Building the Parsed Corpus of Scottish Correspondence

Repository location

urn:nbn:fi:lb-2024070601

Repository name

Kielipankki – the Language Bank of Finland (FIN-CLARIN)

Object name

‘The Parsed Corpus of Scottish Correspondence, source’

Format names and versions

corpus files: .psd; metadata: .csv; documentation: .txt

Creation dates

2018-12-01 to 2021-05-05

Dataset creators

Lisa Gotthard (creator/annotator), University of Edinburgh; Rob Truswell (training and supervision), University of Edinburgh; Beatrice Santorini (training and supervision), University of Pennsylvania; Sarah Einhaus (manual correction of a subset of PoS tags), University of Konstanz.

Language

Older Scots

License

CLARIN_ACA-NC

Publication date

2024-07-06

(3.1) Source data and preparation

ScotsCorr preserves original spelling, punctuation, and markup of some manuscript features, making it an excellent resource for analysing a historical variety while also presenting challenges for automatic annotation. Rather than modernising the texts, the guiding principle during data preparation was to preserve the linguistic properties of the source material wherever possible. Editorial interventions therefore focused on improving compatibility with automatic annotation tools rather than normalising the language itself. Data preparation comprised five principled decisions and stages:

1. Metadata embedded within individual text files were extracted into a separate spreadsheet linked to each document through a unique identifier.¹

2. Formatting markup, including HTML line-break tags and layout information, was removed, while HTML character entities were converted into their corresponding characters (e.g., ‘&amp:’ in ScotsCorr is replaced by ‘&’ in the PCSC). Parentheses were replaced with placeholder strings (<lparen>/<rparen>) to prevent conflicts with later automated parsing routines.

3. Editorial comments describing manuscript damage, deletions, or alternative readings were retained only where they contributed to the interpretation of sentence structure; comments describing purely palaeographical features were discarded.

4. Existing conventions representing contractions and superscripts were preserved because they form part of the diplomatic transcription and can provide linguistically relevant information.

5. No spelling normalisation or lexical modernisation was performed. Retaining the original orthography ensures that the corpus remains suitable not only for syntactic research but also for investigations of orthographic and lexical variation and change.

The example in (1) serves as illustration of a piece of text from ScotsCorr before (1a.) and after (1b.) the clean-up steps described in 1–5.

1. 1. (1)

   1. a.

2. 1. 1. writtin

      2. at

      3. Edinburgh

      4. ye

      5. xxiij

      6. day

      7. of

      8. \

      9. </line>

      10. <line>

      11. november

      12. Be

      13. {space}

      14. {a wide space vertically}

      15. \\

      16. </line>

      17. <line>

      18. Zour

      19. gracis

      20. hvmill

      21. s*er%uatris

      22. \\

      23. </line>

      24. <line>

      25. Katherine

      26. Bellenden

      27. {address>}

1. 1.  

   1. b.

2. 1. 1. writtin

      2. at

      3. Edinburgh

      4. ye

      5. xxiij

      6. day

      7. of

      8. november

      9. Be

      10. Zour

      11. gracis

      12. hvmill

      13. s*er%uatris

      14. Katherine

      15. Bellenden

Because complete manual correction of the entire ScotsCorr exceeded the available project resources, annotation was completed for a representative subset comprising 932 letters (ca. 270,500 words²). The subset was designed to improve the gender balance of the seventeenth-century material by sampling male-authored correspondence from the largest sub-corpora (1600–49, 1650–1699); that is, all the material from female authors were included from the seventeenth century, while only a subset of the male-authored material from that century was included in order to make the male-female proportion more balanced (cf. Tables 1 and 2).

The excluded ScotsCorr texts have undergone automatically generated annotation (“parse skeleton”, as described in Section 3.4, see (2b.)) and provide a foundation for future manual correction and corpus expansion.

(3.2) Annotation framework

The PPCHE annotation framework has been applied to numerous historical corpora covering English as well as a range of other languages. The PPCHE framework was designed to support reproducible syntactic research through a consistent phrase-structure representation that minimises theoretical assumptions. One distinctive feature is its relatively flat treatment of verbal structure, in which verbs are represented as immediate daughters of the clause node rather than within a deeper verb phrase (see Truswell et al. 2019, p. 19–22; Walkden, 2016). This design facilitates automatic searching across corpora and reduces ambiguity during annotation.

