Literary & Corpus Stylistics

1. What this workspace does

Literary & Corpus Stylistics helps you locate textual patterns, inspect their exact evidence, compare selected texts, and record your interpretation. It combines close reading with reproducible corpus observations. It is not an academic-writing quality grader, an automatic literary critic, or a test of an author's merit.

A target is a text you want to investigate. A reference corpus is a deliberately chosen comparison set. A finding is a candidate observation: a repeated opening, for example, or a word that is relatively frequent in the target.

The workspace does not produce an overall foregrounding score. Counts cannot establish beauty, intentionality, reader response, or cultural meaning by themselves. Those claims need your contextual argument and, where appropriate, further research.

You can explore the public Load read-only demonstration without signing in. Its review and export controls are disabled; it does not save work. Sign in to create and use your private workspace.

2. Your first study, step by step

1. Sign in and open the workspace. Navigate to Literary & Corpus Stylistics from ADAR's home page or dashboard.
2. Start small. Choose one poem, speech excerpt, or short prose passage you have permission to use. Preserve the punctuation, paragraphing, and line breaks you intend to study.
3. Add the target. Under Add to corpus, paste the text, supply a useful title and metadata, select its language and the Target role, then choose Add document(s) to corpus.
4. Inspect the saved text. Use Edit in the corpus library to check the source and metadata. Correct extraction problems before running a study.
5. Optionally add a reference. Add an appropriate comparison text with the Reference role. You can begin without one, but cannot then establish deviation against an external corpus.
6. Select the study. Under Prepare an analysis, enter a descriptive report title and select the intended target and reference documents. Do not rely on a document's role label alone; check the actual selections.
7. Run it. Choose Analyze selected texts. Open the evidence, methodology, comparisons, warnings, and internal-validation information.
8. Review a few findings carefully. Read the excerpt in its wider context, check its baseline, and save an accepted or rejected judgment with reasons.
9. Export a record. Download the PDF or HTML report for reading, CSV for the coding sheet, and JSON for the fuller reproducibility record.

3. Frame a useful research question

Choose a question that links an observable feature to a context you can investigate. For example:

• How do repeated clause openings organize collective identity in this speech?
• How does marked punctuation structure the pacing of this short story?
• Which lexical contrasts distinguish a selected novel chapter from comparable prose?
• Does a recurring phrase occur across several works, or chiefly in one document?

Write down why each text belongs in the study, the edition or source you used, the period and genre, and any exclusions. Avoid selecting only passages that support your preferred answer. Keep exploratory observations distinct from hypotheses specified before analysis.

For a reusable source library and a selection record that travels into the report, assemble the study in Corpus Builder first.

4. Prepare and add texts

Paste or upload

The direct workspace accepts DOCX, searchable PDF, TXT, Markdown, CSV, and JSON. Save legacy DOC files as DOCX first. You can select multiple files. If you both paste a document and upload files in one submission, check the library afterward: the pasted document can succeed even if the file batch fails.

Record a meaningful title, author, year, genre, language, and corpus role. Metadata is part of your research record; it does not automatically make a reference comparable or verify an attribution.

Inspect extraction

Check the beginning and end of the saved text, chapter boundaries, paragraph breaks, punctuation, columns, and verse lineation. Headers, page numbers, hyphenated line endings, tables, or editorial notes can create misleading patterns. Word formatting, text boxes, footnotes, and page layout are not guaranteed to survive conversion.

Scanned PDFs: the direct Stylistics uploader does not automatically recognize image-only pages. Use Corpus Builder's OCR and review workflow, or upload a reviewed searchable PDF. Do not assume that an invisible text layer is accurate.

CSV and JSON batches

CSV requires a text column. Optional columns include title, author, year, genre, language, and role. Quote cells containing commas or line breaks. Use en or other for language and target or reference for role.

title,text,author,year,genre,language,role
Practice passage,"We will rebuild our homes. We will rebuild our schools.",Teaching example,2026,speech,en,target

JSON accepts an array of document objects, or an object with a documents array. Use valid JSON; a reference role is not a bibliographic reference.

{"documents": [
  {"title": "Practice passage", "text": "We will rebuild our homes.",
   "language": "en", "role": "target"}
]}

5. Choose targets and references deliberately

The question determines the comparison. Match—or explicitly justify differences in—genre, period, language, medium, topic, audience, and edition. A speech compared with fiction may reveal genre differences rather than an individual speaker's distinctive style.

