Literature Fingerprinting:
`A New Method for Visual Literary Analysis
`Daniel A. Keim∗
`Daniela Oelke†
`University of Konstanz
`University of Konstanz
`
`ABSTRACT
`In computer-based literary analysis different types of features are
`used to characterize a text. Usually, only a single feature value or
`vector is calculated for the whole text. In this paper, we combine
`automatic literature analysis methods with an effective visualiza-
`tion technique to analyze the behavior of the feature values across
`the text. For an interactive visual analysis, we calculate a sequence
`of feature values per text and present them to the user as a character-
`istic fingerprint. The feature values may be calculated on different
`hierarchy levels, allowing the analysis to be done on different reso-
`lution levels. A case study shows several successful applications of
`our new method to known literature problems and demonstrates the
`advantage of our new visual literature fingerprinting.
`Keywords: Visual literature analysis, visual analytics, literature
`fingerprinting
`Index Terms:
`J.5 [Computer Applications]: Arts and
`Humanities—Linguistics, Literature; I.6.9 [Visualization]: Infor-
`mation Visualization—Visualization Techniques and Methodolo-
`gies
`
`1 INTRODUCTION
`Traditional literary analysis is mostly done without the use of a
`computer. One of the reasons is obvious: to properly understand
`a text not only the words that are used are important but also the
`context they are used in, and understanding them in the context is
`difficult to achieve algorithmically. However, there are some fields
`of literary analysis in which computers have already proved useful
`in the past. This includes the classification of texts and some aspects
`of literary criticism. Often these methods are based on features that
`are supposed to characterize the text. Feature extraction methods
`can be as simple as calculating the average sentence length or as
`complicated as estimating the vocabulary richness of a text. In the
`case of text classification often conventional classification methods
`such as Support Vector Machines or Bayesian Networks are used,
`that work fully automatically. In other applications, for example
`in the case of authorship attribution, more transparency is required.
`Then, nearest neighbor classification is a popular approach in which
`an unclassified document is attributed to the author with the most
`similar features given some reference documents with known au-
`thorship. For methods that are based on multidimensional feature
`vectors often a transformation to a low dimensional space is done
`(using PCA, SVD or Karhunen-Lo`eve transform) and the results
`are visualized in a two-dimensional scatterplot [3, 9, 12] or with
`barcharts [9]. All the approaches have in common, that a single
`feature vector or value is used to characterize the whole text. This
`means that a lot of information is disregarded, since the change of
`the values as the text proceeds can reveal characteristic traits of an
`author or show interesting patterns (see fig. 1 for a first impression).
`∗e-mail: keim@inf.uni-konstanz.de
`†e-mail: oelke@inf.uni-konstanz.de
`
`IEEE Symposium on Visual Analytics Science and Technology 2007
`October 30 - November 1, Sacramento, CA, USA
`978-1-4244-1659-2/07/$25.00 ©2007 IEEE
`
`Figure 1: Visualization of the two novels “The Iron Heel” and “Chil-
`dren of the Frost” by Jack London. Color is mapped to vocabulary
`richness.
`It can easily be seen that the structure of the two novels
`is very different with respect to this measure. This would be camou-
`flaged if only a single value for each book would be calculated.
`
`Our idea presented in this paper is to calculate the features for
`different hierarchy levels of the text (such as words, sentences,
`chapters, . . . ) and create a characteristic fingerprint of the text
`which contains significantly more information than just a single
`number and therefore enables the user to gain a deeper understand-
`ing. We successfully apply the method to known literature prob-
`lems (e.g. authorship attribution) and show that by combining the
`automatic analytical methods of literature science with an effec-
`tive visualization technique new insights into literary texts can be
`gained.
