Classifying and Contextualizing Edits in Variants with Coleto: Three Versions of Andy Weir’s The Martian

1. Introduction

This paper began with the desire to combine close and distant reading to compare three versions (or variants) of Andy Weir’s science fiction novel, _The Martian _ (2011, 2014, 2016). We wished to formally describe the changes to the texts made by two professional edits, then interpret how these thousands of textual changes, some big but mostly small, alter readings of the texts. While a number of automatic collation and variant comparison tools exist, none of them easily provide information on the textual changes that we were most interested in when analyzing The Martian , including: (1) a visualization of edits across the course of the entire texts; (2) all of the edits aligned and isolated from their original textual contexts; and (3) automatic classification of these thousands of edits. As no available tool performs all of these functions, we created one, and now present Coleto, a Python suite which inputs two variant text files (in English, German, or French) and creates a number of outputs: a visualization of edits across the progression of the texts, which indicates sections of the texts that contain greater or fewer edits; a .tsv table containing all of the textual changes presented side-by-side and categorized by type of edit; and visualizations of frequencies of edit types.¹

Interested users, even those with minimal coding experience, may use Coleto to quickly gain a variety of information about their variant texts. Comparing variant texts has a long scholarly history, yet to perform genetic/textual criticism requires a great amount of manual work: scholars place the variants side by side and read them through closely, pen or highlighter in hand. Coleto can reduce the massive effort of manual comparison and provide faster and easier access to all of the edits in variant texts, classified and presented in easy-to-use formats. With Coleto, philologists and scholars of genetic/textual criticism may gain speed, new information and new precision, at worst, and new insights, at best.

Coleto’s primary methodological improvement to collation is our suggested classification of edits in variants. Collation tools have previously provided information on insertion, deletion, and (sometimes) transposition, but the study of some variant texts may be improved by additional rule-based classification: expansion major/minor, condensation major/minor, as well as common orthographic features including punctuation (e.g. commas), hyphenation (e.g. large-ish → largish ), whitespace (e.g. in to → into ), capitalization (e.g. earth → Earth ), and italics (e.g. again → again ). We also defined another type of edit, numbers, for edits common in The Martian , e.g. 4 → four , or abbreviations involving numbers, e.g. 80 km → 80 kilometers .

2. Use case: The Martian

The Martian by Andy Weir is a best-selling science fiction novel about an astronaut named Mark Watney who is stranded alone on Mars. Using science and ingenuity, Watney secures food, water, oxygen, and shelter, establishes communication with NASA back on Earth, and finally reaches a spaceship that enables his rescue. The text of The Martian exists in three complete variants. Weir originally self-published The Martian on his personal website in 2011 (hereafter, Martian1 or M1 ), then began selling it on Amazon.com in 2012, for 99 cents, where it sold 35,000 copies in three months ². A major New York publisher, Crown, no doubt impressed by this self-publishing success, bought the rights from Weir, edited the text, and re-released it in 2014 (hereafter, Martian2 or M2 ). Describing this edit from Martian1 to Martian2 in interviews, Weir has said, “The editing process was pretty smooth. It was not a lot of changes at all” ³, as well as, conversely, “…there were a lot of edits and changes… No significant plot changes, nothing like that, but a lot of the wording. It’s much more polished” ⁴. A third variant was published by Crown in 2016, The Martian: Classroom Edition , which removes the text’s extensive profanity to market the novel to educational audiences (hereafter, _Martian3 _ or M3 ). Explaining this decision, Weir stated, “I got a lot of emails from science teachers who said, ‘Man I’d love to use your book as a teaching aid, but there’s so much profanity in it that we can’t really do that’” ².

