1. Introduction
Diesner & Carley (2005: p. 83) employ the term network text analysis (NTA) to describe a wide variety of “computer supported solutions” that enable analysts to “extract networks of concepts” from texts and to discern the “meaning” represented or encoded therein. The key underlying assumption of such methods or solutions, they assert, is that the “language and knowledge” embodied in a text may be “modeled” as a network “of words and the relations between them” (ibid, emphasis added). A second important assumption is that the position of concepts within a text network provides insight into the meaning or prominent themes of the text as a whole.
Broadly considered, creating networks from texts has two basic steps: 1) the assignment of words and phrases to conceptual categories and 2) the assignment of links to pairs of concepts. Approaches to NTA differ with re- gard to how these steps are performed, as well as other dimensions like the level of automation or computer support, the linguistic unit of analysis (e.g. noun or verbs), and the degree and basis of concept generalization. That said, the four building blocks of text networks—concept, relationship, statement, and map—are the same. Specifically, a concept is “an ideational kernel” often represented by “a single word··· or phrase” while a rela- tionship is “a tie that links two concepts together”. A statement is “two concepts and the relationship between them” while a map is simply a “network formed from statements” (Carley & Palmquist, 1992: pp. 607-608, emphasis added). A simple example will help to clarify and illustrate these terms.
Figure 1, below, is adapted from Carley (1997) and it is typical of many network representations of texts. The network itself was constructed from the following two sentences: “Organizations use information systems to handle data. Information is processed by organizations who are interested in locating behavioral trends.” Several things about the network are noteworthy. First, observe that there are seven concepts depicted as nodes in the network, each of which appears only once. They are “organizations”, “information systems”, “process”, “infor- mation”, “interested”, “locating”, and “trends”. Secondly, see that there are also seven statements, i.e. pairs of concepts: 1) “information systems” and “process”, 2) “information systems” and “organizations”, 3) “process” and “information”, 4) “process” and “organizations”, 5) “interested” and “organizations”, 6) “interested” and “locating”, and 7) “locating” and “trends”. Third, note that the map itself is comprised of the network formed by all seven statements. Typically, the analyst must read some or all of the statements in a map in order to extract the meaning of the text as a whole. In this regard, it is then notable that the seven concepts are implicated in va- rying numbers of statements. Specifically, the concepts labeled “organization” and “process” are found in three statements while all other concepts are found in either two or one. Differences of this kind have important implications for the extraction of meaning from a text network, especially in larger networks. However, as the number of concepts and statements increases, so does the complexity of that task (Popping & Roberts, 1997).
The remainder of this paper is organized as follows. In the next section, I review the relevant literature on network text analysis with an eye toward those methods that rely most heavily on network analytical methods for the extraction of meaning. In the third section, I describe a novel method for the creation and analysis of text networks. In the fourth section, I apply the method to a sample of seven Academy Award-winning screenplays— Little Miss Sunshine, Juno, Milk, The Hurt Locker, The King’s Speech, Midnight in Paris, and Django Un- chained.
2. Literature Review
As noted above, network text analysis (NTA) involves the conversion of words into concepts, the creation of linkages between pairs of those concepts, and the subsequent analysis of the resulting text network. In her doc- toral thesis, Diesner (2012: pp. 90-91) compares 17 groups of methods for constructing networks of words along several dimensions, the most relevant four of which are―1) automation, i.e. “the ability to automatically collect one-mode and multi-mode network data”, 2) abstraction, i.e. whether and how terms are abstracted to “concepts or higher level aggregates”, 3) generalization, i.e. “the ability to identify new and unseen instances of entity classes and entity attributes”, and 4) “steps needed to reason about meaning of network data”. Among the ap- proaches she reviewed are several that are similar to the method described later in this paper. Specifically, these include hypertext (Trigg & Weiser, 1986); semantic webs (Berners-Lee, 2001; van Atteveldt, 2008); concept maps (Novak & Gowin, 1984), centering resonance analysis (Corman et al., 2002; McPhee, 2002), mental models
![]()
Figure 1. A simple text network (adapted from Carley, 1997).
(Collins & Loftus, 1975), mind maps (Buzan, 1974) semantic grammars (Roberts, 1997), network text analysis in the social sciences (Carley & Palmquist, 1992), semantic networks in communication sciences (Danowski, 1993), and event coding in political science (King & Lowe, 2003; Schrodt et al., 2008).