`
`Outline
`The rest of the paper is organized as follows: In section 2 the dif-
`ferent types of computer-based literary analysis are introduced and
`different variables for literary analysis are briefly reviewed. Using
`some of the variables, in section 3 we then test their discrimina-
`tion power with respect to the authorship attribution problem using
`novels of Mark Twain and Jack London. In section 4, we locally an-
`alyze the literature fingerprints of two novels of Twain and London
`and of a much bigger and more diverse text - the bible. Section 5
`introduces our framework and finally, section 6 concludes the paper
`and outlines some interesting future applications of our new tech-
`nique that are planned together with literature scientists.
`
`2 BASICS FROM LITERATURE SCIENCE
`2.1 Computer-based literary analysis
`In a number of recent digital library projects, huge amounts of
`literature have already been digitized.
`In order to be able to
`computationally support the analysis of these texts algorithms are
`needed that can cope with all levels of natural language, namely
`the lexical, syntactic, and semantic aspects. Although the field of
`natural language processing has made significant progress over
`the last years, there are still a number of aspects the algorithms
`cannot cope with properly. This is especially true if the semantics
`or meaning of a text has to be taken into account, because of the
`vast amount of words, which have different meanings in different
`contexts, and the impressive flexibility and complexity of natural
`language. Still, computers are of great help whenever the lexical or
`syntactic structure of a text needs to be analyzed. In this section,
`we introduce two fields of computer-based literary analysis, in
`
`115
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`EX1062
`Roku V. Media Chain
`U.S. Patent No. 9,715,581
`
`

`

`which computers have already been successfully applied, namely
`text classification and different types of literary criticism.
`
`types of text classification can be distin-
`Three different
`guished:
`topic-oriented classification, classification into genres,
`and authorship attribution. For topic-oriented classification often
`TF-IDF vectors are used to characterize a text. TF-IDF (term
`frequency - inverted document frequency) vectors are made up of
`the frequency of each term in the document corpus weighted by
`the importance of that term with respect to the other documents
`in the corpus.
`Intuitively, a term is seen as characteristic for a
`document if its frequency within the document is much higher than
`its frequency in the rest of the corpus [10]. A typical application
`of topic-oriented classification is, e.g., labeling newspaper articles
`as Politics, Business, Sports, or Entertainment. For discrimination
`between different literary genres, such as Fiction / Non-Fiction,
`Children’s literature, Travel literature, Poetry, etc.,
`it is useful
`to consider the grammatical structure, the parts of speech used,
`and even the layout of the text in addition to TF-IDF vectors. In
`contrast, a special requirement of authorship attribution is that
`the extracted features should not be consciously controllable by
`the writer to prevent the method from being misdirected by a
`forged text. Note that for all mentioned methods the quality of
`the classification highly depends on the suitability of the features
`for discriminating the objects of the given categories. Therefore,
`enabling the user to understand the discrimination power of the
`features with respect to the classification task is of high importance.
`
`Computer-assisted literary criticism is a rather young field in
`the studies of literature. According to [6], a frequently mentioned
`objection is that the words and sentences of a text cannot be
`analyzed without properly taking the context they are used in
`into account.
`Therefore, most researchers in literary studies
`only use computers to collect data that is afterwards analyzed
`conventionally. Yet, there are some cases in which the computer
`has already proven useful, e.g., for comparing an author’s revisions
`(from version to version) or for the analysis of prosody and poetic
`phonology. Computer-assisted studies have also been performed
`in the context of sequence analysis, such as assigning quoted
`passages to speakers and locating them in the sequence of the
`text [6]. Another interesting field for computer-assisted analysis
`is translation criticism, in which metrics, rhythm, style, and other
`variables of the original text and the translation are compared to
`evaluate the quality of the translation.
`
`All mentioned approaches have in common that one feature vec-
`tor or value is calculated per text or per text block. Even if the text is
`split into chapters and paragraphs, usually the value for each chap-
`ter or paragraph is considered as a single object. In most cases, the
`values are averaged over the whole text, which leads to a smoothing
`of passages with an unusual trend, camouflaging interesting pat-
`terns. None of the computer-based literature analysis methods used
`so far deals with the behavior of the values across the text, which
`means that this very important information is completely ignored.