Our experiments first seek to formally describe the textual changes in the professional edits to The Martian , especially as Weir’s inconsistent public statements on the first edit make it unclear how extensive this editing process was. Figure 1 summarizes the textual variation between Martian1 , Martian2 , and Martian3, as performed by the statistics feature of Wdiff ⁵.⁶

Quantitatively, Wdiff’s collation statistics lends weight to Weir’s statement that there were “not a lot of changes at all” in the first edit to The Martian , as 92.4% of his self-published version is shared with the Crown Publishers edition. But 7.6% is not an insubstantial textual change, as changing even a few words in a literary text can potentially alter its interpretation. The basic collation provided by Wdiff thus leaves many questions about both edits to The Martian unanswered. Are there passages or sections of The Martian that received especially high amounts of edits? How many edits comprise the 6,319 words transformed to 9,205 words in the first edit, or the words changed in the second edit? Are the edits mostly minor corrections to e.g. punctuation, or more substantial changes such as added paragraphs? Are the edits long or short? How many of these edits can be automatically classified, and classified how? Did Martian3 , the “Classroom Edition” (intended to remove profanity for student readers) alter any text besides the profanity? And moving on to interpretation, how do the hundreds of minor textual changes, such as capitalization and the rendering of numbers and scientific abbreviations, alter readings of the texts? To begin to answer these, we apply Coleto to the texts.

3. Related Work

A number of tools exist to assist in comparison of textual variants. Wdiff compares text files on a word-per-word basis and provides statistics of insertions, deletions, and changes ⁵. Archival and philological websites allow parallel visual inspection of variants with annotations such as linguistic information or highlighting ⁷. The Versioning Machine is a long-running open-source tool that displays variant texts and manuscript images side-by-side, with a number of changes highlighted ⁸. Juxta and its web-based interface Juxta Commons provided user-friendly collation and visualization of changes in plain text files ⁹. Juxta Commons as a web service was shut down in September 2020, although its source code remains online.¹⁰ CollateX is advanced software for textual collation providing users with alignment tables as well as a visualization of variant changes as a word-level graph ¹¹. TRAViz provides sophisticated collation and aims to improve on CollateX’s visualization methodology ^{12 13}. Elisa Nury has introduced the PyCoviz tool to visualize collation based on CollateX ¹⁴. Other tools not specifically focused on variant comparison could also be useful for variant comparison tasks, for instance TRACER, a suite for detection of text re-use ¹⁵.

Different users and their research tasks will be drawn to one or more of these collation tools over the other. While existing collation suites generally approach collation visualization through side-by-side comparison of text chunks or sentence-level comparison, Coleto takes a different approach: to identify and separate all of the edits in variant texts, classified by rule-based methods and presented in output documents and visualizations. This assists our close and distant reading of The Martian , as we wished to interpret the cumulative effect of hundreds of minor textual changes. As digital humanities collation tools can and do build upon one another, Coleto’s open-source code and methodology for classifying edits can be incorporated into any other collation suite.

In addition to work in digital genetic criticism (e.g. ^{16 17}), previous studies which apply digital methods to the interpretation of variants of relatively recent fiction include Yufang Ho’s comparison of the 1966 and revised 1977 versions of John Fowles’s novel The Magus ¹⁸. Martin Paul Eve examined differences in the U.S. and U.K. editions of David Mitchell’s Cloud Atlas , including collation and comparison of a single character’s narrative, arguing that “textual variance [is] an element of Cloud Atlas that can and should be read ” ¹⁹, see also ²⁰), an assertion that we follow. In addition, Thomas Crombez and Edith Cassiers have analyzed Luc Perceval’s adaptation of Dostoyevsky’s The Brothers Karamazov into a theater play, based on extensive materials from the director’s Dropbox account and using Neil Fraser’s “diff_match_patch” library to great effect ²¹.

