A sentence such as 'She would not, surely, be expected to give way now', an example constructed for this article, shows the phenomenon in small. Its third person and past tense belong to a narrator. Its 'surely', its 'now' and its note of indignation belong to a character. This is free indirect discourse. It has been described as the characteristic technique of the nineteenth-century novel for rendering minds and voices (Pascal, 1977; Cohn, 1978), and arguments about irony, sympathy and narrative authority in Austen, Eliot, Dickens and Trollope regularly turn on whether a given sentence is the narrator's judgement or a character's.

Large-scale study of the technique is held back by a practical gap. Automatic recognition of direct speech is comparatively easy because quotation marks do most of the work. Free indirect discourse has no such marker, and systems that attempt to find it need hand-labelled examples for training and testing (Brunner, 2013). The most thorough annotated corpus of discourse presentation in English covers twentieth-century writing (Semino and Short, 2004). For Victorian fiction, such annotation as exists has been produced for the purposes of individual projects, with category systems that differ in detail, so that results cannot easily be pooled or compared.

This article proposes a tagging scheme and an accompanying workflow designed to be shared. The aim is modest: a small set of categories, defined closely enough that trained annotators in different places would be likely to apply them alike, and an explicit way of recording the cases where they would not. The scheme is presented as a design. It has not been applied to a body of texts for this article, and no agreement figures are reported.

Two strands of earlier work inform the design: the narratological debate about what free indirect discourse is, and the experience of projects that have annotated discourse presentation.

Victorian novels present particular difficulties. Their narrators are talkative and opinionated, given to general reflections in the present tense, so that an evaluative sentence may be the narrator's wisdom, the character's self-justification, or the first mimicking the second. Gunn (2004), writing on Austen, argued that free indirect discourse is best understood as the narrator's imitation of a character within the narrator's own discourse, a view that fits Eliot and Trollope well and that makes a sharp boundary hard to draw. Dickens uses the form for collective and satirical voices, the opinions of a dinner table or a government office, which have no single source. Bleak House is partly narrated in the present tense, which removes the backshifted tense on which many definitions rely. Dialect and long sentences with several embedded viewpoints add further complications. A scheme tuned to twentieth-century prose will not transfer without adjustment.

The scheme rests on one decision: the properties of a passage that annotators are asked to judge are separated into layers that can be decided, and disagreed about, independently. Figure 1 shows the four layers and their values.

A tagging scheme is only as good as the process around it. The workflow proposed here follows the iterative model usual in language annotation, in which a draft scheme is applied to a sample, revised in the light of the problems found, and applied again before large-scale work begins (Pustejovsky and Stubbs, 2012). A pilot should use chapters from at least three novelists with different narratorial habits, so that the guidelines are not shaped by one author's manner. The guidelines themselves should be a public, versioned document containing definitions, the decision procedure, worked examples of clear and ambiguous cases, and a record of each change.

Every chapter in the main phase should be annotated independently by two people who have read the whole novel, since attribution by idiom depends on knowing how characters speak. Agreement is then measured. Two questions have to be kept apart: whether annotators marked the same stretches of text, and whether they gave matching labels to the stretches both marked. Chance-corrected coefficients such as kappa (Cohen, 1960) address the second, and Krippendorff's alpha has been extended to the first, where units are not given in advance (Krippendorff, 2004). Artstein and Poesio (2008) review the choice of coefficient and its interpretation. The verbal benchmarks proposed by Landis and Koch (1977) are widely quoted, but they were offered as conventions and should not be treated as pass marks.

Disagreements are then adjudicated in discussion, and each is classified, following the distinctions drawn by Gius and Jacke (2017), as a slip, a defect in the guidelines or a real difference of interpretation. Slips are corrected, guideline defects lead to a revision, and interpretive differences are kept in the released data as ambiguous spans with both readings noted. The released resource should include the individual annotations as well as the adjudicated version, and should meet the FAIR principles of findable, accessible, interoperable and reusable data (Wilkinson et al., 2016). Because Victorian texts are out of copyright, open release of text and annotation together is straightforward.

This article describes a scheme and does not demonstrate it. Its central claim, that separating mode, form, attribution and certainty will yield acceptable agreement on clear cases while preserving the disputed ones, is a hypothesis. It would be tested by a pilot of the kind outlined above, reporting agreement for span boundaries and for each layer, and by a comparison with a single-label scheme applied to the same passages. It is possible that the certainty layer will itself be applied inconsistently, with one annotator's clear case being another's ambiguous one. If so, the layer would need firmer criteria, or would have to be derived from disagreement between annotators instead of being judged directly.

Other limits are built in. The four-way form distinction discards finer categories that some research questions need. The cue list is drawn from descriptions of standard literary English and may serve dialect speech poorly. The requirement that annotators read the entire novel makes the work slow and expensive, which will restrict the size of any corpus produced. And the scheme takes no side in the dispute between dual-voice and single-voice accounts, which means that it records where readers locate the form without explaining what the form is. That neutrality is deliberate, but theorists of either persuasion may find the resulting data less decisive than they would wish.

Free indirect discourse matters to the study of Victorian fiction because so much of the novels' irony and moral judgement is conducted through it, and it resists annotation for the same reason: the form works by leaving the reader to decide whose words are on the page. The scheme proposed here accepts that condition. It asks annotators for four separate judgements, gives them stated cues and an ordered procedure, records uncertainty as data, and is expressed in a standard encoding so that work done in one place can be used in another. Whether it delivers reliable and useful annotations is a question for a pilot study, which the workflow section is intended to make easy to conduct.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.