`In our approach, we therefore consider the literature in more de-
`tail by analyzing the texts on different hierarchy levels (i.e. calcu-
`lating one value per sentence, paragraph, chapter, or text block).
`By visualizing the results of the detailed literature analysis together
`with their position in the text, even local analyses become possi-
`ble. Moreover, the comparison of the visualizations for different
`variables leads to insights about the discrimination power of the
`different literature analysis variables. Since the success of each of
`the methods highly depends on an appropriate choice of the feature
`analysis variables, the possibility to efficiently compare the effec-
`tiveness of the variables with respect to a specific task provides new
`ways of an in-depth literature analysis.
`
`2.2 Variables for literary analysis
`Different variables for literary analysis have been proposed. They
`can roughly be classified into three groups: statistical measures,
`vocabulary measures, and syntax measures.
`In [8], a compre-
`hensive survey on variables for literary analysis with a focus on
`authorship attribution can be found.
`Information about variables
`for text theme classification can be, for example, found in [7]. In
`this subsection, we briefly introduce some important text analysis
`measures to give the reader an overview of the field and provide
`the necessary background knowledge for the following sections.
`The focus will be on variables which measure the stylistic traits of
`literary texts in general and the style of an author in particular.
`
`Statistical measures
`Calculating the average word length or the average number of
`syllables per word are two simple variables to characterize a text.
`While the first one does not provide reliable results, the second
`one can be useful to distinguish different genres. This is intuitively
`plausible because in poetic texts the number of syllables of a word
`is much more important than in prose texts.
`Sentence length is an indicator of style that can be used to estimate
`how good the rhythm of a text is preserved in a translation of the
`text. It is also used for authorship attribution studies, although in
`the context of authorship attribution it can be problematic since the
`length of the sentences is consciously controllable by an author
`and is not meaningful if the text has been edited by someone else.
`It has been shown that the distribution of sentence length is a more
`reliable marker for authorship than the average sentence length.
`Yet, it is also more difficult to evaluate. Here our technique proves
`useful because the visualization of the results allows an effective
`comparison of the distribution.
`Instead of working on the words directly, it is also possible to
`analyze the proportions of certain parts of speech (POS) (such
`as nouns, verbs, adjectives . . .) in the text. By this, the degree of
`formality of a text can be measured or the style of a text can be
`compared to its translation in another language.
`
`Vocabulary measures
`Vocabulary measures are based on the assumption that authors
`(and their texts) differ from each other with respect to vocabulary
`richness (how many words are in the vocabulary of the author and
`is s/he able to use his/her vocabulary by applying new words as the
`text proceeds) and with respect to word usage (which words are
`preferred if several can be applied).
`
`To measure the characteristic word usage of an author the
`frequencies of specific words are counted. The success of this
`method highly depends on the appropriate choice of words for
`which the frequencies are compared. Different approaches have
`been suggested, e.g., to group the words into categories such as
`idiomatic expressions, scientific terminology, or formal words,
`and count the number of occurrences for each group or compare
`the frequency distributions of the words. Good results have been
`reported for function words such as ”the, and, to, of, in . . . ” as the
`set of words. According to [5], function words have the advantage
`that writers cannot avoid using them, which means that they can be
`found in every text and almost every sentence. Furthermore, they
`have little semantic meaning and are therefore among the words
`that are least dependent on context. With the exception of auxiliary
`words they are also not inflected, which simplifies counting them.
`Finally,
`the choice of specific function words is mainly done
`unconsciously which means that it is an interesting measure for
`authorship attribution.
`
`Measures of vocabulary richness are mainly based on the evalu-
`ation of the number of tokens and different types. In the following,
`
`116
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`

`

`used when there is doubt whether the person that claims to have
`written the text is really the creator. One example for such a doubt-
`ful situation is the assignment of the 15th book of the series of the
`Wizard of Oz. The book was published after the death of its au-
`thor L. Frank Baum and was said to have been only edited by his
`successor Ruth Thompson who wrote the next books of the series.