4. Advancing the classification of edits in variant texts

Existing schemas to classify edits in variant texts have so far been limited. The established methodology for classifying edits in scholarly editing is based on a distinction between accidentals and substantives, as defined by the widely-used Greg-Bowers tradition, which is also included, for instance, in the MLA Committee on Scholarly Editions’ Guidelines for Editors of Scholarly Editions ²².²³ Importantly for digital texts, there is no widely-applicable or widely-followed typology of edits in digital scholarly editing and collation, with different materials calling for different typologies ²⁴. We therefore suggest the following categories of edits in variant texts, as integrated in Coleto (Table 1).
Coleto’s classification schema for types of edits in variants of text. Script-Identifiable Edits Semantically Open Edits Numbers , e.g. 4 → four , or abbreviations involving numbers, e.g. 80 km → 80 kilometers Hyphenation , e.g. large-ish → largish Punctuation , e.g. commas ** Whitespace** , e.g. in to → into ** Capitalization** , e.g. earth → Earth Italics , e.g. again → again ** Insertion** ** Deletion** ** Expansion (Minor)** Expansion (Major) Condensation (Minor) ** Condensation (Major)**
Script-identifiable edits are those that can be confidently classified into types based on formal linguistic features alone; these include changes related to capitalization, whitespace, hyphenation, spelling of numbers, and changes to scientific abbreviations. For the rest of the edits that cannot be easily defined in a rule-based way, we classify them based on the nature and extent of the edit, under the proposed term, Semantically Open Edits: the well-established categories in existing collation tools of deletion and insertion, but supplemented by expansion and condensation categories. For the latter two, a distinction between small and large edits is beneficial, but the choice is (perhaps inevitably) an arbitrary one. For the following experiments, we defined major edits as having a Levenshtein distance greater than 5,²⁵ but users of Coleto may select their own preferred Levenshtein value in the configuration file.

The Gothenburg model of textual collation is described as Tokenization → Normalization → Alignment → Analysis/Feedback → Visualization (Birnbaum and Spadini 2020).²⁶ Coleto’s primary contribution to general collation methodology is to the Analysis step of the Gothenburg model (Figure 2).

5. Applying Coleto to The Martian

As its target audience is humanities researchers, Coleto is designed for ease of use, so users need only download the code from GitHub, place two variant texts in .txt format in the Data folder, choose configurations in the config.yaml file, and run the run_coleto Python script. Coleto then performs the following tasks: (1) Preprocessing of texts: splits texts into one sentence per line to prepare for alignment; (2) Align texts using Wdiff; (3) Perform analysis: rule-based detection and classification of edits; (4) Generate overview table; and (5) Generate statistics and visualizations.

Coleto also generates a number of outputs, the first being a data table in the .tsv format of all the edits in the variants, aligned in one convenient place (Table 2).
Sample of table generated by Coleto comparing edits from Martian1 to Martian2 . itemid version1 version2 category main-type lev-dist lev-dist-class lendiff-chars 9298-1 495 494 script-identifiable numbers 1 minor 0 9299-1 5 five script-identifiable numbers 4 minor 3 9299-2 Tau Event dust storm other tbc 9 major 1 9307-1 Adicalia other deletion 8 major -8 9308-1 3000m 3000 meters script-identifiable numbers 6 major 6 9308-2 500m 500 meters script-identifiable numbers 6 major 6
Coleto also creates three visualizations. A bar chart visualizes frequencies of Semantically Open Edits, or edits as classified by the statistics feature of Wdiff (insertion and deletion) alongside our suggested Levenshtein-based edit schema for Wdiff’s changes, which we subdivide into Expansion Major/Minor and Condensation Major/Minor (Figure 3).

This first visualization provides researchers with a bird’s-eye view of changes between two variant texts. And if a researcher wishes to examine, for instance, all of the Expansion Major edits, these are tagged in the .tsv table generated by Coleto. The next visualization is of our suggested, more narrowly defined, script-identifiable edits. We have so far implemented: numbers, punctuation, capitalization, hyphenation, whitespace, and italics (Figure 4).

Figure 4 displays that 2,226 script-identifiable minor textual edits were detected in the editing process from _M1 _ to M2 , which provide evidence for a commentator to reconcile author Andy Weir’s conflicting statements on the first edit: “It was not a lot of changes at all,” and “…there were a lot of edits and changes… No significant plot changes, nothing like that, but a lot of the wording.” While changes to hyphenation, capitalization, etc. may be of interest to some research tasks, they are undoubtedly of little interest to others, and the assignment of edits into finer classifications assists researchers in isolating the specific edits they wish to examine.

