charities connected to lotteries
charities connected to lotteries
at
te
nt
io
n
at
te
nt
io
n
a
t
ct
t
ra
iv
e
a
t
ct
t
ra
iv
e
n
y
m
o
en
y
m
o
e
What’s next, a hyperlink or text?
The influence of hypertext writing on the writing process
What’s next, a hyperlink or text?
The influence of hypertext writing on the writing process
Master’s thesis of Melanie Jolyn Hof for the research master programme Linguistics
of the University of Groningen
Supervisors: dr. V.M. Baaijen (University of Groningen) and dr. M.A.H. Braaksma (University of Amsterdam)
Date of submission: October 28, 2013
Contact: Melanie Hof
CONTENTS
1. INTRODUCTION 4
1.1. Learning possibilities during writing 5
1.2. Development in writing 6
1.3. The value of hypertext writing 8
1.4. Processes in hypertext writing: this study 10
2. METHOD 15
2.1. Participants 15
2.2. Materials and procedure 15
2.3. Data preparation 16
2.4. Data analysis 20
3. RESULTS 27
3.1. The text production process of eleventh grade students 27
3.2. The effect of hypertext writing on the text production process 30
3.3. The distribution of activities over the writing process 32
3.4. The general character of the writing process 38
4. DISCUSSION 42 4.1. Conclusions 43 4.2. Methodological discussion 45 4.3. Theoretical discussion 46 BIBLIOGRAPHY 48 APPENDICES 52
Appendix A. Overview of the lesson series 52
Appendix B. The assignment in argumentative text writing 54
1. INTRODUCTION
Writing about a topic can be a valuable tool to promote the understanding of a topic, when it is accurately employed (Klein, 2000; Klein, 2004; Klein & Rose, 2010). According to Bereiter & Scardamalia (1987), an increase in understanding can hardly be expected, when a writer immediately transcribes every association that pops up in his mind while writing a text about a particular subject. The writer needs to actively reflect on the topical information that he has at his disposal, based on concerns as: Which information should be provided to convince a reader of my opinion? In which order can this information best be presented to reach that aim? And which additional information should be given to underpin the arguments for my opinion?
The physical shape of the texts that students generally write can impede such reflections considerably (Braaksma, Rijlaarsdam, & Janssen, 2007; Coirier, Andriessen, & Chanquoy, 1999). The texts require linear presentation of information instead of hierarchical. Whereas a hierarchical information structure usually results from thorough information reflections with the text goal in mind, a linear structure is most easily obtained by transcribing a string of succeeding associations.
Hypertexts, on the other hand, are characterized by a hierarchical presentation of information (Bromme & Stahl, 2002a; Lehrer, 1993). The information is chunked in separate units, which ought to be logically related. Usually, the starting point is a home page with the most prominent information, and one or more hyperlinks to pages with additional information. These pages can be linked together as well as to other pages, which contain supportive material for the information discussed at the page. Each embedded page can again have embedded pages, which can have other embedded pages. Finally, a tree-like text is created with many intertwined branches.
1.1. Learning possibilities during writing
During writing, three main processes can be distinguished: planning, text production, and revision (Hayes, 1996). While planning, a writer applies general reflective processes to determine which goal he wants to achieve with his text, and how he can achieve that. The resulting plans are the input for the text production process, in which the abstract ideas are translated and subsequently transcribed in running text with the help of lexical and grammatical knowledge. The abstract ideas, translated text, and even transcribed text can constantly be changed by revision.
Bereiter & Scardamalia (1987) expect that the acquisition of content knowledge primarily occurs during the planning process, since writers then explicitly reflect on their topic knowledge in order to determine what information should be incorporated, and how that should be done. During text production and revision, contemplation of topic knowledge can be triggered, but the subsequent contemplation is done in a planning process (Hayes & Nash, 1996; Sharples, 1996). For example, the first content plans are generally rather abstract, and usually elaborated while producing the sentences that have to fulfill that plan (Kaufer, Hayes, & Flower, 1986). Assuming like Hayes (1996) that text production is just the encoding of meanings into language, a planning process should be activated in case a potential meaning elaboration seems necessary. While producing text, writers clearly reflect on just produced text in order to check whether the text they produce at that moment fits the context (Kaufer et al., 1986). When it does not, a revision of the form as well as of the meaning can be considered necessary. For the latter, again a planning process needs to be activated.
Whereas the text produced so far can trigger some content reflection during text production, Bereiter & Scardamalia (1987) consider rhetorical constraints the most important triggers for content reflection, and then especially before text production starts. The genre a writer intends to write usually imposes these constraints (Klein & Kirkpatrick, 2010). As Klein (1999) states: “a genre is distinguished by a rhetorical intention, expressed through discourse elements that form particular relationships with one another” (p. 230). In order to come up with the content for the different elements, and be able to establish the required relationships, the writer often has to draw new relations between the information about the topic that he has at his disposal. This could lead to new insights in the matter at hand, and subsequently the transformation of knowledge.
The result of a content planning process based on rhetorical constraints usually is an overarching plan of which content has to be used for the different genre elements, in which order the elements have to be presented, and in which order the content within the different elements has to be presented (Bereiter & Scardamalia, 1987; Hayes & Nash, 1996; Kaufer et al., 1986; Walvoord et al., 1995). Generally, such plans are generated in early stages of the text production, as reflected by the inclination of many expert writers to produce a writing plan, before text production is initiated (Kaufer et al., 1986; Walvoord et al., 1995).
Writers tend to pause most often and longest at paragraph boundaries, successively followed by sentence, clause, constituent, and word boundaries (Schilperoord, 1996; Spelman Miller, 2000). Pauses at word, constituent, clause, and – to some extent – sentence boundaries generally seem to be a consequence of formulation and transcription problems (Kaufer et al., 1986; Schilperoord, 2001b).