In Table 1 below, a subset of the methods reviewed by Diesner is summarized along four dimensions that are relevant to this study―1) selection, i.e. the basis upon which words are selected for or excluded from the sub- sequent analysis, 2) conceptualization, i.e. to which conceptual categories are the selected words mapped, 3) re- lationships, i.e. the basis upon which pairs of concepts are linked to one another and 4) the approach taken to the extraction of meaning. Four groups of methods are summarized therein―Centering Resonance Analysis (CRA), Network Text Analysis in the social sciences, Semantic Networks, and Semantic Mapping. In the second column of the table are described the various approaches to selection employed in the studies under examination. Three distinctions are worthy of note. The first of these concerns parts of speech. CRA, for example, relies exclusively upon nouns and noun phraseswhich they define as “a noun plus zero or more additional nouns and/or adjectives” (Corman et al., 2002: p. 174). All methods exclude determiners (a, an, the), as well as some or all of the follow- ing—prepositions, conjunctions, transitive verbs, verbs of being, acronyms, place names and proper nouns, and other words that “distort the description of the text” (Doerfel & Barnett, 1999: p. 592). As for pronouns, some approaches convert them to their corresponding nouns or noun phrases (e.g. Carley, 1997) while others simply eliminate them from consideration altogether (Doerfel & Barnett, 1999). A second distinction concerning unit of analysis is the importance of word frequency as a basis for inclusion. Notably, all three of the “semantic” me- thods select words based on their frequency of occurrence, excluding of course the “stop words” just described.
The third column in the table indicates whether and how the resulting selected words and phrases are aggre- gated into higher-order conceptual categories. Two methods that do not aggregate are CRA (Corman et al., 2002) and semantic mapping (Leydesdorff & Welbers, 2011). The former applies “minimal affix stemming” which changes only plural forms of words to the singular, e.g. drivers → driver. The semantic mapping approach does not employ any kind of stemming and as such, its word networks include multiple related variants of key words and concepts, e.g. both chemical and chemistry or all of the following—science, scientific, scientists, and the abbreviation Sci. The other methods aggregate to broader categories of concepts. Carley (1997) aggregated words through the use of a conceptual thesaurus based upon two generalizations—one contextual and the other grammatical. In the former condition, concepts with similar meanings were recoded into the same concept, e.g. information system and system both were coded as information system. In the second instance, concepts with different tenses and endings were stemmed, e.g. retrieved, retrieves, retrieving, etc. all become retrieve. The se- mantic network studies appear to use only the second approach.
![]()
Table 1. Selected methods for generating and analyzing text networks.
The fourth column of the table describes the nature of the relationship that links concepts. The most common of the approaches is co-occurrence of two or more concepts within some researcher-defined window. Typical of this approach is Carley (1997) who analyzed mental maps derived from answers to two open-ended questions by 41 undergraduate students enrolled in an information systems course. The questions were “What is an informa- tion system?” and “What leads to information systems success and failure?” After eliminating stop-words, maps were extracted using five windows—four, six, eight, ten, and 100 words. A linkage was established between pairs of words occurring within those windows. Leydesdorff & Welbers (2011) used a much larger window— their entire sample of the 195 documents—to establish co-occurrence and thus relationship. Neither Carley & Palmquist (1992) nor Corman et al. (2002) relied upon co-occurrence within windows of any kind. The former mapped relationships between concepts that were primarily causal or definitional in nature and had such attributes as strength, sign, direction, and meaning. For example, in response to questions such as “What is the purpose of research writing?” or “What steps would you follow in writing a research paper?” an answer such as “decide specifically what you are going to write-down” was coded as a relationship between the concepts “topic” and “writing”. CRA, on the other hand, focused on essentially hierarchical relationships derived from the se- quentially linking “noun phrases, which are potential centers in an utterance” (Corman et al., 2002).
As shown in the fifth column of Table 1, what all these methods have in common is their reliance upon a network analytic approach to extracting meaning from the textual data. More specifically, each uses a combina- tion of qualitative and quantitative techniques to accomplish that end. In the former instance this involves, at a minimum, reading some or all of the statements and perhaps relating them back to parts of the text. For example, Corman et al. (2002) found that “chains of concepts” in their network graph corresponded directly to “specific chains of nouns in specific sentences” in the text (p. 199) and that different parts of the graph captured the “dis- cursive division” between the two parties whose utterances comprised the map (p. 186). In contrast, quantitative analysis of the map usually involves counting the number of unique concepts and statements, especially when comparing maps associated with different individuals/groups (Carley, 1997). Enumeration is also key in deter- mining the most influential concept in a network (Leydesdorff & Welbers, 2011). Other quantitative methods for discerning meaning within or across text networks include correlational (Corman et al., 2002) and factor analy- sis (Leydesdorff & Welbers, 2011). It is worth noting that CRA stands out among these approaches as the only method that developed its own metric—resonance—for measuring the “mutual relevance of two texts” (Corman et al., 2002: p. 189).
Although the sample is relatively small, it is possible to generalize broadly about these approaches to NTA. Firstly, all methods, select some words for further analysis while excluding others from further consideration. Articles, prepositions, and pronouns are among the words most often excluded. The remaining words then are aggregated to higher order conceptual categories, typically grammatical or logical in nature. Next, pairs of con- cepts are then related to one another, most frequently on the basis of their co-occurrence within some window of words. Finally, an analysis of the structure or pattern of these relationships—through network analysis and/or statistical methods like factor analysis—are employed to extract meaning.
