Methodology

1. Research scope and textual sources

The base texts are public-domain Project Gutenberg transcriptions. Corresponding GITenberg text versions were searched to confirm the occurrence of candidate forms and their whole-text frequencies. Five additional works were introduced when the project expanded from six to eleven: Our Mutual Friend, Barnaby Rudge, Hard Times, Little Dorrit and Sketches by Boz.

Titles are italicised in accordance with standard conventions for English-language books. The years indicate approximately when publication began; they are not controlled publication dates for quantitative analysis. As a collection of individual sketches, Sketches by Boz must eventually be stratified by sketch and scene before comparison with the ten novels.

The results may be affected by edition differences, historical spellings, subsequent modernisation, punctuation conversion, OCR or keyboard errors, and whether apostrophes are encoded as ' or ’. Such editorial variation must be examined before drawing diachronic conclusions.

2. Extracting orthographic candidates

This is a selectively curated, exploratory candidate list, developed with reference to characters’ speech contexts rather than by a comprehensive automatic out-of-vocabulary search alone. It includes apparent phonetic representation, regional vocabulary, non-standard elision, socially marked forms, and individual patterns of speech. Candidate strings were checked against the source texts. Because the original six-work list is inherited, coverage is not necessarily uniform across works.

The dataset’s category field contains provisional descriptions. The categories actually employed and the considerations relevant to them are outlined below.

Ordinary contractions such as don’t and I’m do not in themselves demonstrate regional dialect. Forms that are also standard vocabulary, such as pint, or very short strings such as id, cannot be interpreted out of context. Eye dialect, phonetic difference, colloquial lexis and historically variable spelling should remain analytically distinguishable.

3. Attributing forms to speakers

Each network edge represents a candidate spelling linked to a speaker label within a particular work. Preliminary attribution draws on explicit speech tags, surrounding narration, conversational turn-taking and known representations of an individual’s pronunciation. An edge is not a validated annotation of every occurrence of the form.

• sample-confirmed: at least some example contexts support the proposed speaker association; not all attestations have been assigned to that speaker.
• speaker review required: the proposed speaker–form association still needs contextual verification.
• Other speakers, episodic speakers, and similar labels: convenience groupings containing more than one individual, not unique persons.

The observation that a character uses a form and the number of times that form appears in the whole work answer different questions. Reliable attribution requires examination of each occurrence, explicit reporting constructions (such as said X), turn boundaries and embedded quotations.

4. Counting units and interpretation

whole_book_frequency records case-insensitive candidate-string matches across the entire electronic source text. It does not guarantee that matches occur in direct speech, are uttered by the linked speaker, or are dialectal. Narration, prefaces, reproduced documents, embedded quotations, repetition by other people and homographic forms may all be included.

Distinguish the following counting units:

• 555 work–form types: the sum of distinct candidate forms recorded separately in each work. A form found in two works contributes two units.
• 429 distinct forms: unique orthographic strings across all eleven works, following basic case and apostrophe normalisation.
• 729 candidate speaker–form links: proposed associations between speaker labels and spellings within each work; these are neither token frequencies nor conversational turns.
• Normalised frequency: only possible once speaker-specific speech spans and their denominators have been verified. The present data do not support reliable per-speaker frequencies per 10,000 spoken words.

A future speaker-normalised frequency could be calculated as 10,000 × speaker-specific form tokens / all tokens in the same speaker’s verified direct speech. Both the numerator and denominator must be drawn from comparable textual media, annotation layers and editions.

5. Speech, written discourse and quotation depth

Quotation marks do not necessarily identify direct spoken discourse. They may delimit letters, advertisements, read-aloud documents, documentary evidence, directly represented thought, or retrospectively reconstructed dialogue. Treating all quoted material as dialogue produces false positives. The outer narrator, an embedded storyteller and an inner conversational exchange may constitute distinct attribution levels.

1. Medium: spoken / written / represented thought / uncertain.
2. Presentation: direct quotation / indirect report / free indirect discourse.
3. Discourse function: conversational exchange, extended monologue addressed to a listener, soliloquy, quoted document, reading aloud, etc.
4. Attribution and embedding: speaker identifier, addressee, quotation depth, source and confidence of identification.

Future analysis should align speech-span boundaries and form offsets, avoid splitting a long monologue into fictitious turns, and independently assign another speaker’s direct utterance embedded within the current speaker’s words.

6. Comparative principles for eleven works

The aim is to compare Dickens’s literary representation of regional, class-associated and individual patterns of speech—not to treat the novels as audio records of historical dialect. Character function, volume of dialogue, chronology, fictional setting, genre, exaggerated pronunciation and social status may all shape the visible distribution.

7. Interpreting the network

The all-works view displays candidate work–form connections for spellings shared across texts; selecting one work displays a speaker-label–form bipartite graph. These are provisional associations. Display limits and filtering may hide nodes, so consult the exported CSV for the complete inventory.

The presence, thickness or geometric location of an edge is not a measure of the number of times a character speaks, distinctiveness of a spelling, relative frequency in dialogue, or distance between characters. Use the distinction between sample-confirmed and speaker review required.

8. Data, documentation and reproducibility

The following files are available for download alongside the website:

• dataset.json: book metadata, candidate speaker–form edges, and selected examples.
• data/book_summary.csv: eleven-work aggregate counts.
• data/dialect_character_links.csv: candidate associations, one edge per row.
• Research report: detailed observations for the five most recently added works.

A reproducible study should record the electronic edition and text hash (e.g. SHA-256), search expressions, Unicode normalisation, word-boundary rules, case handling, source offsets and the evidence for speaker attribution. The present release does not include full token-level positional annotations or a parallel alignment of critical editions.

9. Validation programme

1. Segment dialogue and other represented discourse throughout all eleven works, distinguishing nested quotation, letters and long monologues.
2. Record speaker IDs, chapter or sketch identifiers, token offsets, surrounding contexts and direct/indirect presentation mode.
3. Inspect concordances of every candidate form to exclude homographs, false matches, and forms that occur only in standard usage.
4. Test inter-annotator agreement and classification reliability using a written coding manual and independently coded samples.
5. Normalise by speaker-specific word counts, then control for work, chapter, occupation, region and genre.

Re-check individual quoted examples against the source before citation. This is an exploratory research resource, not an official publication of the University of Osaka.