Based on the pausing behavior of professional writers, Kaufer et al. (1986) and Schilperoord (2001b) concluded that professional writers translate an idea into language clause by clause. According to Chenoweth and Hayes (2003), the translation of ideas into language is more subdivided, when a writer has less linguistic knowledge and working memory resources. They theorized that a proposer creates and sends an idea package to the translator, which transforms it into language. They assumed that “the proposer will not send, and the translator will not accept, larger packages than the translator can typically handle” (p. 114). Consequently, the more linguistic knowledge and working memory capacity a writer has, the larger the idea package that the translator receives to convert into language, and the larger the language string he produces. The transcriber subsequently writes down that language string. When the transcription of such language burst ends in a pause, “the production process appears to run out of steam” (Hayes, 2009, p. 4). All planned language is written down. The translator needs to receive and translate a new information package.
The translation process can be interrupted by the evaluator/reviser, which checks the correctness of the translated package. The frequency with which the reviser interrupts the translation process also depends on a writer’s linguistic proficiency and working memory capacity. When one finds the words to communicate an idea relatively easy, one can better apply grammatical knowledge when formulating the language strings, and consequently does not need the reviser to do that (Chenoweth & Hayes, 2001). In the beginning of the translation process, the retrieval of words and the application of grammatical rules is most easy, since a writer does not require memory space to remember text that is already translated, but not yet transcribed. Consequently, the chance that the reviser interrupts the translation process increases, when the translated language bursts get longer.
1.2. Development in writing
Harris, 2000; Kellogg, 2008). According to Kellogg (2008), at least twenty years of experience have passed before a writer is able to do that.
It already takes four to six years to automate the formulation and transcription processes to such an extent that a writer can devote some cognitive capacity to planning and revision activities (McCutchen, 2006). To generate content, inexperienced writers fall back on a strategy of knowledge-telling, as Bereiter and Scardamalia (1987) defined it. They generate one or two ideas, translate them in one or two sentences, and take that as starting point for a new generation process, in which again just one or two ideas are generated. The writing process is usually paused rather long at sentence boundaries, to be able to devote memory resources to content generation (Wengelin & Strömqvist, 2004; Wengelin, 2006). The associative manner of content generation does not lead to the generation of ideas that are completely off-topic. Unskilled writers seem to have sufficient working memory sources to keep the text topic in mind. Based on the analysis of a corpus of 270 essays, Hayes (2011) concluded that more than 90% of the students in Grade 1 to 5 are able to write essays of which the whole content is related to the main topic of the text. Less than 10% of the students shifts to another topic, while writing a text.
When the burden reduces that text production activities place on the working memory, writers become able to keep subtopics in mind too. After Grade 5, they start to elaborate on subtopics that become activated while writing about the main topic (Hayes, 2011). However, as usual for knowledge tellers, these elaborations are immediately generated and transcribed at the moment the subtopics are activated. The transcription is not guided by a structural plan. They actually do not even generate one. Reflection at the beginning of text production still does not reach any further than the first sentence: a pause before the first sentence is equally long as a pause for any other sentence (Strömqvist, Nordqvist, & Wengelin, 2004). Presumably, they are still incapable to produce a text plan. When Coirier, Favart, & Chanquoy (2002) provided seventh graders with ideas that needed to be clustered, they were at best able to relate two ideas to each other. Differences in importance between the ideas were, moreover, hardly recognized.
In most writing process studies, ninth grade students are the eldest high-school students under scrutiny. Generally, they are compared to university students, who are often considered proficient writers. Since the results of Strömqvist et al.’s study (2004) reveal that ninth graders are far from equal to university students, we may expect that their proficiency still increases a lot during the last years of high school. The research of Coirier et al. (2002) into the structuring capacities of writers shows that these capacities especially improve between grade 9 and 11. The eleventh graders were as good as the university students in thematically clustering provided ideas, and ordering them in order of importance. Moreover, they could correctly apply many linguistic devices that can be used to show the reader how ideas are related, but far from all. The university students knew better which relations could be indicated with the distinctive kinds of punctuation markers. In addition, university students tend to produce more elaborate sentences, with several dependent clauses. By comparing the texts of 17-year-old and university students, Johansson (2009) concluded that 17-year-old students write texts that have the same length as the texts of university students, but a lower lexical density.
Regarding the results of the text analyses of Coirier et al. (2002) and Johansson (2009), eleventh grade students seem to lack at least the linguistic proficiency of a proficient writer. In Johansson’s (2009) study, additional support for this assumption can be found in the writing process of the 17-year-old students. The 17-year-old students needed more time to write the same amount of text as the university students. One reason for that was that they deleted more text. According to Johansson (2009), “the vast majority of the deletions are due to typos or other spelling errors, which are (almost immediately) detected and corrected.” (p. 130), since there existed a high positive correlation between the number of key presses and the number of characters in the final text. Because this correlation was even high for the university students, Johansson (2009) expects more revisions at word and phrase level, when students have more writing time.
Literary competent eleventh grade writers indeed seem able to do more higher-level revisions. Groenendijk, Janssen, Rijlaarsdam, & Van den Bergh (2008) studied the poem writing process of strong and weak readers of literature. They discovered that most weak readers produced their texts in a very linear manner, whereas several of the strong ones produced their text non-linearly. The linearly produced texts were of low quality, and some of the non-linearly produced texts were of high quality. The latter is presumably especially a consequence of the moment revisions elsewhere in the text were executed. In the middle phase of the writing process, low-level revisions led to quality improvements. In the final phase, sentence level revisions led to quality improvements. When revisions were done at the leading edge of text production, the revisions always influenced the text quality negatively.
