vs. normalized historical text compared to modern SUC benchmarks.

A surprising finding, termed the "Archaic Advantage," emerged during complement extraction. The system achieved a higher exact-match rate for full verb constructions in historical text (38.7%) than in modern journalistic text (30.8%). This is attributed to the formulaic and structurally simpler nature of historical legal and administrative records compared to the dense prose of modern news.

Comparison of complement extraction accuracy between historical (Raw/Norm) and modern (SUC) datasets.

Furthermore, the stability of verb valency over 400 years allowed modern dictionaries to effectively refine the parser's output. Adding valency filtering boosted the percentage of fully correct constructions from 34.5% to 38.7%, although it occasionally introduced "missing" complements if the dictionary was overly restrictive.

Conclusion

The study demonstrates that orthographic variance is a more significant hurdle than grammatical change for parsing historical Swedish. While the current rule-based normalization is effective, its performance varies by period, suggesting a need for "period-aware" adaptations. Future work will focus on refining these rules for specific genres and integrating the tool into interactive interfaces for historians to verify and accelerate their data extraction processes.

Summary updated

BibTeX
@inproceedings{pettersson2012parsing,
  author = {Eva Pettersson and Beáta Megyesi and Joakim Nivre},
  title = {{Parsing the Past - Identification of Verb Constructions in Historical Text}},
  booktitle = {LaTeCH@EACL},
  year = {2012},
  url = {https://openreview.net/forum?id=RE1IiJuPRHu},
}
Figure