Taken together, the four dimensions outlined above—selection, conceptualization, relation, and extraction of meaning—establish criteria upon which any claim to novelty in the NTA field must be understood. Considered in turn they are as follows:
· Selection: A new method might be novel because it includes or excludes different (kinds of) words and/or uses different rationales for doing so.
· Conceptualization: A new method might be novel if it employs a different approach to aggregating from words to concepts.
· Relation: A new method might novel for the kind of relationship it uses to link concepts.
· Extraction: Finally, a method might be novel for the techniques and/or metrics it uses in the extraction of meaning from the network data.
The next section of this paper details a new method of network text analysis that meets these four conditions.
3. A Novel Method
As suggested in the introduction, at the heart of the method described herein is etymology or word history, a term defined in the Cambridge Advanced Learner’s Dictionary as “the study of the origin and history of words”. Over the past several decades, academic etymologists have traced thousands of words in a wide variety of Eu- ropean languages back to their Indo-European and other etymological roots. Among the most well-known con- temporary sources in this field are the American Heritage Dictionary of Indo-European Roots (Watkins, 2011); the Leiden Indo-European Etymological Dictionary series (Lubtosky, 2006-2013) which is a compilation of etymological databases using Sergei Starostin’s STARLING software technology; and the Indogermanisches Etymologisches Woerterbuch (Pokorny, 1959) which is an “updated and slimmed-down reworking” of Walde & Pokorny’s (1927-32) three-volume Vergleichendes Wörterbuch der Indogermanischen Sprachen (Wikipedia, 2013).
Several aspects of contemporary research on etymology are potentially useful for network text analysis. First, there is the obvious fact of the network-like relationship that underlies the field itself. By definition, from every etymological root descends or originates at least one word, otherwise it is not a root. That relationship is genitive, i.e. a relational case typically expressing source, possession, or partition. It is hierarchical and directed—from the root (parent) to word (descendant). The name used to relationship of pair of words descending from the same root is “paronym”. Figure 2, below depicts the Indo-European (IE) root saewel―which means “sun” and five words descending from it as indicated in the American Heritage Dictionary of Indo-European Roots (AHDIER) —south, solar, solstice, helium, and sun.
What is significant about Figure 2 is that it represents a bridge between the second and third stages of the NTA process—Conceptualization and Relation. More specifically, while conceptualization involves assigning words to their etymological roots, that relationship is not just categorical: there is also a network-like relation- ship—as shown above. But that relationship is actually not sufficient to construct the kind of network map like other methods of NTA. The issue is that the text networks are built on relationships between concepts, not be- tween words and concepts. The following example will explain how this hierarchical and directed relationship between words and higher-order concepts can be converted into one that links concepts.
Consider a text that contains the following nine words: the closed compounds sunlamp, manpower, sunlight, lamplight, and gentleman; the open compounds solar power and native tongue; the hyphenated compound self- possessed; and the proper noun Secretary General. As shown in Table 2, below, these words are all multi-mor- phemic compounds, each element of which descends from a different etymological root. And in that fact lies the solution to the linking of concepts. Recall that a statement in NTA is comprised of two concepts and the rela- tionship that links them. In Figure 3, below, each of these words appears on the link between the two etymologi- cal roots—the concepts—that co-occur within the word. Put another way, the relationship is the co-occurrence of two different etymological roots in the same multi-morphemic compound or multi-word expression—co-oc- currence in what is essentially a window of one word. For example, the etymological roots gene—(to give birth; beget) and man-1 (man) are linked by their co-occurrence in the compound word gentleman. Taken together, that word and those two roots comprise a statement. And as shown below, it is possible to construct an entire map or network from these interconnected statements. That network is comprised of eight concepts—namely, the Indo-European roots dnghu, gene, man1, s(w)e, poti, sawel, leuk, and lap—and nine statements which in- clude the words native tongue, gentleman, Secretary General, self-possessed, manpower, solar power, sunlight, sunlamp, and lamplight, as well as the pair of roots to which each element belongs.
![]()
Figure 2. An etymological root and its derivatives.
![]()
Table 2. Selected Indo-European roots and their derivatives.
Note: Source: American heritage dictionary of Indo-European roots.
![]()
Figure 3. A text network based on etymological relationships among selected words contained in Table 2.
Once we recognize that networks can be constructed in this manner, it becomes apparent that compounds words are not the only ones that can or should be included. For example, one can include abbreviations and acronyms, e.g. USA (United States of America), FBI (Federal Bureau of Investigation) and CNN (Cable News Network). Also relevant are portmanteau or blend words which are formed by joining the beginning of one word and the end of another. Examples of blend words include brunch which is a blend of breakfast and lunch and motel which is a blend of motor and hotel. A third category would include certain clipped words, which are de- fined in the Random House Dictionary as ones “formed by dropping one or more syllables from a longer word or phrase with no change in meaning”. Examples of relevant clipped words include internet which is a clipped form of internetwork and slo-mo which is a clipped form of slow-motion.