1.3. The value of hypertext writing
indicating the structural relationships between ideas. Being unable to accurately communicate a generally hierarchical information structure to the reader, eleventh graders tend to fall back on a knowledge-telling writing strategy, when writing a linear text. Braaksma et al. (2007) suggest that this tendency can be prevented, by first teaching “students to rework their ideas in a writing plan”, and subsequently showing them “how they can sequence ideas in a linear fashion, and how they can use textual signs to guide readers in their process of reconstruction and regeneration, when the text can be converted into a hierarchical network of ideas” (p. 102).
Since the production of a writing plan seems a necessary step in hypertext writing, it is expected that writing these texts can learn students to better develop an information structure before text production is started (Braaksma et al., 2007). After all, the production of a good hypertext starts with a reflection on which separate information units can be distinguished, and how these can and should be related (Bromme & Stahl, 2002b, Lehrer, 1993). When hypertext writers have access to an overview map of their hypertext, these relations can easily be checked and changed during the hypertext production. The overview map is a graphical representation of all hypertext pages and the relations between them. When the hypertext environment does not contain an overview map, the writing process considerably benefits from a predefined structure.
Especially the drawing of relations is expected to be a rather time-consuming process, in which many content reflections are triggered. A hypertext reader does not necessarily read the text in the order the author intended, because several reading routes are possible (Bromme & Stahl, 2002b; Carter, 2003). A reader can decide not to read the content of a node, when it seems to contain information that he is not interested in or already acquainted with. Furthermore, a reader can decide not to open the hyperlinks in the order they are presented in the hypertext. For example, when a homepage has several hyperlinks to other pages, a reader can choose which one to read first. That could be the second one, since that one seems to lead to more interesting information. Subsequently, he may read all pages that page is linked with, before returning to the home page. In order to ensure that the text remains coherent regardless of the route a reader chooses, the writer needs to take care that the reader has or has had access to all relevant further information. Therefore, a writer could, in the first place, refer to all relevant pages with hyperlinks. In the second place, the writer could decide to restrict the possible reading routes. According to Carter (2003), this is especially recommendable for hypertextual argumentative texts. To convince a reader of one’s opinion, generally, a particular reasoning pattern is developed in argumentative texts. To ensure that the reader will follow that, a specific reading route needs to be established.
2005; Stahl et al., 2007), it cannot be claimed that hypertext writing is more beneficial than linear text writing. Moreover, as long as the interventions encompass more potential learning activities than just hypertext writing (Lehrer, 1993), it is hard to determine whether the learning effects are just a consequence of hypertext writing, or also of or especially of other activities like collaborative inquiry.
The most reliable results seem to come from the two experimental studies Braaksma et al. (submitted) report: one among tenth grade students in higher general secondary education and one among eleventh grade students in academic secondary education. In both studies, the students followed a lesson series in argumentative text writing. At the end of the lesson series, the students had to write an argumentative text in a hypertext or linear text format. Braaksma et al. (submitted) found mixed results. In the study among the tenth grade students, they discovered that students with a low level of prior content knowledge gained from hypertext writing, whereas students with a high level of prior content knowledge gained from linear text writing. In the study among the eleventh grade students, they discovered no differences in content knowledge acquisition between linear and hypertext writing.
Besides that it is not evidently proven that hypertext writing can enhance topic knowledge acquisition, it can also be questioned whether writing in that genre necessarily triggers the reflection processes on topic knowledge that lead to learning gains. When hypertext writers have access to an overview map of their hypertext, topic knowledge reflection is certainly to be expected. Pohl and Purgathofer (2004) logged the hypertext production process of 95 writers. They discovered that 87 of the writers often rearranged the structure of their hypertext at the overview map: 39 of them did that primarily by deleting nodes and creating new ones, 48 by moving nodes to other positions at the overview map. Yet, the results of a series of experiments of Bromme and Stahl (and colleagues) show that these reflection processes do not necessarily lead to learning gains. They compared the topic knowledge increase of hypertext writers who were aware of a particular structural constraint of a hypertext with the knowledge increase of hypertext writers who were not. Those who were aware of a hypertext’s non-linear nature showed more content knowledge increase than those who were not (Bromme & Stahl, 2005). Similarly, those who took into account multiple audience perspectives showed more content knowledge increase than those who did not (Bromme & Stahl, 2002; Stahl et al., 2007).
1.4. Processes in hypertext writing: this study
graphically. In the hypertext like writing task, students had to produce such argument scheme based on a provided linear text. While performing the latter task, the students planned and analyzed more than during the performance of the linear writing task. Yet, Braaksma, Rijlaarsdam, & Van den Bergh (2010) found different effects of hypertext and linear text writing in an exploratory analysis of the pause and production behavior of 8 hypertext writers and 8 linear text writers. Planning processes are expected to occur during pauses (cf. Schilperoord, 2001a; Spelman Miller, 2006). Braaksma et al. (2010) discovered that the linear text writers paused more and longer in general, and especially between words in the beginning of the writing process and between sentences in the middle of the writing process. The hypertext writers showed more production activities during the entire writing process.
Taking into account the results of Braaksma et al. (2010), one can wonder whether hypertext writing triggers more reflection processes. Yet, it is rather hard to get a complete account of a writing process based on just the frequency and duration of pausing and production activities. Production activities can encompass regular text production as well as revision activities. Moreover, a pause in the text production process is not necessarily an indication of planning (Spelman Miller, 2006). During a pause, a writer can also translate previously planned content, read previously written text, evaluate previously produced text, and be off-task. In this study, I use the methodology of Baaijen et al. (2012) to establish more elaborately what the effect of hypertext writing compared to linear text writing is on the writing process.