The PPCHE framework is accompanied by two purpose-built software tools that were used throughout this project: CorpusSearch2 (CS; Randall, 2010; Randall et al., 2004), which was used for automatic revision and corpus querying, and Annotald (Beck et al., 2011), an annotation editor for manual correction of parses. Although CS remains widely used within the PPCHE community and does not require ad hoc workarounds to run, it is a mature software package whose continued maintenance is community-led. Annotald, by contrast, required a number of ad hoc workarounds to run even at the time of corpus construction; although it is no longer actively maintained, community-developed patches have recently improved its compatibility with modern systems, and it remains the annotation editor used for many existing PPCHE corpora. Following annotation, CS was used in a series of studies using the corpus data, both to query the parsed corpus and to generate coding-string outputs which were converted into .csv data frames, merged with the corpus metadata, and imported into R for quantitative analysis (see Gotthard, 2022, pp. 44–6; 2024b; 2026).

(3.3) Part-of-Speech tagging

PoS-tagging combined automatic dictionary-based annotation with iterative manual refinement and correction. Automatic tagging was implemented using custom Python scripts³ that compared each token in the corpus against two lexical resources: a tag dictionary derived from the Linguistic Atlas of Older Scots (LAOS; Williamson, 2024) and the Penn Treebank tag dictionary. Because neither LAOS nor the Penn Treebank employed the PPCHE tagset, both dictionaries were converted to PPCHE-compatible labels before annotation. Using both resources accommodated the mixture of Scots and increasingly anglicised forms present within the corpus.

Tokens that could not be matched automatically were initially assigned an ERROR label. These unresolved forms were subsequently addressed through an iterative correction process involving (i) expansion of the dictionaries, (ii) regular-expression matching of predictable orthographic patterns, and (iii) context-sensitive revision queries implemented in CS (e.g., an ERROR form occurring between a determiner and noun received a default tag for adjective (ADJ)). Remaining tagging errors were corrected during manual review of the syntactic annotation.

(3.4) Syntactic annotation

Syntactic annotation likewise combined automated procedures with manual review and correction. CS revision queries⁴ were first used to construct an initial parse by wrapping individual PoS-tagged words into phrase structures and progressively expanding these structures to include complements and adjuncts. Functional information, such as probable subjects, was also added automatically where reliable structural cues existed – for example, a Wh-form could be assumed to be followed by an empty C category.

Automatic parsing was necessarily conservative. Historical punctuation, preserved in ScotsCorr, does not consistently indicate sentence boundaries, and many constructions require contextual interpretation that cannot be inferred reliably through rule-based procedures. Consequently, automatically generated structures were regarded as a first-pass analysis rather than a completed parse. This stage generated a form of “parse skeleton” that could then be corrected manually – in (2) the progress from PoS-tagged, unparsed, text (2a.) to parse skeleton (2b.) is illustrated.

1. 1. (2)

   1. a.

2. 1. 1. (PRO I)

      2. (DOP doe)

      3. (NEG not)

      4. (VB denie)

      5. (PRO I)

      6. (MD will)

      7. (DO doe)

      8. (PRO it)

1. 1.  

   1. b.

2. 1. 1. (IP (NP-SBJ (PRO I))

      2. (DOP doe)

      3. (NEG not)

      4. (VB denie))

      5. (IP (NP-SBJ (PRO I))

      6. (MD will)

      7. (DO doe)

      8. (NP (PRO it)))

Manual correction was undertaken using Annotald. During this stage, constituent boundaries, phrase labels, functional annotations, and attachment decisions were reviewed and corrected individually. The combination of automated parsing and manual validation provided a balance between efficiency and annotation quality while maintaining consistency with other corpora using the PPCHE framework. The final parse of the sentence in (2) can be seen in (3).

1. 1. (3)

2. 1. 1. ((IP-MAT

      2.         (NP-SBJ (PRO I))

      3.         (DOP doe)

      4.         (NEG not)

      5.         (VB denie)

      6.                 (CP-THT (C 0)

      7.                         (IP-SUB

      8.                                 (NP-SBJ (PRO I))

      9.                                 (MD will)

      10.                                 (DO doe)

      11.                                 (NP-OB1 (PRO it)))))

      12. (ID 293_2_F1700,.16))

The example in (3) also illustrates an advantage of a parsed corpus over an unannotated one; the construction contains an embedded clause with an empty complementiser, (CP-THT (C 0)), as well as both auxiliary and lexical uses of do, (DOP doe) and (DO doe) respectively – features that would be difficult and time-consuming to identify reliably using only surface-string searches.