And though it may seem otherwise, these groups of words are no random assortment. Rather, they comprise a well-defined, inter-related set that is extensively-studied in the field of morphology. Specifically, they all belong to the branch of morphology known as word-formation, the study of creation of new words principally through changes in their form (Wisniewski, 2007). Morphologists described several kinds of word-formation (Stekauer, 2000; Plag, 2003, Lieber & Stekauer, 2009), all of which are hierarchically related to one another as shown in Figure 3, below (Plag, 2003: p. 17).
As the diagram suggests, the two major branches of morphology are word-formation (“the ways in which new complex words are built on the basis of other words of morphemes”) (Plag, 2003: p. 13) and inflection (the mod- ification of words to express a broad range of grammatical categories, e.g. tense, person, number, etc.) Word- formation can be further sub-divided along two lines—derivation (forming new words through changes in form) and compounding (the creation of novel words through the combination of existing ones). Derivation is also of two varieties—affixation (changes in form through the addition of affixes) and non-affixation (changes in form through means other than affixes). The former involves the addition of prefixes, suffixes, and infixes while the latter includes the creation of abbreviations, acronyms, blend words, truncated or clipped words, as well as the process conversion. Table 3, below, provides examples of each type of word-formation described in Figure 4. Word formation is, then, the basis for the selection of words in this method of NTA.
To summarize briefly, the method described above 1) filters words based on their morphological properties— only multi-morphemic compounds are analyzed, 2) maps them to higher-order concepts on the basis of etymol- ogy—elements of each compound are mapped back to their respective etymological root, and 3) relates those concepts on the basis of their co-occurrence within the same word. As far as we are aware, no other method of NTA performs these three basic steps in this way. However, there remains to be considered the question of the extraction of meaning from the overall network. In short, this method achieves that goal through an approach not unlike other methods—through reading and through enumeration of concepts and statements. There are, how- ever, important differences—differences detailed in the following section.
4. Methods and Data
In this Section I demonstrate the application of the method described in the previous section to a selection of Academy Award-winning screenplays, namely the seven that received the Oscar for Best Original Screenplay between 2006 and 2012. As shown in Table 4, below, the seven winners of the Academy Award for Best Origi- nal Screenplay in the years 2006-2012 were Little Miss Sunshine (Arndt, 2006), Juno (Busey-Mario, 2007), Milk (Black, 2008), The Hurt Locker (Boal, 2007), The King’s Speech (Seidler, 2009), Midnight in Paris (Allen, 2011), and Django Unchained (Tarantino, 2012). Electronic copies of these seven screenplays were found in
![]()
Table 3. Examples of seven types of “novel” word-forms in the sample.
![]()
Table 4. Seven academy-award winning, best original screenplays, 2006-2012.
Note: Source: Oscar.com, IMDb.com.
![]()
Figure 4. Tree diagram of morphology and its domains.
numerous places online and where more than one version existed, the latest publication date was selected. All but one of the screenplays was machine readable—Woody Allen’s Midnight in Paris. It was also the only of the seven to be categorized by the International Movie Database (IMDb, 2013) as comedy and not a drama or dra- ma-comedy. Notably, all seven screenplays were solo-authored and only one author is a woman—Brook Busey- Mario, the author of Juno. Three of the authors were nominated for the Academy Award for Best Original Screenplay either in years before or during the time periods under study. They are Mark Boal, who was also no- minated for Zero Dark Thirty in 2012; Quentin Tarantino, who was also nominated for The Inglorious Basterds in 2009 and Pulp Fiction in 1994; and Woody Allen who has received a total of 15 nominations. Finally, note that the page range varied substantially around the 111 page average with the longest Django Unchained (167 pages) being almost twice the length as Midnight in Paris (89 pages) and The King’s Speech (92 pages).
The first step in the text analysis was to generate a list of all relevant words and phrases for each screenplay. This was accomplished with the help of a software program called Automap 3.0.10 (Carley, 2001-2013). The process was as follows. First the screenplay was converted to a text file and uploaded to Automap. After remov- ing single letters, extra spaces, and spurious characters, four routines were run within Automap—Extract Hy- phenated Words, Suggest N-Grams, Identify Possible Acronyms and Concept List. As their names suggest, the first routine extracted all words containing hyphens, many of which turned out to be hyphenated compounds. The second routine selected from the screenplay all multi-word phrases and combinations that were not named entities but which could be open compounds, e.g. post office or parking lot. The third routine identified and ex- tracted all words that were capitalized. Several of these turned out to acronyms or abbreviations, e.g. CD (compact disk) and HUMVEE. The fourth routine used was Concept List (Per Text) which generated a list of all unique words in the text. Because Allen’s Midnight in Paris was not machine-readable, I read and coded the screenplay manually.