Baaijen et al. (2012) are the first who explicitly “tried to derive measures from the raw output of keystroke logs that could be related to cognitive models of writing” (p. 247). Therefore, they isolated “text production in its “purest,” most linear form from other activities” (p. 253). They just focused on the text that was produced during the writing of the first draft of a text. Post-draft revision was considered a different process. Within the activities that were registered during initial draft production, a distinction was made between text production activities and other activities like cursor movements and text deletions. Like Chenoweth & Hayes (2003), they considered the production of translated text ended when a pause of two seconds or longer occurred, and interrupted when another activity was executed. Consequently, Baaijen et al. (2012) followed Chenoweth & Hayes (2001 & 2003) in defining a piece of text produced before a pause as P-burst, and a piece of text produced for another activity as R-burst. When producing text on a computer, it is relatively easy to move the cursor from the leading edge of text production to a position elsewhere in the text. Therefore, Baaijen et al. (2012) considered it important to add the distinction that an R-burst could end with a revision at the leading edge (RL-burst) or elsewhere in the text (RI-burst).
In order to obtain an accurate measure of a P-burst’s length with keystroke logging data, Baaijen et al. (2012) considered it important to make a distinction among the P- and R-bursts based on after what they were initiated: a pause (PP- or PR-burst) or a revision (RP- or RR-burst). Chenoweth & Hayes (2003) based their theory on studies of writers’ think-aloud protocols. In these protocols, the translation process is verbalized. In keystroke logs, just the transcribed results of the translation process are registered. The revision process during translation is hidden. It could be that the transcribed P-bursts are revised during translation. In that case, they can be shorter than transcribed P-bursts, which are not revised during translation. Chenoweth & Hayes (2003) expect that the translation process of ideas into language stops, when a revision process is started. When this revision occurs half-way the translation process, the finally translated and subsequently transcribed burst is shorter than a burst that is produced in an uninterrupted translation processes. According to Baaijen et al. (2012), it depends on writers’ writing strategy whether they mainly transcribe P-bursts that are revised during translation. Other writers “transcribe the output of the translator directly and then revise as they go along” (p. 252). These writers generally transcribe longer PP-bursts, since the reviser does not often interrupt the translation process. Furthermore, they transcribe more RP-bursts that are relatively short in length. After they have interrupted the transcription of a burst for revision purposes, they can produce an RP-burst which just involves the modification of that burst. These RP-bursts are considerably shorter than PP-bursts, which only involve the production of new content.
If one wants to measure writers’ fluency with keystroke logging data, one should consequently not take into account the mean length of all P-bursts, but just the mean length of the PP-bursts. This mean PP-burst length cannot be used to compare the fluency of two individual writers. When one writer produces shorter PP-bursts, this is not necessarily an indication of a smaller translation capacity. He can also mainly transcribe PP-bursts that are revised during translation. However, the mean PP-burst length of group can be considered an accurate indication of the fluency of that group. When taking the mean PP-burst length of a group of writers, individual differences in writing strategies are leveled out. Thus, it seems worthwhile to compare the mean PP-burst length of eleventh grade students writing an argumentative text with the mean length of the PP-bursts, which the graduate students in Baaijen et al.’s (2012) methodological study produced while writing an argumentative kind text. Their mean PP-burst length was six words.
Since an analysis of the different bursts and pauses P-bursts end in can reveal the planning, translation and revision behavior during the writing process, I decided to determine what the effect of hypertext writing compared to linear text writing is on the writing process, by analyzing what the effect is on:
1. the frequency with which the different bursts occur; 2. the length of the bursts;
3. the frequency with which is paused at different textual levels; 4. the duration of pauses at the different text locations;
6. the overall character of the writing process, as revealed by a principle component analysis based on a variety of measures derived from the keystroke logs.
It is first established how frequently eleventh grade linear text writers produce the different language bursts, how long the produced bursts are, and how frequently and long they pause at the different textual levels. I consider it necessary to unravel what the pause, translation, and revision behavior of eleventh grade linear writers is in light of Baaijen et al.’s (2012) methodology, in order to accurately describe what the effect is of writing a hypertext on the writing process. It is rather difficult to derive the unaffected state from the two previous studies of keystroke logging data of eleventh grade students: Groenendijk et al. (2008) and Johansson (2009). In the first place, Groenendijk et al. (2008) studied writers who wrote a completely different genre, a poem. In the second place, Groenendijk et al. (2008) and especially Johansson (2009) used a very different methodology. Johansson (2009) counted the number of presses on the delete and backspace button; she did not determine whether the students did a revision at the leading edge of text production or elsewhere. Groenendijk et al. (2008) and Johansson (2009) did not distinguish language bursts, and subsequently did not determine the length of it.
It is hypothesized that eleventh grade linear writers produce more R-bursts than P-bursts. Since previous research has revealed that their linguistic proficiency is not matured (Coirier et al., 2002; Johansson, 2009), they probably often have to interrupt the translation as well as the transcription process for revision (cf. Kaufer et al., 1986; Chenoweth & Hayes, 2001, 2003). These revisions are expected to be done at the leading edge, leading to a higher frequency of RL-bursts than RI-bursts. The insufficient linguistic competence presumably also negatively influences their translation capacity (cf. Kaufer et al., 1986; Chenoweth & Hayes, 2001, 2003). Therefore, it is supposed that the eleventh graders produce PP-bursts, which are smaller in length than the bursts the graduate students in the study of Baaijen et al. (2012) produced. Being unable to translate large idea packages into language, eleventh grade students assumably do not translate language clause by clause like professional writers (cf. Kaufer et al, 1986; Schilperoord, 2001b). I rather expect them to translate a few words at a time, and consequently pause rather frequently between words, besides between paragraphs, between sentences, and between subsentences.