The third visualization generated by Coleto renders all of the edits across the progression of the texts (Figure 5). Script-identifiable and “other edits” are visualized as different series, with the value of each series calculated by the sum of the absolute Levenshtein distances for each sentence of the text, with Savitzky-Golay smoothing applied.²⁷

This visualization assists the researcher in obtaining a sense of how edits are distributed across a text, with notable spikes highlighting passages or sections which feature especially high amounts of edits, which may warrant closer inspection.

6. Investigating The Martian with Coleto

The data and visualizations generated by Coleto may assist in a variety of formal, hermeneutic, and mixed-method studies. In this section, we return to our original goal of investigating textual issues in variants of The Martian , which can also illustrate Coleto’s benefits for researchers.

6.1 Quantifying and interpreting reductions in profanity

Profanity is a key stylistic feature of _ The Martian_ , as exemplified by the novel’s opening lines in M1 and M2 : “I’m pretty much fucked. That’s my considered opinion. Fucked.” Profanity may be the most important stylistic feature in the textual history of The Martian , as the raison d’être for an entire variant, M3 , the “Classroom Edition,” was to remove the novel’s copious profanity for educational audiences. Profanity is also a touchstone for issues of censorship and commercialization in discourse around the film adaptation of The Martian . While the film only explicitly includes two verbal instances of fuck , the offending word is also written on screen and alluded to a number of other times, as Jacob Brogan writes: “[ fuck is] often suggested in the way that sex might be in a different sort of all-ages movie.” ²⁸. Brogan reads The Martian film’s limited use of fuck as a metafictional device, the film “pointedly nodding to its own limitations [… and] making reference to one of the most bizarre guidelines of the MPAA’s PG-13 rating, the principle that a film can include only one, and in some very rare cases two, (non-sexual) uses of the word.” Brogan also suggests Watney’s frequent use of fuck is emblematic of the character’s psychology: “Indeed, Watney’s access to the one thing he’s not supposed to say — and his willingness to keep saying it — indexes his indomitability.” As another impression, The Martian foregrounds the use of conversational American English to render almost all of its narrative, a tone underscored by its frequent profanity. An improved formal understanding of profanity in The Martian texts could thus support a variety of research questions.

While locating profanity in texts is a matter of simple query, Coleto’s generated tables bring additional speed and convenience to comparing the profanity in M2 and M3 (Table 3), while the generated visualizations provide additional context (Figure 6).

Note that the y-axis is much smaller here than in Figure 5, as the edit from _Martian2 _ to Martian3 (removing and replacing profanity) was much smaller than the edit from Martian1 to Martian2 . The large spike around sentence 3600 in Figure 6 demonstrates how Coleto highlights areas of large edits which scholars may wish to investigate further. In this case, however, the explanation is mundane: small textual variation surrounding a particularly long string of code that NASA sends to Watney on Mars.
Table 3. Selection of table generated by Coleto comparing edits from Martian2 to Martian3 . itemid version1 version2 category main-type lev-dist lev-dist-class lendiff-chars 1-1 fucked. screwed. other tbc 5 minor 1 3-1 Fucked. other deletion 7 major -7 4-1 Screwed. other insertion 8 major 8 5-1 two months month other condensation 5 minor -5 48-1 lets let other tbc 1 minor -1 56-1 shit crap other tbc 4 minor 0 95-1 fucking other deletion 7 major -7 185-1 fucked. screwed. other tbc 5 minor 1 241-1 piss. pee. other tbc 3 minor -1 280-1 damned stupid other tbc 5 minor 0 309-1 botanist, damn it. botanist! other condensation 10 major -9
Manually reviewing the separated edits in Coleto’s generated .tsv table reveals that over 290 of the edits to Martian3 remove or soften profanity, depending on how profanity is defined,²⁹ in ways presumably acceptable to school boards: fuck → crap , _ass _ → butt , piss → pee , shit → crap , etc. As some examples beyond corporeal profanity, Jesus Christ is changed to a secular Wow, Thank God is removed, and a line involving the epithet gay is also altered:

[08:31] JPL: Good, keep us posted on any mechanical or electronic problems. By the way, the name of the probe we’re sending is Iris. Named after the Greek goddess who traveled the heavens with the speed of wind. She’s also the goddess of rainbows.