The output of each of the four routines was consolidated into a single list for each screenplay and then scanned by the author to identify all multi-morphemic compounds—described in Table 3—acronyms, abbrevia- tions, anacronyms, blend words, clipped words, compounds, and pseudo compounds. Excluded from considera- tion were all proper nouns (Green Zone, Hollywood), place and organization names (South Pole, Scotland Yard, Burger King), product names (Land Rover), holidays (New Year’s Eve, Christmas Eve), as well as any other word or phrase connoting a specific person, place, or thing through capitalization. Also eliminated were all ref- erences to screenplay and film jargon, e.g. ECU (extreme close-up), off-screen, VO (voice-over) and POV (point of view). In order for suggested n-grams to be included in the word list, the word combination had to appear in one of seven on-line dictionaries—the American Heritage Dictionary of the English Language, the Cambridge Advanced Learner’s Dictionary, the Cambridge Dictionary of American English, Collins English Dictionary, the Compact Oxford English Dictionary, the MacMillan Dictionary, or Merriam Webster’s Online Dictionary. Finally, multi-word exclamations and interjections such as good night, goodbye, OMG were also eliminated.
The components of each of the remaining multi-morphemic compounds were then mapped to their corresponding etymological roots, over 95% of which were roots found in the American Heritage Dictionary of Indo- European Roots (AHDIER). Next, all pairs of roots in a given screenplay were converted into a symmetrical matrix which was then uploaded into version 6.487 of the UCINet software program (Borgatti, Everett, & Free- man, 2002). Text networks were then generated using version 2.118 of the Net Draw software program embedded in UCINet. As shown in Table 5 below, there was considerable variation across the seven screenplays in terms of the number of unique concepts (etymological roots) and the number of statements contained in their text networks. In Figure 5 is depicted the entire text network for Dustin Black’s Milk. It was selected because among the seven screenplays in this sample, its text network is closest to the average in terms of the number of concepts (248 vs an average of 251), number of statements (252 vs and average 258), the number of concepts in the network’s main component (154 vs an average of 142), and the number of statements in the main component (176 vs an average of 176).
As noted previously, the extraction of meaning from text data is one of the primary concerns of NTA. As also mentioned, there is not a consensus about how this is to be done. One thing that all reviewed methods have in common, however, is reading statements, particularly those associated with the most central or influential nodes. In network analysis there are dozens of measures of centrality and influence, the majority of which are based ei- ther on the number of connections a node has and/or the position and related role it occupies within the network.
![]()
Table 5. Structural characteristics of seven academy award winning, best original screenplays.
Nodes colored in green represent concepts—etymological roots—whose constraint values are less than or equal to 0.25. Blue nodes have constraint values greater than 0.25. The unfilled or white nodes also have constraint scores above 0.25 and are notable for not being mutually reachable in rela- tion to the blue and green nodes (the network’s main component), as well as in relation to most other white node.
Figure 5. Text network map of Dustin Lance’s Milk.
The method adopted here is to rely on the position/role, namely a measure of structural holes known as con- straint. In short, nodes that have low constraint in a network are those that bridge or connect its otherwise dis- connected parts. In social network analysis this role is known as brokerage which, in management and organiza- tional studies has been linked to superior firm-level, business-unit, managerial, and individual performance (Burt, 2005; Cross & Cummings, 2004; Hunter, 2011; Hunter & Chinta, 2013).
Table 6, below, shows the statements associated with the least constrained nodes in the text network for the screenplay of Milk. As calculated in UCINet, constraint values range from 0 (unconstrained) to 1.0 (completely constrained). In this study “least constrained” was defined as values less than or equal to 0.25. There were 22 nodes or concepts with values in this range, just under 9% of the 248 concepts in screenplay’s entire text net- work. Notably, all of the most influential nodes are in the network’s main component, i.e. the largest grouping of nodes that are mutually reachable. Also noteworthy is that 21 of the 22 most central nodes are concepts traced to Indo-European roots found in the AHDIER.
The synopses provided by IMDb for the film Milk reads as follows: “The story of Harvey Milk, and his strug- gles as an American gay activist who fought for gay rights and became California’s first openly gay elected official.” Wikipedia describes Milk as “a 2008 American biographical film based on the life of gay rights activist and politician Harvey Milk, who was the first openly gay person to be elected to public office in California, as a member of the San Francisco Board of Supervisors”. The site’s summary of the film’s plot provides information on the difficult relationship that existed between Milk and a fellow supervisor, Dan White, who eventually shot and killed Milk in 1978:
After two unsuccessful political campaigns in 1973 and 1975 to become a city supervisor and a third in 1976 for the California State Assembly, Milk finally wins a seat on the San Francisco Board of Supervisors in 1977 for District 5. His victory makes him the first openly gay man to be voted into major public office in California and in the top three in the entire US. Milk subsequently meets fellow Supervisor Dan White, a Vietnam veteran and former police officer and firefighter. White, who is politically and socially conserva- tive, has a difficult relationship with Milk, and develops a growing resentment for Milk when he opposes projects that White proposes.
![]()
Table 6. Statements associated with the 22 least constrained nodes in the text network of the screenplay for Dustin Black’s Milk.