2. METHOD
The data dealt with in this study were collected in an experimental study of Braaksma et al. (submitted) into the effects of hypertext writing and observational learning on content knowledge and writing skill acquisition. Four eleventh-grade classes in academic secondary education of a Dutch city high school followed a lesson series in argumentative writing. By random assignment, it was determined whether the students had to write the final argumentative essay as a linear text or a as a hypertext, or whether they just had to observe other students writing that essay as a hypertext. 22 Students were placed in the linear writing condition, 26 in the hypertext writing condition, and 30 in the observational learning condition.
2.1. Participants
For this study, fifteen students were randomly selected from the eleventh-grade students in the linear writing condition (LIN), and fifteen of the students in the hypertext writing condition (HYP). All these 17-year-old students had some experience with writing linear argumentative texts.
2.2. Materials and procedure
The students had to write an argumentative essay of approximately 500 words. The essay had to address the issue: Should charities be disconnected from commercial lotteries?. In the four preceding lessons, the students had done several (collaborative) inquiry activities (cf. Hillocks, 1982 and 1995), as to be able to write the essay as a well-structured linear text or hypertext with an attractive introduction (see Appendix A for an overview of the complete lesson series). In the fourth lesson, the students had developed an argumentation plan, which had to function as basis for the content of the essay. In this plan, the students had defined their opinion, and had listed three main and subordinate arguments to substantiate it. Most of the arguments were based on provided documentation. While writing the essay, students had access to the documentation too.
The production of the essay was divided in two phases. In the final ten minutes of the fourth lesson, the students had to write a draft version of the introduction and had to think of a title for the whole essay. Main aim of the introduction production task was the application of recently acquired knowledge about attracting readers’ attention. Except for the requirement to already make a hyperlink to the arguments in the HYP condition, the task was the same for the students in both conditions.
During the production of the argumentation and conclusion, participants’ writing process was registered by logging their keystrokes with Inputlog (cf. Leijten & Van Waes, 2013). Inputlog registers all keystrokes, mouse movements, and mouse clicks that a writer makes in a Microsoft Word environment, likewise as the duration of all moments of inactivity – i.e., pauses. The ability to use Microsoft Word had several advantages. The text production had a relatively natural and familiar character, since all students were well-acquainted with that word processing application. Furthermore, students could use the rather user-friendly hypertext creation possibilities of that application. In contrast to hypertext writers in other hypertext writing studies, the students had no access to an overview map (cf. Bromme & Stahl, 2002 & 2005; Lohr et al., 1996; Pohl & Purgathofer, 2004; Stahl et al., 2007).
2.3. Data preparation
As to be able to analyze the process of writing the argumentation and conclusion of the argumentative essay with the keystroke logs, I followed the analysis methodology of Baaijen et al (2012). First of all, the production of text was automatically discriminated from the other registered activities: mouse movements, mouse clicks and presses on a navigation key, the backspace button or the delete button. Text production was considered ended, when such activity was executed. Except for the cases, the backspace button was pressed to erase a typo. In conformity with conventional practice (cf. Wengelin, 2006), I did not consider a typo as an interruption of the text production process. Furthermore, the production of text could end in a pause of two seconds or longer. The pieces of text that were produced between the termination points were coded as language bursts, following Chenoweth and Hayes (2001 & 2003).
A typo was defined as a correction within a word, which does not last two seconds or longer. As a measure for the duration of a correction, I took the time between the last key press for the deletion of the wrongly typed letters and the first key press afterwards. This can be illustrated with the excerpt of a keystroke log in Table 1. The backspace presses in row 10-12 were considered a typo correction, and therefore ignored. It took less than two seconds to delete ‘ens’: 149.859-149.109=0.750 seconds. The writer seems to consider her subsequent revision of the typo ‘dens’ incorrect, since she immediately deletes it after production (row 17-20). This deletion again last less than two seconds: 152.172-150.406=1766 seconds. Yet, since she actually is struggling with the correct production of ‘tijdens’ from second 149.109, the continuous text production has already been stopped for 3.063 seconds (152.172-149.109). If a typo was revised several times, it was, therefore, not considered a typo anymore, when the time between the last key press for the first erasure (row 9 in Table 1) and the first key press after the last erasure (row 21 in Table 2) exceeded the two seconds threshold.
Keystroke Input time Output time Action time Pause time 1 Movement 137.422 137.531 0.109 0.656 2 Left Button 138.078 138.140 0.662 8.344 3 SPACE 146.422 146.500 0.078 1.750 4 T 148.172 148.234 0.062 0.359 5 i 148.531 148.672 0.141 0.094 6 j 148.625 148.750 0.125 0.218 7 e 148.843 148.922 0.079 0.063 8 n 148.906 149.015 0.109 0.109 9 s 149.015 149.109 0.094 0.125 10 BS 149.140 149.218 0.078 0.110 11 BS 149.250 149.297 0.047 0.140 12 BS 149.390 149.453 0.063 0.469 13 d 149.859 149.937 0.078 0.156 14 e 150.015 150.093 0.078 0.157 15 n 150.172 150.250 0.078 0.156 16 s 150.328 150.406 0.078 0.625 17 BS 150.953 151.000 0.047 0.156 18 BS 151.109 151.265 0.078 0.156 19 BS 151.265 151.423 0.078 0.158 20 BS 151.423 151.595 0.062 0.577 21 d 152.172 152.234 0.062 0.156 22 e 152.328 152.390 0.062 0.156 23 n 152.484 152.562 0.078 0.172 24 s 152.656 152.734 0.078 0.172 25 SPACE 152.828 152.890 0.062 0.156 26 d 152.984 153.047 0.063 0.172 27 e 153.156 153.234 0.078 0.172 28 SPACE 153.328 153.390 0.062 4.256 29 s 157.490 157.568 0.078 0.166 30 h 157.734 157.812 0.078 0.188 31 o 157.922 158.000 0.078 0.078 32 w 158.000 158.078 0.078 0.156 33 s 158.156 158.265 0.109 0.109
time at a particular unit boundary. I used the conventional threshold level of two seconds, to distinguish cognitive processing pauses from normal letter and word transition time (cf. Wengelin, 2006).