[08:47] WATNEY: Gay probe coming to save me. Got it.
(³⁰ and ³¹)

In Martian3 , “Gay probe” is altered to “Pride parade probe.” These changes to Martian3 were previously noted by Susan Ohanian (2016), who criticized the spirit of the edit and noted inconsistencies, such as another instance of _Jesus Christ _ that remains in the text ³².

None of the bowdlerization in Martian3 is surprising, as the express purpose of the text was to remove or reduce profanity and other potentially offensive content. But we wished to also know whether Martian3 altered any text besides potentially offensive content — perhaps Weir or his editors used the occasion of a new edition to make some final corrections? — and the table generated by Coleto can quickly answer this question, as well. By manual inspection and sifting of the dataframe, it emerges that Martian3 also makes factual corrections to scientific details, as well as occasional replacement of continuous verbs and other minor stylistic features (Table 4).
Selection of edits from Martian2 to Martian3 , including manually added comments. M2 M3 Comment Comment2 Levenshtein Automatic Classification eighteen twenty-two science? 9 most part science? 3 sixty seconds five minutes science? 10 thirty-two-minute twenty-two-minute science? 3 two months month science? 5 condensation grumbled, crossing grumbled and crossed passive 7 said, rubbing rubbed passive 9 condensation
By separating and classifying edits, Coleto can also serve as an exploratory research tool by assisting in the spotting of non-obvious trends. For instance, manually inspecting the data table of changes from the earlier Martian1 to Martian2 , we stumbled upon the discovery that profanity was cut in this first edit, as well. This was a surprise, given the large frequency and foregroundedness of profanity in Martian2 . It emerges that in the first edit of The Martian , from self-published novel to major publishing house release, fuck and shit (and their word forms) were substantially reduced, by about 33% and 15%, respectively. Numerous other words and phrases were also softened with lesser profanity, more varied profanity, or non-profanity (e.g. the shit hits the fan → all hell breaks loose) in the first edit.

The Martian thus has the dubious distinction of having been bowdlerized three times: in M2 , which reduced profanity significantly, the thorough scrubbing of profanity in M3 , the “Classroom Edition,” and finally in the film adaptation, which abided by a (reported) film board’s code of two fuck s allowed, but winked at the audience through additional implied profanity. This textual history should inform readings of profanity in any variant or adaptation of The Martian , and perhaps any discussion of the protagonist Mark Watney’s nonconformist characterization.

6.2 Interpreting changes to numbers and abbreviations

When comparing variants manually, or even with existing collation tools, it can be difficult to gain a sense of the frequency and distribution of edits, especially very minor stylistic edits. For instance, in the first edit of The Martian , it is readily apparent that numerous minor stylistic/orthographic changes have been made to the way scientific abbreviations and numbers are written, such as L → liter and 8 → eight, but existing variant analysis tools provide no simple way to quantify these and gain a sense of this type of edit’s distribution.

From the script-identifiable edits of The Martian , Coleto classifies 868 edits as the rendering of numbers and scientific values in the first edit, from M1 to M2 . While individual instances of such numerical changes may seem interpretively insignificant, hundreds of them scattered across the text have a cumulative effect upon readings of the text. In protagonist Mark Watney’s narration in Martian1 , the use of cardinal numbers and short scientific abbreviations support the fiction that Watney is a scientist working in dangerous conditions, too focused on surviving the life-threatening conditions on Mars to bother writing kilometers or forty-one out in full. Plausibility is a central concern of science fiction ³³, and when such abbreviations and numerals are expanded in Martian2 , this may increase readability, but weakens the stylistic realism of Watney’s voice.