• For a within-genre question: use texts from the same relevant genre and comparable communicative setting.
• For an author comparison: include multiple independent works where possible. A single passage cannot represent a writer's entire style.
• For a historical comparison: document period differences and avoid interpreting spelling conventions as stylistic innovation.
• For a single-text close reading: run without references, but restrict claims to local patterns and their stated internal baselines.

Keep targets and references separate. Do not use the same source, another edition of it, or a heavily overlapping excerpt as its own comparison. Equal corpus size is not mandatory because rates are normalized, but extreme size imbalance, sparse evidence, and unequal diversity still affect interpretation.

Record the number of original works as well as token counts. Chapters cut from one novel are not independent authors or independent works.

6. Run and manage a report

In Prepare an analysis, confirm the target and reference selections, give the report a useful title, and check Analysis language. This is a default; each document retains its recorded language. Choose Analyze selected texts.

Wait for the saved report rather than treating an input page or an error response as success. The report should display findings and report details. You can open saved reports from the workspace, review them, and export them. A saved report can also be opened using its report link.

Local stylistics analysis is separate from the main ADAR analysis pipeline and does not consume its analysis allowances. Importing an existing ADAR document brings available results into the study; it is not a fresh run of that pipeline. Owner manuscript generation is a separate provider-backed action.

7. Read the report in the right order

1. Check the study identity. Confirm titles, document counts, target/reference selection, language, and analyzed word counts.
2. Read methodology and warnings. Understand which rules, linguistic resources, and baselines were available.
3. Inspect the analytical narrative. Treat it as a synthesis of the measured findings and current reviews, not as an independent source of truth.
4. Open exact evidence. Check quotation, document, location, source offsets, contextual passage, baseline, and proposed interpretation.
5. Compare alternative explanations. Repetition may be a meaningful pattern, a genre convention, a quotation, a heading, or an extraction error.
6. Review and export. Your decisions should remain visible beside the machine-generated candidates.

A large number of findings can reflect text length, repetition, segmentation, or extraction artifacts. It is not a ranking of literary quality. Few findings do not mean that a text lacks stylistic interest.

8. Understand candidate findings

Keep phonological interpretation, broad semantic claims, author intention, and audience effects separate unless you have the additional evidence needed to support them. The software does not replace those investigations.

9. Understand corpus comparisons

Comparisons use pooled target and reference tokens. They answer questions about relative lexical occurrence in the selected material, not about the whole language or a population of writers.

Always inspect the actual concordance and original-work distribution. A contrast concentrated in one long document is different from a pattern recurring across several independent texts. Avoid reporting statistical significance as aesthetic significance.

10. Review and annotate evidence

Use the evidence filters to narrow by search text, category, document, or review status. Open each relevant candidate and read the source context, not just the highlighted words.

1. Check that the quotation and location match the source.
2. Check that the stated rule or baseline actually applies.
3. Choose Accept candidate, Reject candidate, or leave it Unreviewed.
4. Use the notes field for your rationale, ambiguity, extraction concerns, or an alternative explanation.
5. Use the interpretation field for your evidence-grounded reading, then choose Save review.

Acceptance records your adjudication; it does not turn an exploratory observation into a universally validated claim. Rejection is useful research data, especially when explaining false positives or genre conventions. Review changes regenerate the analytical narrative and can mark an existing manuscript as needing revalidation.

11. Use concordance to test your reading

Enter a literal word or phrase in the concordance search and choose Search corpus. Read occurrences with their surrounding context and note which documents supply them.

The search is a literal substring search, not a lemma, whole-word-only, regular-expression, grammatical-pattern, or semantic search. A short search string can occur inside a longer word. Results show up to 250 occurrences; a visible result list is not proof that you have inspected every occurrence in a large corpus.

Search variants explicitly when needed. For a claim about “build,” for example, consider relevant inflections and compounds separately and explain your search procedure. Concordance supports close reading; it does not automatically calculate a formal collocation association measure.

12. Connect an existing ADAR analysis

1. Run the relevant text through the main ADAR workflow.
2. In the same authenticated browser session, open Literary & Corpus Stylistics.
3. Choose Import current ADAR document.
4. Confirm the imported text and metadata, then run or rerun the stylistics study.