(3.5) Quality assurance

Quality assurance was integrated throughout the annotation process rather than performed as a single validation stage. Three complementary procedures were used to maximise annotation consistency.

First, repeated cycles of automatic annotation and CS revision queries were used to identify and correct systematic errors. As recurring tagging and parsing errors were identified during manual inspection, new CS revision queries were written to apply the same corrections across the corpus automatically, reducing repeated manual intervention.

Second, all completed parses were manually reviewed using Annotald. During this stage, constituent structure, phrase labels, functional tags, and attachment decisions were checked individually and corrected where necessary. Manual review was particularly important for constructions that could not be resolved reliably through rule-based procedures, such as matrix clause boundaries and ambiguous constituent attachment.

Third, once manual correction had been completed, the corpus was validated using a suite of CorpusSearch “sanity check” queries developed for PPCHE-style corpora (provided by Beatrice Santorini⁵). These queries identify inconsistencies that may not be obvious during manual annotation, such as unmatched co-indexation, providing a final check of annotation consistency before release.

Finally, as with any manually annotated corpus, some annotation errors or inconsistencies are likely to remain despite the procedures described above. While no revision has been required as of yet, users are encouraged to report any errors they encounter so that these can be incorporated into future revisions of the corpus.

(4.1) A new parsed corpus of Early Modern Scots

The principal output of this project is the PCSC, the first and currently only parsed corpus of Scots; the completed corpus comprises 270,553 words distributed across 932 letters written by 328 authors between 1543 and 1747, preserving the rich metadata inherited from ScotsCorr while adding PoS tags and full syntactic annotation.

(4.2) Balancing automation and manual annotation

At the time the corpus was developed, no parser trained on historical Scots was available, and extensive orthographic variability limited the usefulness of existing NLP tools. Consequently, a semi-automatic workflow combining rule-based annotation with manual correction provided the most reliable strategy. One of the principal methodological outcomes of the project was the adaptation of established annotation workflows to the particular characteristics of ScotsCorr and Early Modern Scots. This experience suggests that similar hybrid workflows may continue to be useful for future corpus-building projects for historical and under-resourced language varieties.

(4.3) Preserving historical variation

The decision to preserve the diplomatic transcription of ScotsCorr, despite the additional challenges it posed for automatic annotation, reflects a broader principle in historical corpus construction: While normalisation may improve automatic processing, it can also remove evidence relevant to linguistic research. By preserving original orthography, the PCSC can be used for a wider range of research questions than syntactic analysis alone, particularly studies of orthographic change, lexical variation, and scribal practice (such as visible corrections, or other textual elements), albeit at the cost of increased annotation effort.

(4.4) Limitations

Several limitations should be considered when using the PCSC.

First, only a subset of the original ScotsCorr has undergone complete manual correction; although the remaining material has been automatically tagged and parsed, it does not currently achieve the same level of annotation reliability and is not yet made available.

Second, the corpus inherits the biases of the source material; some chronological periods contain substantially more correspondence than others, and gender representation remains uneven despite efforts to improve balance through targeted sampling. These imbalances reflect the historical survival of documents rather than annotation choices, but they nevertheless influence the types of quantitative analyses that can be undertaken.

Third, there is a learning curve involved with interpreting and using the PPCHE annotation scheme and adjacent software; users unfamiliar with PPCHE-style corpora may require some initial training in these methods in order to use the PCSC.

Additionally, adopting the PPCHE framework necessarily involves certain theoretical commitments; although the annotation scheme was designed to minimise theory-specific assumptions, its relatively flat constituent structure differs from many contemporary syntactic analyses. Furthermore, some annotation conventions, such as the use of empty categories to represent movement dependencies, may not align with all theoretical frameworks and can influence the interpretation of syntactic structures. Researchers using the corpus should therefore interpret the annotation as a practical representation optimised for corpus searching rather than a theoretical account of syntactic structure.

Finally, unlike some recently developed historical corpora, the PCSC does not include lemma annotation. The primary aim of the project was to produce a syntactically parsed corpus compatible with the PPCHE framework, which traditionally does not incorporate lemmatisation (although recent work has added partial or full lemmatisation to some of these corpora, see Trips & Percillier (2020)). Due to the time and resource limitations of the project, priority was given to accurate part-of-speech and syntactic annotation. Nevertheless, the dictionary-based tagging workflow provides a potential basis for future lemmatisation – however, reliable lemma annotation would require additional manual correction because of the extensive orthographic and lexical variation present in the corpus.