2.3.1. Burst coding
The bursts that the participants produced during the writing process were classified with the classification system in Table 2. I adapted the classification system of Baaijen et al. (2012, p. 258) in several respects.
First, the hypertext writers did not just move the mouse for revision purposes. They also had to do that to create a hyperlink. Since I considered it important to give an account of hyperlink creation, I decided not only to distinguish P- and R-bursts, but also H-bursts. Likewise as Baaijen et al. (2012), I classified bursts as P-burst when it ended in a pause of at least two seconds, and as R-burst when it ended in a mouse movement, mouse click, navigation key press, backspace key press and deletion key press. A series of mouse movements could certainly not be considered a revision, when it was followed by the typing of the name of the file under which the next page of a hypertext was saved. The typing of such file name is a necessary step in the production of a hyperlink in Microsoft Word. Since the bursts that were produced before hyperlink creation ended with it, they were coded as H-bursts.
Second, it appeared to be far from easy to determine whether an R-burst ended in a revision at the leading edge (RL-burst) or a revision elsewhere in the text (RI-burst). In case text production is immediately followed by backspace presses, it is quite simple to conclude that revision at the leading edge of text production is done. However, when mouse movements are registered before another language burst is produced, we cannot straight away conclude that such piece of text is inserted at another position in the text. Some students delete the last produced words by selecting them and then deleting them or directly overwriting them with other words. In the latter case, the revision takes place at the leading edge. To determine whether the mouse was moved to do a revision at the leading edge or elsewhere in the text, it was of help to consult the final texts. In some cases, even then, no decision could be made. Therefore, I added the residual category RM-burst. Following Baaijen et al. (2012), I classified a burst as a RLI-burst, when it terminated in a revision at the leading edge that was immediately followed by a revision insertion.
production was continued at a different text location, the first burst produced at that different text location was flagged as preceded by a leading edge shift. Until text production was resumed at another text location, the leading edge of text production was there. The production of a title later in the text production process was also considered normal text production preceded by a leading edge shift. This decision was, in the first place, made to generalize over the students: some produced a title of a few
Burst type Definition
PP Burst initiated after and terminated in a pause of at least two seconds.
PRL Burst initiated after a pause of at least two seconds and terminated by a revision at the leading edge.
PRI Burst initiated after a pause of at least two seconds and terminated by a revision insertion – that is, a revision carried out elsewhere in the text either within the same sentence or elsewhere in the text.
PRLI Burst initiated after a pause of at least two seconds and terminated by a revision at the leading edge as well as a revision elsewhere in the text.
PRM Burst initiated after a pause of at least two seconds and terminated by a revision of which the character could not be distinguished.
PH Burst initiated after a pause of at least two seconds and terminated by the creation of a hyperlink.
RP Burst initiated after a revision and terminated in a pause of at least two seconds. RRL Burst initiated after a revision and terminated by a revision at the leading edge. RRI Burst initiated after a revision and terminated by a revision insertion – that is, a
revision carried out elsewhere in the text either within the same sentence or elsewhere in the text.
RRLI Burst initiated after a revision and terminated by a revision at the leading edge as well as a revision elsewhere in the text.
RRM Burst initiated after a revision and terminated by a revision of which the character could not be distinguished.
RH Burst initiated after a revision and terminated by the creation of a hyperlink.
HP Burst initiated after the creation of a hyperlink and terminated in a pause of at least two seconds.
HRL Burst initiated after the creation of a hyperlink and terminated by a revision at the leading edge.
HRI Burst initiated after the creation of a hyperlink and terminated by a revision
insertion – that is, a revision carried out elsewhere in the text either within the same sentence or elsewhere in the text.
HRLI Burst initiated after the creation of a hyperlink and terminated by a revision at the leading edge as well as a revision elsewhere in the text.
HRM Burst initiated after the creation of a hyperlink and terminated by a revision of which the character could not be distinguished.
HH Burst initiated after the creation of a hyperlink and terminated by the creation of another hyperlink.
IH Insertion burst terminated by the creation of a hyperlink.
IG Insertion burst that is initiated in the middle of sentence production and involves a revision within the current sentence.
IR Insertion burst that is initiated at the end of sentence completion and involves a revision within the sentence that just has been produced.
IB Insertion burst that is initiated either in the middle of sentence production or after sentence completion but involves a revision carried out over sentence boundaries – that is, elsewhere in the text and no longer than a sentence in length.
words, others of a whole sentence. Furthermore, the production of a title resembled more common text production. It often took several bursts to produce one.
Fourth, following Baaijen et al. (2012), I made a distinction among the P-, R- and H-bursts based on their initiation type. In this study, H had not only to be added as termination type, but also as initiation type: a burst could be initiated after a pause longer or equal to two seconds, a revision, or a hyperlink creation. Consequently, HP-, HR- and HH-bursts were distinguished, besides PP-, PR-, PH-, RP-, RR-, and RH-bursts. In addition, the category IH-burst was distinguished. In some cases, hyperlink creation directly followed an insertion burst production. In order to be then able to indicate that a hyperlink was created, I added that category. That category did not serve as a substitution of the categories that Baaijen et al. (2012) distinguished among the I-bursts. I considered it important to distinguish IG-, IR-, and IB-bursts too, since “(a) IG-bursts reflect within-sentence revision; (b) IR-bursts reflect end-of-sentence revision; and (c) IB-bursts reflect revision over sentence boundaries” (p. 257). In case, an I-bursts was terminated by hyperlink creation, it received as additional code ‘IH-burst’.