Import requires an unexpired analysis belonging to the current browser session. An exact-text match updates the existing document rather than creating a duplicate. A report made before the import must be rerun to include the new attachments.

Existing schemata, speech acts, pragmemes, practs, culturemes, and lexical-cohesion information are linked only when their quoted evidence overlaps a stylistic passage. A plain upload has no attached ADAR classifications; missing layers are not invented.

13. Understand internal validation

New reports run internal validation automatically. Validate with ADAR reruns the checks on a saved report, including an older report.

The audit checks source quotations and offsets, reproduces supported rule detections and baselines, examines attached ADAR provenance, and recomputes corpus quantities and comparison statistics. Missing framework layers are marked unattached. Failed checks remain visible and block new owner manuscript generation.

Researcher review and internal validation answer different questions. A source-consistent candidate may still deserve rejection as analytically unhelpful. An accepted interpretation must still have accurate source evidence and a defensible baseline.

14. Export and preserve your study

Use filenames that identify the study and revision. Keep source files, corpus-selection criteria, any exclusion record, review decisions, and the final report together. If the study came from Corpus Builder, the report's pinned corpus snapshot records the selected versions and provenance.

Download essential records before deleting reports or modifying source documents. Deletion confirmations are consequential, not merely cosmetic.

15. Owner-only manuscript studio

This section applies only to ADAR's designated owner with the required administrator role. Ordinary users and other administrators do not receive access simply by opening a report.

1. Open a saved report and resolve failed internal checks.
2. Accept at least one verified evidence candidate.
3. Supply a research topic, research question, and one bibliography entry per line. A target-journal field is optional.
4. Choose a target of 3,000–8,000 words and request generation.
5. Review the actual word count, evidence identifiers, citations, argument, and limitations. Edit and save the draft as needed.
6. Export DOCX, PDF, text, or HTML only after authorial review.

Generation sends measured results, accepted evidence, and supplied references to the configured manuscript provider. It is not a purely local action. Supplied bibliography entries and journal claims are unverified. Do not include confidential material unless its transmission is permitted.

The workflow uses the owner's supplied academic voice guidance for manuscript prose, not to rewrite quoted literary evidence. Short or incomplete outputs can be rejected rather than presented as complete. Review changes can require manuscript revalidation.

A draft is not a guarantee of indexing, journal acceptance, accurate references, scientific validity, or publication readiness. The author remains responsible for claims, disclosure, permissions, and journal requirements.

16. Languages, limits, and troubleshooting

Current workspace limits

• 30 documents; each document must contain 3–20,000 words.
• 100,000 total words in the workspace.
• 5 MB per file and 12 MB per upload request.
• Eight saved reports; concordance displays up to 250 occurrences.

These are the limits of this dedicated research workspace, not a description of the main ADAR speech-analysis route. For long works, Corpus Builder can retain larger originals and prepare bounded, source-preserving analysis parts. The selected study still has its own limits; no unlimited-size claim is implied.

Language coverage

Set each document's language correctly. English grammatical observations use an offline English tagger when available; its absence is disclosed. Other-language texts retain supported Unicode lexical and positional observations but do not receive English grammatical or semantic rules. Tokenization is not designed for unsegmented Chinese or Japanese. Cross-language lexical comparisons need particular caution.

17. Before you present or publish a study

18. Plain-language glossary

Baseline: the expectation against which an observation is compared; it may be local or based on a reference corpus.

Candidate: a detected observation that still requires contextual review.

Concordance: occurrences of a search string shown with surrounding text.

Foregrounding: interpretively significant prominence or deviation; not something established by a frequency count alone.

Keyness: a relative-frequency contrast against a specified comparison corpus, not an automatic measure of literary importance.

Provenance: the recorded origin and processing history of evidence.

Source offset: a recorded position in the analyzed text used to locate exact evidence.

Target / reference: the material being investigated / the selected comparison material.

Token / type: an occurrence of a word-like unit / a distinct unit under the tokenizer's rules.

Validation / review: automated internal consistency checks / the researcher's recorded assessment and interpretation.

Next: open the Stylistics workspace, or use the Corpus Builder guide to prepare a documented collection.