A comparison of the keywords in Table 6 with the information provided by the plot excerpt and synopses re- veals large areas of overlap. Specifically, we have descending from the IE root dhe (to set, put) are the words elected official and public office. From dheue (to close, finish, come full circle) descend the words showdown and town hall. From the root kei-1 (to lie; bed, couch; beloved, dear) descend city council, city hall, city-wide, civil rights, and civil war. The root kel-2 (to cover, conceal, save) gives rise to city hall and town hall. From the root man-1 (man) we have several words related to Dan White. Specifically there are policeman and fireman (White’s professions prior to entering politics), as well as manslaughter (the crime of which White was convicted for shooting Milk). From the root reg-1 (to move in a straight line; to direct in straight line, lead, rule) descend the words human rights, right wing, and civil rights. Finally, descending from the root wi (apart, in half) descend the words citywide, statewide, and nationwide.
Taken together these fourteen words—or better yet statements because they represent the relationship existing between two concepts/roots—neatly encapsulate the story. Notably, several of them appear in the film synopses provided by IMDb and Wikipedia. Specifically, the single-sentence synopsis from IMDb contains the phrases “gay rights” and “elected official” while Wikipedia’s contains “gay rights” and “public office”. Notably the phrase “gay rights” does not appear in Table 6. Even though it appears several times in the screenplay, the two- word combination was not found in any of the seven online dictionaries. If the list of sources had included Wi- kipedia or Wiktionary or Dictionary.com, it would have been included and, of course, been part of a very central statement.
The same approach was applied to the six other screenplays in the sample and largely similar results were ob- tained, as summarized in Table 7, below. Specifically, in four of the other six screenplays the statements asso- ciated with the least constrained nodes clearly overlap with the synopses provided in the right hand column. A typical example is The Hurt Locker, a film about the exploits of US Army bomb squad deployed to Iraq. Several of the statements associated with the least constrained nodes evoke war and explosives. These include army-is- sue, body armor, fireball, first aide, gunfire, gunshot, half-destroyed, headquarters, HUMVEE, IED (Improvised Explosive Device), machine gun, shell-shocked, state-of-the-art, suicide bomber, and USA, as well as a robot called TALON (Threat and Local Observation Notice).
![]()
Table 7. Selected statements associated with the least constrained nodes in six academy award-winning, best original screenplays.
Legend: IMDb = International Movie Database, http://www.imdb.com/; RT = Rotten Tomatoes, http://rottentomatoes.com.
Little Miss Sunshine, the winner in 2006, is about a dysfunctional family that sets out on an interstate road trip to get their youngest member, seven-year-old Olive, to the finals of the eponymous beauty pageant. Several of the multi-morphemic words associated with the least constrained nodes suggest the family’s vehicle (back seat, hatchback, seat belt, side door, SUV, and under-revving), the trip that they undertake together including the sev- eral stops along the way (convenience store, gas station, service station, newsstand, state trooper) and the road itself (US Highway, turn-off). Another large set of statements reference the beauty pageant itself and/or the par- ticipants (back-stage, butt-wagging, half-dancing, make-up, outfit, over-coiffed, over-dressed, runners-up, run- way, and upbeat).
Juno, the 2007 winner, a story about “offbeat” and “an independent-minded teenager confronting an un- planned pregnancy” (Wikipedia, 2013). Among the statements associated with the least constrained nodes are ones concerned with conception (bedroom, bedside, make-out), pregnancy and childbirth (birthing room, birth- day, knocked-up, waiting room), high school (backpack, blackboard, dropouts), recreational drug-use (drugstore, dime bag, burnout) and disaffected young people (teenage, teenwave, runaways), some of whom are not making the most productive use of their time and ability (burnout, dropouts, time suck).
The winner in 2012 for best original screenplay was Django Unchained, Quentin Tarantino’s Civil-War-era western about a freed slave’s efforts to rescue his wife from a brutal plantation owner. Among the statements associated with the least constrained nodes are several commonly associated with slavery and slaves (free man, boss man, overseer, bullwhip, bare chest, bare feet, runaway), as well as plantations and the work performed thereon (family unit, big house, farmhouse, blacksmith, barnyard, bunkhouse, livestock, stableboy). There are also several statements associated with Django’s cooperation with a bounty hunter who is pursuing dangerous fugitives from justice (courthouse, outlaws, track-down, peace officer), the means of transportation typical of the times (wagon wheels, stagecoach, horseback, country road), and, of course, the violent means to which each resorts to obtain their ends (bullet hole, gunfight, gunbelt, gunman, gun powder, hangman, lawman, marksman- ship, shotgun, showdown).
The two screenplays whose statements that do not correspond as well to the film’s synopses are The King’s Speech and Midnight in Paris. The problem with the latter is a lack of data. Recall this screenplay, as well as The King’s Speech were outliers regarding the number of statements in the network map, both having 194 where the average was 258. Midnight in Paris also has a highly fragments screenplay with only 14 of its 124 (11.2%) concepts being mutually reachable. By comparison, every other screenplay has at least 50% of its concepts mu- tually reachable. In fact, so fragmented is the text network for Midnight in Paris that none of its constraint val- ues are less than 0.25, the threshold set for assessing statements in this study.