2.3.2. Pause coding
In the pause analysis, only the pauses in which the P-bursts ended were taken into account. I agreed with Baaijen et al’s (2012) assumption that one can only certainly know that a writer pauses for the planning or formulation of the next text unit at such linear transitions. When a writer pauses two seconds or longer before, between, and after mouse movements, navigation key strokes, and backspace or delete button presses, he could also be concerned with the revision of already produced text.
The pauses in linear text production were coded based on the textual location at which they occurred: within words, between words, between subsentences, between sentences, and between paragraphs. Inputlog automatically distinguishes pauses within words, pauses between words, pauses between sentences, and pauses between paragraphs. I manually discriminated pauses between subsentences from pauses between words, based on the definition of Baaijen et al. (2012): a pause between words can be considered a pause between subsentences, when a comma, colon, or semicolon is produced during that pause.
2.4. Data analysis
leading edge shifts occurred in a changing distribution over the writing process, and whether this distribution was different for the linear and hypertext writers. Finally, I examined whether the overall character of the writing process was different for the linear and hypertext writers with a principle component analysis. The methodology of the different analyses is presented in the order the analyses have been presented here. In the same order, the results of the analyses are presented in the result section (§3).
2.4.1. The text production process of eleventh grade students
As stated, the text production process of eleventh grade students was analyzed by assessing the frequency with which the linear text writers produced the different kinds of bursts, and what the length of the bursts was in words. Their pause behavior was determined based on the frequency in which they paused at different text locations, and the percentage of the total pause time that they spent at the different text locations.
I used the probability that a particular burst or pause type occurred during the writing process of a writer as a measure of the frequency of occurrence of the distinct burst and pause types. By taking the probability of occurrence instead of the exact frequency, it was possible to circumvent the influence that the length of the writing process had on the frequency in which the different bursts or pauses occurred. A student who produced 200 bursts presumably produced much more R-bursts than a student who just produced 60 bursts, although the proportion could be equal.
Repeated-measures ANOVAs were executed in order to determine whether the burst and pause types differed in occurrence and length. Burst or pause type was taken as the independent variable, and probability of occurrence, mean burst length, or percentage of total pause time as the dependent variable. When the normality of distribution assumption was violated, the non-parametric variant ‘Friedman’s ANOVA’ was used. For two- or three-way ANOVAs, no non-parametric variant was available. When I had to use such analyses to analyze a non-normally distributed dataset, I log transformed the data. For the sake of interpretability, the normal ‘untransformed’ means and standard deviations are reported. When the sphericity assumption for a parametric within-subjects test was violated, the degrees of freedom were adjusted with the Greenhouse-Geisser correction.
Post-hoc test were carried out to break down an overall effect. To control for the familywise error rate, Bonferroni correction was applied. In case of post-hoc tests after a parametric within-subjects test, SPSS automatically corrects the p-values of the separate analyses, when one has opted for a Bonferroni correction. Consequently, it is then correct to stick to an α-level of 0.05. When Wilcoxon’s tests are done after a Friedman’s ANOVA, the α-level has to be adjusted manually, by dividing it through the total number of tests.
frequency of occurrence or length. If they did, it was established whether the initiation type of the P- or R-burst affected that pattern. Finally, it was tested whether the RL- and RI-bursts differed in frequency of occurrence and length, likewise as the IB-, IR-, and IG-bursts. In the analysis among the different R-bursts, the RLI- and RM-bursts were not included. In order to assess the exact difference between revisions at the leading edge and revision insertions, I considered it best to just take into account the ‘pure’ variants, RL- and RI-bursts.
2.4.2. The effect of hypertext writing on the text production process To establish whether hypertext writing influenced the pause, text production, and revision behavior of eleventh grade students, it was investigated whether the hypertext and linear text writers (1) produced the different language bursts in a different frequency, (2) produced language bursts with a different length, (3) paused in a different frequency at the different text locations, and (4) spent a different percentage of the total pause time at the different text locations. For the first analysis, a MANOVA was executed with condition as independent variable, and the probability of occurrence of each burst type as dependent variables. For the other analyses, mixed ANOVAs were carried out with burst or pause type as within-subjects factor, condition as between-subjects factor, and probability of occurrence, burst length, or percentage of total pause as dependent variable. Because there is no non-parametric alternative for a mixed ANOVA, a log transformation was applied on data that was not normally distributed.
2.4.3. The distribution of activities over the writing process
Generalized additive mixed effects regression was used to investigate whether the probability that a leading edge shift, burst type, or pause type occurred, changed over the writing process of both linear and hypertext writers, and whether that distribution pattern was different for the linear and hypertext writers. By means of generalized additive modeling, monotomic smooth functions could be developed, which estimated the probability of occurrence of a particular activity at every moment of the writing process for all students in general, for every student individually, and all students in one condition.
Variables
New binary variables had to be created, to be able to predict the probability that a leading edge shift, a particular burst or a particular pause occurred over the writing process. For each burst type, a variable indicated which produced bursts were of that type (1), and which were not (0). For each pause type, a variable indicated which bursts were produced after a pause of that pause type (1), and which were not (0). Finally, one variable indicated which bursts were produced after a leading edge shift (1), and which were not (0).
in case a student produced 200 bursts. However, for such a student, burst 50 would occur in the beginning of the writing process, whereas it would occur in the middle for a student who produced 100 bursts. To standardize the time indication, I, therefore, divided the number of each burst by the total number of bursts a student produced. Then, the text production process of every student ran from 0 to 1, and, for example, the time indication .5 (the middle of the process) was given to burst 50 in case a student produced 100 bursts, and to burst 100 when a student produced 200 bursts.