`However, some literature specialists think that Ruth Thompson also
`wrote the 15th book and that the attribution to Baum was only due
`to commercial motives to ease the transition from one author to the
`next without losing sales. See [5] for an interesting analysis on the
`problem.
`Authorship attribution has also been named stylometry, because
`the classification is based on the distinct stylistic traits of a docu-
`ment and is independent of its semantic meaning. To measure style,
`certain features of the text are extracted that clearly discriminate
`the literary work of one author from another author. Classical
`authorship attribution is mostly done on a pure statistical basis,
`excluding non-numeric measures. To get reliable results, enough
`texts of the potential writers with known authorship have to be
`available as basis for attributing the text in doubt to one of them.
`
`3.2 Case study with literature of Mark Twain and Jack
`London
`In this subsection, we will present the results of a study with
`literature of Mark Twain and Jack London. Our goal was to test the
`existing literature analysis measures and see whether our detailed
`visual representation leads to new insights.
`
`In our study we used the following texts, that are all publicly
`available from Project Gutenberg [1]:
`• Jack London:
`- The Call of the Wild
`- Children of the Frost
`- The Iron Heel
`- Jerry of the Islands
`- The Sea-Wolf
`- The Son of the Wolf.
`• Mark Twain:
`- A Connecticut Yankee in King Arthur’s Court
`- A Tramp Abroad
`- Chapters From My Autobiography
`- Following the Equator
`- The Adventures of Huckleberry Finn
`- The Innocents Abroad
`- Life on the Mississippi
`- The Prince and the Pauper
`- The Gilded Age: A Tale of Today
`- The Adventures of Tom Sawyer.
`We preprocessed the texts by removing the preamble and other
`Gutenberg specific parts of the document and by replacing short
`forms with the corresponding long forms (e.g. isn’t → is not). Af-
`terwards we used the Standford POS tagger to annotate the texts
`[2]. For that we had to remove the chapter titles, since the tagger
`is only able to cope with complete sentences (though it is fault-
`tolerant with some grammatical errors). Finally, we split the docu-
`ments into blocks with a fixed number of words each to be able to
`show the behavior of the variable values across the text. The num-
`ber of words per block can be chosen by the user. For this paper, we
`set the number of words per block to 10,000, but similar results are
`obtained for a wide variation of this number as long as the blocks
`are not too small (> 1,000); since some literature analysis measures
`will provide unstable results when applied to short texts. To obtain
`a continuous and split-point independent series of values, we over-
`lap the blocks with the neighboring blocks by about 9,000 words.
`
`let N denote the number of tokens (that is the number of word oc-
`currences which form the sample text, i.e. the text length), V the
`types (the number of lexical units which form the vocabulary in
`the sample, i.e. the number of different words), and Vr the number
`of lexical units that occur exactly r times. A simple measure for
`vocabulary richness is the type-token ratio (R) defined as
`
`.
`
`V N
`
`R =
`
`This measure has one severe disadvantage, namely its dependency
`on the length of the text. A more sophisticated method to mea-
`sure vocabulary richness is the Simpson’s Index (D) that calculates
`the probability that two arbitrarily chosen words will belong to the
`same type. D is calculated by dividing the total number of identical
`pairs by the number of all possible pairs:
`r=1 r(r− 1)Vr
`D = ∑∞
`N(N − 1)
`While the Simpson’s Index takes the complete frequency profile
`into account, there are also measures that focus on just one specific
`part of the profile. For example, [8] reports that Honor´e suggested
`a measure that tests the tendency of an author to choose between a
`word used previously or utilizing a new word instead, which can be
`calculated as
`
`.
`
`R =
`
`100logN
`1−V1/V
`and is based on the number of Hapax Legomena (V1) of a text, that
`means the number of words that exactly occur once. The method
`is said to be stable for texts with N > 1300. Similar to this, the
`Hapax Dislegomena (V2) (the words that occur exactly twice) can
`be used to characterize the style of an author. According to [8],
`Sichel found that the proportion of hapax dislegomena (V2/V ) is
`stable for a particular author for 1,000 < N < 400,000. At first this
`seems counterintuitively but with increasing text length not only
`more words appear twice but also words that formerly occurred
`twice now occur three times and therefore left the set of hapax
`dislegomena.