As for The King’s Speech, there are a few statements that evoke the period in which the story takes place, i.e. the period after first World War and build-up to the second. These include goose-step (a reference to Nazi sol- diers), wartime, WWI (World War I), and perhaps naval officer. There are also a few statements that, when taken together, place the story in Great Britain or some part of the United Kingdom, e.g. tea time, prime minister, and pub which is clipped word for the open compound public house. But noticeably absent from the rather limited set of statements is anything to do with speech, language, public speaking, stage fright, or speech impediments and speech therapy. Nor is there anything that suggests that the story takes place in and among members of Bri- tain’s Royal Family. That said, if the rules used to include words had allowed for proper nouns, then the several of the following statements might have figured more prominently in the resulting text network—Whitehall (both a wide thoroughfare in London and the name for the British Civil Service), Englishman, Great Seal, Great War, Scotland Yard. But again, there are no terms among these related to speech and speaking. One conclusion that one might draw from this omission is that the screenplay’s author did not have a well-developed mental model of the subject. Specifically, his lack of use of either lexicon or terminology specific to the field and/or less morphologically complex words might suggest a commensurate lack of understanding. Alternatively, the author may have known but consciously decided not to use this language. Interestingly, Wikipedia provides some interesting information about the matter. It states that the author first···
··· read about George VI’s life after overcoming a stuttering condition he endured during his youth. He started writing about the relationship between the monarch and his therapist as early as the 1980s, but at the request of the King’s widow, Queen Elizabeth The Queen Mother, postponed work until her death in 2002. He later rewrote his screenplay for the stage to focus on the essential relationship between the two prota- gonists. Nine weeks before filming began, Logue’s notebooks were discovered and quotations from them were incorporated into the script.
What all this tells us is that the writing of the screenplay was carried out over a period of more than 20 years. The long period of inactivity was presumably due to unspecified sensitivities on the part of the King’s widow. We also learn that what began as a screenplay became a stage play and before becoming a screenplay once more. Finally, we learn that just nine weeks before filming began in the fall of 2009, new material concerning the work of the speech therapist Logue was rapidly incorporated into the script. These conditions are highly irregular and contradict much of the conventional wisdom concerning screenwriting. As such, it is not surprising that the re- sulting text network appears under-developed and highly fragmented. Nor is it surprising that terminology and jargon on speech therapy is absent and/or poorly integrated with other themes. That said, that screenplay’s net- work is not nearly as fragmented as that of another winner, Midnight in Paris, a screenplay whose numerous ec- centricities will have to await explanation in another time and place.
5. Discussion
The objective for this paper has been to describe and demonstrate a novel method of network text analysis. The basis for claiming novelty, it has been argued, rests on four pillars. The first of these concerns the unit of analy- sis, particularly the rationales for including or excluding the words that are to be further analyzed. In this method the only words that are included are multi-morphemic compounds, e.g. abbreviations, acronyms, blends or port- manteau, clipped words, and several varieties of compounds including pseudo-compounds. As was discussed at some length, this is by no means a random assortment. Rather, these categories represent the majority of the processes that morphologists refer to as “word-formation”, the creation of new or novel and complex word forms from existing words. To my knowledge, no existing method of NTA uses morphology as a primary or secondary basis for the selection process. And precisely because this criterion is so novel, it is possible to antici- pate an important objection: it allows for the inclusion of pronouns and prepositions and many other words typically excluded by almost every other form of text analysis, whether network or frequency-based. This justify- able concern is addressed in the discussion of the three remaining bases for claiming novelty—aggregation, rela- tionship or linkage, and the extraction of meaning.
As noted previously, after selecting words from among the many appearing in a text, the analyst typically as- signs or aggregates them to high-order categories. That process may be as simple as stemming words back to their base forms or may involve the use of synonyms or other conceptually-based categories and relationship. The method described in this paper relies upon etymology. Specifically, each element of the multi-morphemic compound is aggregated to its corresponding etymological root, most often an Indo-European (IE) root. For example, for the compound word firewood the modifier, fire, was assigned to the IE root paewr which means “fire” while the head of the compound, wood, was assigned to another IE root widhu which means “tree, wood”. The result of this aggregation process was the creation of a linkage between the two roots—paewr and widhu— due to the fact that they both co-occur within the same word, firewood. As far as I am aware, no existing method of network text analysis aggregates words to their etymological roots and/or creates linkages among concept based on the co-occurrence of the descendants of etymological roots, either within or between words.
The final and arguably most important stage in NTA involves the extraction of meaning from the text network. The most common approach to extraction involves reading some or all of the statements of which the entire network is comprised. Recall that statements are the two concepts or nodes and the relation between them. In the method outlined above, only a subset of statements were used to extract meaning—those associated with the least constrained nodes or concepts in the network. And as was shown in some detail, in five of seven screen- plays analyzed this small subset of statements mapped fairly well—in some cases very well—onto the major themes of the narrative. One import of this approach was that while prepositions or pronouns do exist in the network as statements, they do no distort the meaning or adversely affect the extraction process. In short, there is no compelling reason that statements containing prepositions or pronouns even be read, let alone treated as substantive. As such, these words and related statements are retained for both completeness’ sake and because of the possibility that they may function as bridges between otherwise disconnected segments of the overall network.