In Martian1 , however, Weir did not only write L for liter or 8 for eight in the voice of astronaut Mark Watney, but also in the third-person narration of the NASA scientists back on Earth, which has a jarring effect on readability (Table 5):
Selection of passages containing edits classified as numbers by Coleto. Martian1 Martian2 Watney’s first-person narration in personal journal style That extra 18kwh of storage will be tough. I’ll have to take 2 of the Hab’s 9kwh fuel cells and load them on to the rover or trailer. ³⁰ That extra 18 kilowatt-hours of storage will be tough. I’ll have to take two of the Hab’s 9-kilowatt-hour fuel cells and load them onto the rover or trailer. ³¹ Character dialogue in the context of third-person narration (NASA scientists) What’s the biggest gap in coverage we have on Watney right now? Um, Mindy said. Once every 41 hours, we’ll have a 17 minute gap. The orbits work out that way. ³⁰ What’s the biggest gap in coverage we have on Watney right now? Um, Mindy said. Once every forty-one hours, we’ll have a seventeen-minute gap. The orbits work out that way. ³¹

6.3 Formal identification of fixed scientific details

In addition to plausibility, scientific rigor is another central concern of science fiction ³³, especially the subgenre of hard science fiction, and Weir pays careful attention to the accuracy of scientific and mathematical explanations throughout The Martian . Weir went as far as to code a computer program that plotted the trajectory of the fictional Hermes spacecraft to ensure scientific accuracy in The Martian . Weir’s dedication to scientific accuracy was so great that a NASA scientist later reviewed Weir’s trajectory calculations, and concluded that Weir had gotten them right ³⁴.

In the process of writing and publishing Martian1 online, readers occasionally pointed out corrections to mathematics and science presented in the novel, which Weir then incorporated ³⁵. While locating these scientific corrections in the variants of The Martian would certainly be possible manually, it would be a time-consuming task of scanning the text for single-character changes to numerical digits. The side-by-side data table generated by Coleto, however, greatly speeds up this manual search (Table 6).
Examples of scientific corrections in the edit from Martian1 to Martian2 , as identified by Coleto’s numbers edit classification. Martian1 Martian2 The answer is a cool 1000 The answer is about 1100 The atmosphere is 98% CO2 The atmosphere is 95 percent CO2 It’s a closed system. It’s a closed system. Okay, technically I’m lying. The plants aren’t entirely water-neutral. They strip the hydrogen from some of it (releasing the oxygen) and use it to make the complex hydrocarbons that are the plant itself.

6.4 Interpreting the altered epilogue

The most substantial spike in Coleto’s visualization of edits between M1 and M2 in Figure 5, above, occurs at the very end of the novel: the removal of the 263-word epilogue at the end of Martian1 and a 255-word addition at the end of Martian2 , which substantially alters the novel’s closure. In both M1 and M2 , Watney, finally safely aboard the rescue spacecraft, is happy to be alive and states, “This is the happiest day of my life.” In Martian1 , however, a brief epilogue follows this. Some time after his rescue in space, Watney, now safely back on Earth, is employed by NASA to train the crew of an upcoming space mission. Watney sits upon a bench, killing time, when a young boy recognizes him as the famous astronaut. Watney tolerates the boy’s attention, until the boy poses a question:

So Mr. Watney, the boy said, If you could go to Mars again, like, if there was another mission and they wanted you to go, would you go?

Watney scowled at him. You out of your fucking mind?

Ok time to go, the mom said, quickly herding the boy away. They receded in to [sic] the crowded sidewalk.