`Many other methods to measure the vocabulary richness exist. The
`interested reader should consult [8] for a deeper investigation of
`the topic.
`
`Syntax measures
`Syntax-based measures analyze the syntactical structure of the text
`and are based on the syntax tree of the sentences. As the syntactical
`structure contains additional information, syntax measures have
`a high potential in literature analysis and have already been used
`in some projects.
`In [4], an experiment is reported in which a
`new syntax-based approach was tested against some word-based
`methods and was shown to beat them. In another approach [11],
`the authors build up syntax trees and develop different methods
`to analyze the writing style,
`the syntax depth, and functional
`dependencies by evaluating the trees. Note that – to a certain
`extend – the usage of function words also takes the syntax into
`account, because some function words mark the beginning of
`subordinate clauses or connect main-clauses. They therefore allow
`inferences about
`the sentence structure without analyzing the
`syntax directly.
`
`3 AUTHORSHIP ATTRIBUTION
`3.1 The concept of authorship attribution
`The goal of authorship attribution is to determine the authorship of
`a text when it is unknown by whom the text has been written or
`when the authorship is disputed. Authorship attribution can also be
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`117
`
`

`

`(a) Function words (First Dimension after PCA)
`
`(b) Function words (Second Dimension after PCA)
`
`(c) Average sentence length
`
`(d) Simpson’s Index
`
`(e) Hapax Legomena
`
`(f) Hapax Dislegomena
`
`Figure 2: Fingerprints of books of Mark Twain and Jack London. Different measures for authorship attribution are tested. If a measure is able
`to discriminate between the two authors, the visualizations of the books that are written by the same author will equal each other more than
`the visualizations of books written by different authors. It can easily be seen that this is not true for every measure (e.g. Hapax Dislegomena).
`Furthermore, it is interesting to observe that the book Huckleberry Finn sticks out in a number of measures as if it is not written by Mark Twain.
`
`This results in a soft blending of the values instead of hard cuts and
`therefore enables the user to easily follow the development of the
`values across the text.
`As visual representation of the results we depict each text block as
`a colored square and line them up from left to right and top to bot-
`tom. Although very simple this is an effective visualization since
`the order of the text blocks is very important and the alignment cor-
`responds to the standard reading direction. We also experimented
`with other shapes such as rounded rectangles, squares with beveled
`
`borders and circles. However, it turned out that the perception of
`a trend is easiest when displayed on a closed area with no borders
`visible. For the comparison of discrete values the other shapes are
`more useful. If a hierarchy has been defined on the text (made up
`of chapters, pages of the book, paragraphs, etc. ), the pixels are vi-
`sually grouped according to that hierarchy. Thereby, the structure
`of the text can be visually perceived and patterns that discern one
`passage of the other become obvious.
`Since function word analysis is known as one of the most suc-
`
`118
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`

`

`cessful methods for discriminating the texts of different authors, we
`started our analysis with this measure. We took a list of 52 function
`words that was also used in [5]. For each text block, a feature vector
`was calculated by counting the frequency of each of the function
`words, resulting in a 52-dimensional feature vector. We then ap-
`plied principal component analysis (PCA) to the feature vectors to
`linearly transform the data to a new coordinate system in which the
`first dimension accounts for the largest variance, the second dimen-
`sion for the second largest variance and so on. Figure 2(a) shows
`the values of the first dimension. We use a bipolar, interactively
`adjustable colormap to map the values to color.
`If a measure is
`able to discriminate the two authors, the books of one author will
`be mainly in blue and the books of the other one will be mainly
`in red. It is obvious that this is not the case here. What sticks out
`immediately is Mark Twains The Adventures of Huckleberry Finn.