6. Conclusion
In light of the above, two distinct but related avenues of future research are immediately and compellingly evi- dent. The first concerns the abstraction of meaning. The second concerns the question of inter-text analysis. Concerning the former, recall that the extraction of meaning from the textual data involved the reading of a sub- set of the statements that collectively comprise the network. The selection of that subset was based on one net- work property—constraint. In particular, the statements used in extraction of meaning were limited only to those associated with nodes having constraint values less than 0.25. Although constraint is widely used in network analytic studies, there is no reason that it should or must be the only one employed at this stage. It is entirely possible that other measures of influence in networks, e.g. centrality, betweenness, closeness, etc., could provide similar or even better results. Future research should seek to determine which measures are better suited, if any, to the extraction of meaning from nodes in text networks like the ones generated here.
As for inter-text analysis, it’s important to know that the network approach outlined here can also be used to compare texts of different authors or the different texts by the same author. Such an application is not at all no- vel. Corman et al. (2002), for example, expressly intended for centering resonance analysis to be used in this way, i.e. to compare similarities and differences in themes present across texts. Carley (1997) compared the mental models of eight project teams, each with 4 - 6 members, enrolled in an information systems project course at a private university. Each team was required to “analyze a client’s need and then design and build an information system to meet that need within one semester.” Five of these teams were eventually deemed suc- cessful and three were not. At three points during the semester, each teams were required to provide responses to two open-ended questions—“What is an information system?” and “What leads to information system success or failure?” Their answers were coded and used as data. Carley reported that, on average, the “cognitive maps” of the members of successful groups had significantly more concepts and more statements compared to maps by members of non-successful groups. Future research should attempt to replicate and extend findings such as these, perhaps for example, using a sample of screenplays written by both experts and amateurs.
Before concluding this section, one broader point concerning network statements in these screenplays is in order. It concerns the matter of lexicon. One definition of that word is lexicon which is defined in the American Heritage Dictionary of the English Language (AHDEL) as “the morphemes of a language considered as a group.” But that is only one of the AHDEL’s three definitions of the term. The other two are “a dictionary” and the “stock of terms used in a particular profession, subject, or style; a vocabulary.” In light of the analysis above, the third definition is also relevant, and highly so. In his best-selling book on screenwriting entitled Story, Ro- bert McKee, argues that not only do films fall into widely-recognized genres, but audiences have “a complex set of expectations” for each one (McKee, 2010: p. 80). These expectations play out at multiple levels—primarily in terms of “subject, setting, role, event, and values”. In short, horror films must scare; westerns are expected to have outlaws and lawmen and men on horseback; wars films must depict battles and battlefields, death and dy- ing, bravery or heroism, and grace or cowardice under pressure; films about sports must depict them being played; love stories need lovers and heartbreak, make-up and break-up; epics need quests; legal thrillers need courtrooms, attorneys-at-law, evidentiary rules, prosecutors, judges, and opposing counsel; hospital dramas need doctors and nurses, EMTs and x-rays, vital signs and wheelchairs, and so on. Audiences know whether their ex- pectations are met or not based on what they see on the screen or hear coming out of the mouths of the actors. But before that can happen, screenwriters must, as McKee says, use words and only words to “put a film in the reader’s head” (p. 394). That, he says, requires words and language that are “vivid” and “descriptive”. And al- though he doesn’t say it directly, a lot of the advice and examples that McKee gives in this matter concern pre- cisely the kinds of words that this method selects and analyzes. Moreover, he emphasizes that the words used must be appropriate to the topic. It’s hard to imagine, for example, a movie about basketball that doesn’t use the words backboard, front court, half-time, back court, slam-dunk, jump shot, power forward, point guard, etc. Those words comprise some—but not all—of the conceptual vocabulary or lexicon of basketball. What McKee and other writers emphasize is that all film genres and sub-genres have such a lexicon and that to fulfill an au- dience’s genre expectations, able screenwriters are expected to know the genre’s lexicon. As an aim to informing screenwriting theory and practice, future research with this method should compare and contrast the lexicon of screenplays within and across film genres, and especially those screenplays that have achieved recognition and awards.
Finally, when compared with existing approaches to network text analysis (NTA) the method described in this study is novel, in part, because of its reliance upon morphology as a basis for filtering/selecting words. As men- tioned previously, a sizable proportion of the remaining words—and arguably the most substantive—is lexicon, bothin the morphemic sense of the word and the conceptual vocabulary or genre-specific sense, as well. The method is also novel, in part, because of its reliance upon both morphology and etymology as the bases for ag- gregating the words into broader conceptual categories and for creating linkages among them. As such, the most fitting acronym for this method is MENTAL, i.e. morpho-etymological network text analysis of lexicon.