Watney snorted in their direction. Then he closed his eyes and felt the sun on his face. It was a nice, boring afternoon.
³⁰

This is a typical epilogue in fiction, as defined by the Russian Formalist Boris Eikhenbaum: it sets the perspective by a shift in time and provides some sort of after-history of the major characters ³⁶. This epilogue of M1 may be read to evince the lack of internal growth by Watney, who remains emotionally disengaged and contemptuous of playing the role of hero. Watney’s harrowing adventures on Mars have not altered his nonconformist characterization, a nonconformity encapsulated throughout the text by — and on the final page of the novel, reiterated by — the most foregrounded stylistic feature in the novel: profanity. The epilogue of M1 also shuts the door on potential sequels, which could be read as authorial and textual nonconformity with commercial expectations for science fiction novels.

Martian2 , however, cuts this pessimistic epilogue and inserts a substantial amount of optimistic text just before the novel’s end. In M2 ’s new ending, Watney expresses gracious appreciation for all the parties involved in his rescue — his crewmates, NASA, the Chinese space program, the billions of people who hoped for his survival — and affirms a widespread faith in human nature:

The cost for my survival must have been hundreds of millions of dollars. All to save one dorky botanist. Why bother?

Well, okay. I know the answer to that. Part of it might be what I represent: progress, science, and the interplanetary future we’ve dreamed of for centuries. But really, they did it because every human being has a basic instinct to help each other out. It might not seem that way sometimes, but it’s true.
³¹

This greatly alters the tone of the novel’s ending, revises Watney’s growth and characterization, and now leaves open the possibility of sequels. The correlation between reduced profanity in M2 and Watney’s character is notable: as the text itself becomes more conformist, so does its protagonist.

7. Conclusion and future work

A traditional scholar who sat down to compare the first two variants of The Martian would possibly, but possibly not, have noticed the reduced profanity, although the changed epilogue is impossible to miss. If such a scholar were sufficiently rigorous and willing to invest the necessary time, they would undoubtedly locate the corrected science and note the changes to abbreviations, numerals, and other minor stylistic aspects, as well. Coleto, like any collation tool, does not seek to replace such diligent genetic criticism, but to supplement it with speedier information and new precision, as well as provide a method that scales to longer texts which may challenge the time and patience of scholars performing manual comparison of variant texts.

Coleto could naturally be adapted per research topic, and may hopefully be useful for critical interest in, inter alia, variants of fiction, the role of the editor, and text re-use. Some tantalizing use cases for comparing variants of fiction come from the rapidly emerging global phenomenon of digital self-published best-sellers,³⁷ in which, like The Martian , texts often exist in a self-published variant as well as a later professionally-edited variant.

A significant weakness of Coleto is its inability to detect transpositions, as it relies on Wdiff for collation. For instance, Wdiff does not detect if a sentence is removed, but then replaced a few pages later in the novel — Coleto would classify this as an unrelated deletion and insertion. While Coleto could be integrated with transposition-detecting software in the future, for now, a pragmatic research suggestion would be to use Coleto in tandem with CollateX ¹¹, as described e.g. in ³⁸. Comparing Martian1 and Martian2 in CollateX results in 98 transpositions detected. Of these, 10 involve punctuation and should be considered artefacts of the method. Another 33 represent transpositions of a single word, showing stylistic preferences on the word-order level. Yet another 21 concern multi-word expressions which change the overall construction of a sentence or paragraph more substantially. Finally, 34 concern segments of more than 5 words, typically in the range of a short phrase to several sentences. In all cases, the scope of movement is rather small, with transpositions remaining very localized phenomena. This inspection of transpositions in this test case shows that, quantitatively and above all qualitatively, transpositions were not a major part of the first edit to The Martian , but naturally this step might play a greater role in variants of other texts.

For further work on Coleto, we plan to pursue the quantitative investigation into the semantics of edits by using word embedding models, thereby quantifying the amount of actual semantic distance between two tokens in an edit, rather than the surface-level Levenshtein distance. Some notable challenges we expect are defining edits concerning spelling mistakes (where misspelled words may not be in a model’s vocabulary) and edits involving different numbers of words (where it is not obvious which words establish semantic distance).

Acknowledgements

Thanks to Martin Paul Eve, the audience of our talk at Digital Humanities 2017 in Montreal, and the anonymous reviewers of this journal for excellent feedback on various versions of this project.