`This novel seems to differ more from all the other writings of Mark
`Twain than the writings of the two authors differ from each other.
`If we visualize the second dimension of the transformed function
`word vectors we can see that the books of the two authors now
`separate from each other (figure 2(b)) - again with the exception
`of Huckleberry Finn (and this time also the book The Adventures
`of Tom Sawyer) which we would rather attribute to London than to
`Twain if its authorship was unknown. To analyze the strange behav-
`ior of Huckleberry Finn, we tested other variables such as Sentence
`length, Simpson’s Index, the Hapax Legomena measure of Honor´e,
`and the Hapax Dislegoma ratio (see section 2.2 for an introduction
`of the variables). Figures 2(c) - 2(f) show the visualizations for
`the different measures. In fig. 2(e) Huckleberry Finn again clearly
`stands apart. The Simpson’s Index shown in fig. 2(d) would again
`mislead us to attribute the book to Jack London, whereas in 2(c) it
`nicely fits to all the other books of Mark Twain. Finally, the Hapax
`Dislegoma shown in 2(f) seems to have no discriminative power
`and is therefore not useful for the analysis. Taking all analysis mea-
`sures into account, it is clear that there is something special about
`Mark Twain’s The Adventures of Huckleberry Finn. The reasons
`for the exceptional behaviour cannot be answered by our analysis.
`The potential explanations range from language particularities such
`as the southern accent of the novel which may irritate some of the
`measures over the editing of the text in Project Gutenberg to the sur-
`prising speculation that a ghost writer was involved in the creating
`of the novel.
`On the more general side, the figures show that not every variable
`is able to discriminate between the books of Mark Twain and those
`of Jack London, and this is also true if the novel Huckleberry Finn
`is excluded from the study. In fig. 2(f) (Hapax Dislegomena), we
`do not see much of a difference between the texts at all. The state-
`ment of Sichel that the proportion of Hapax Dislegomena in a text
`is specific for an author [8] cannot be verified, at least for these two
`authors. Instead, the sentence length measure (see fig. 2(c)) allows
`a very nice discrimination between the two authors. Mark Twain’s
`books in average have longer sentences than Jack London’s books.
`Only one novel per writer, namely Jerry of the Islands of Jack Lon-
`don and The Adventures of Tom Sawyer of Mark Twain break ranks
`and may be attributed to the other author. The second PCA dimen-
`sion of the function word vector (fig. 2(b)) and the Simpson’s Index
`(fig. 2(d)) also provide very nice results. Based on the Simpson’s
`Index, we can observe a trend to a higher vocabulary richness (less
`repetition) in the writings of Mark Twain than in the books of Jack
`London.
`
`Figure 3: The figure shows the fingerprints of two novels that almost
`have the same average sentence length.
`In the detailed view, the
`different structure of the two novels is revealed. The inhomogeneity
`of the travelogue Following the Equator can be explained with the
`alternation of dialogs, narrative parts and quoted documents.
`
`4.1 Detail analysis of two novels
`In this subsection, we will analyze two books, whose average sen-
`tence length is about the same. The images in Figure 3 show the
`result of splitting the text into overlapping text blocks of 10,000
`words each (with an overlap of 9,000 words) and calculating the
`average sentence length block-wise. The visual fingerprints reveal
`that the structure of the two books is totally different despite their
`identical overall average values. While the average sentence length
`in Jerry of the Islands of Jack London does not differ much across
`the novel (and thus the total average value would be meaningful),
`there are significant variations in Following the Equator of Mark
`Twain. Following the Equator is a non-fiction travelogue that Mark
`Twain wrote as an account of his tour of the British Empire in 1895.
`In fig. 3, some passages stick out as they are in dark blue respec-
`tively dark red. Taking a closer look at the text reveals the reasons:
`The long stripe in dark blue in the first line, for example, repre-
`sents a passage, in which Mark Twain quotes the scientific text of
`a naturalist with rather complex and long sentences. On the other
`hand, in the dark red passages in the second and third line Mark
`Twain noted some conversations that he had during his travel with
`the short sentences of spoken language. The second dialog is di-
`rectly followed by the quotation of a written report about a murder.
`One would rather expect such a report as being characterized by
`long sentences. This is probably why Twain himself utters his sur-
`prise about the text in his book. He says:
`
`“It is a remarkable paper. For brevity, succinctness, and
`concentration, it is perhaps without its peer in the liter-
`ature of murder. There are no waste words in it; there
`is no obtrusion of matter not pertinent to the occasion,
`nor any departure from the dispassionate tone proper to
`a formal business statement.” [13]
`
`The dark blue area in the forth line is due to a historical report of
`the black death and an official report of the trail.
`
`4 DETAIL ANALYSIS OF LITERATURE FINGERPRINTS
`
`The task of authorship attribution is an interesting application of
`our visual literature fingerprint. To reveal their full power, in this
`section we will look at the visual fingerprints in more detail.
`
`4.2 Detail analysis of the bible
`In a second study, we analyzed the visual fingerprint of the
`bible.
`In this case, we used the existing hierarchy of the text to
`define the blocks. While every text has an inherent syntactical
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`119
`
`

`

`Figure 5: Visual Fingerprint of the Bible. More detailed view on the bible in which each pixel represents a single verse and verses are grouped
`to chapters. Color is again mapped to verse length. The detailed view reveals some interesting patterns that are camouflaged in the averaged
`version of fig. 4.
`
`Figure 6: The visualization of this paper helped to find the longest sentences and improve the readability by modifying them appropriately.
`
`120
`
`Authorized licensed use limited to: Georgia Institute of Technology. Downloaded on March 25, 2009 at 22:29 from IEEE Xplore. Restrictions apply.
`
`

`

`can discern details that were camouflaged before. In figure 5, each
`pixel represents a single verse. The verses are grouped into chap-
`ters and the chapters are grouped into books. Again, verse length
`is mapped to color. In the detailed view, we may be able to get ad-
`ditional new insights. Note, for example, that a lot of chapters in
`Job start with a short verse. In the book of Job, Job and his friends
`take turns at giving long monologues. As the chapter borders were
`drawn in between two speeches most of the chapters start with a
`short verse like “Then Job / Bildad / Zophar / Eliphaz answered
`and said”. Another interesting observation is the clear division of
`Nehemiah 10 into two parts. The first one, the one with the short
`verses, consists of a list of persons that signed a treaty whereas the
`second part is a historical report. In the coarse representation of fig.
`4, we were not able to discern that because the average value of the
`whole chapter is not much different from other chapters. Another
`interesting observation is the regular pattern of Numbers 7, which
`appears odd. Looking into the text, we find reports how each tribe
`offers its dedication for the tabernacle. Since the offerings of the
`tribes were all similar, almost the same text is used for every tribe
`and therefore the text is repeated twelve times.
`
`4.3 Analysis of this paper
`Since our visual fingerprinting can be applied to any text, we also
`applied it to a first draft of this paper. Figure 6 shows a visualiza-
`tion of the sentence length of this paper on a sentence-by-sentence
`level. While it was thought to be a fun experiment at first, the vi-
`sualization proved to be useful to improve the paper. We used the
`visualization, for example, to find the extraordinary long sentences
`and rechecked them for readability. Other measures help to find re-
`peated words or other redundancies, detect incomplete or complex
`syntactical structures, or determine homogeneity of the text flow
`which is especially relevant for multiple-author papers.
`
`5 THE FRAMEWORK
`To ease the analysis we set up a framework in which all the mea-
`sures are implemented and that allows us to explore a text inter-
`actively. Figure 7 shows a screenshot of the tool. As can be seen
`multiple texts can be loaded at a time and therefore compared to
`each other or - as in this case - one text can be displayed multi-
`ple times to compare different measures. Interaction with a defined
`hierarchy comes in several ways. First of all, th