Generalized additive mixed modeling
Since I used binary dependent variables, logistic regression had to be opted to model the data. With logistic regression, the occurrence of a burst or pause type cannot be modeled directly. It only estimates the probability that a burst or pause occurs given the specified predictors. This probability is given in logits, which are the natural logarithms of odds: the change that an activity occurs divided by the change that it does not occur.
I opted for generalized additive models to model the data instead of linear models, since I assumed that the distribution of leading edge shifts and the different types of bursts and pauses was non-linear. Insertion bursts, for example, have presumably especially a high probability of occurrence at the moments that writers have finished the production of a piece of text, and reread it to check it and/or to get inspiration for the subsequent text part.
Usually, several growth models are applied to a dataset, in order to determine which of them best describes the (non-)linear development of a process over time (cf. Van den Bergh & Rijlaarsdam, 1996). The disadvantage of this method is that a predefined form is imposed on a dataset (Baayen, 2008). With generalized additive modeling (GAM) in R, a more flexible, non-linear trend for a specific dataset can be determined based on the input (Wood, 2006). In GAM, a standard linear model with regression coefficients is combined with smooth functions for one or more predictor(s). The basic smooth function for one predictor is cubic regression splines, which fits cubic polynomials (functions of the form: y=a + bx + cx2 + dx3) to several intervals of the
predictor values, and connects these to form a continuous curve. Each point at which two polynomials are joined is a knot. The more knots a curve has, the more smooth and wiggly the function is. With this method, one runs the risk of under smoothing and especially over smoothing. Therefore, the mcgv-package of Wood (2006) uses the method of cross validation: a model is constructed on one part of the dataset, and tested on another part. The package takes 9 knots as baseline, and consequently checks whether the dataset can be modeled sufficiently accurate with less. The final number of knots can be determined based on the presented effective degrees of freedom. These give an indication of the number of cubic polynomials needed to accurately predict the data.
experimental group. The Chi-square test can reveal for each of the groups whether they use a process in a changing distribution as well. Before the regression model is scrutinized, it should, however, first be checked whether the addition of any factors improves the model. Therefore, the AIC score of the new model has to be compared with the AIC score of the former one that is more basic. When the AIC score of the new model is at least 2 lower than the AIC score of the former model, the model can be considered a significant improvement. The AIC score gives an indication of the goodness-of-fit of a model, taking into account the complexity of the model, i.e., the number of parameters. The lower the score is, the better the model is.
Modeling the distribution of bursts, pauses and leading edge shifts
In order to investigate whether the probability of occurrence of the leading edge shifts and the particular bursts and pauses changed over the writing process, I developed generalized additive mixed logistic regression models for each of them individually. I started with a model that estimated the probability of occurrence based on only a smooth function for time. When the smooth function was significantly different from a flat line with a logit of 0, I concluded that a leading edge shift, the burst or the pause occurred in a changing distribution over the writing process of all participants. In that case, the basic model was extended with the addition of a random subject factor, to account for the non-exhaustiveness of the sample (cf. Baayen, 2008). If the addition of a random subject factor improved the model, I added to that one the predictor that the time course differed for the linear and hypertext writers. If the addition of a random subject factor did not improve the model, I added the condition predictor to the original model with a non-random time curve.
2.4.4. The general character of the writing process
Finally, a principle component analysis was executed to capture the overall character of the text production process of the linear and hypertext writers, and with the students’ component scores, the differences between the conditions in the overall constellation of the writing process. To obtain a complete overview of the writing process with the principle component analysis, variables were included that account for (1) the text production activities of the students, (2) their pause behavior, and (3) the (non-)linearity of their writing process.
Variables
I tried to capture the distinctive text production activities by taking into account the probability of occurrence of P-, R-, I-, and RL-bursts. To limit the number of variables, I decided to just focus on the probability of occurrence of the main burst types. The percentage of H-bursts was not included as a variable, since that variable just accounted for a characteristic of hypertext production. I was interested in whether hypertext and linear text production differed in aspects both had in common. I did consider it important to take into account the distinction between revisions at the leading edge and revision insertions, since Baaijen et al. (2012) concluded that the different kinds of revision differed in character. Whereas the probability of occurrence of I-bursts gives an indication of the frequency in which a burst ended with an revision insertion, a different variable had to be added to account for the frequency in which leading edge revisions occurred. Furthermore, the text modification index of each student was enclosed in the PCA analysis, in order to have an indication of the amount of words students deleted and/or changed during the text production process (Baaijen et al., 2013). A student’s text modification index was calculated by dividing the number of produced words by the number of words in the final text. Since the majority of the students already had produced the introduction of the text, the number of already produced words was subtracted from the number of words of the final text, before the text modification index was calculated.
To account for students’ pause behavior, in the first place, the proportion of writing time spent on reflection was added as variable. This was the sum of all time paused at linear transitions divided by the total writing time. Students’ word and sentence linear transition index were considered to discover at which locations ‘empty’ pauses were most likely (Baaijen et al., 2012). These indices stood for the proportion of linear transitions at word and sentence level, i.e., the percentage of transitions between words and sentences that involved the direct continuation of text production. Hence, the production of the next word or sentence was not preceded by a mouse movement, mouse click, press on a navigation key, or press on the delete or backspace button.
was, it was considered a non-linear burst. Thus, all I-bursts were marked as non-linear bursts as well as the first burst after a leading edge shift. To have a measure of the percentage text that was not produced in the same order as it ended up in the final text, finally, the sentence non-linearity index and the paragraph non-linearity index were included as variables. These indices represented the percentage of sentences or paragraphs in the final text that were not produced linearly, but were inserted after the students had produced following parts of the text.
Principle component analysis
The decision which components had to be retained was made based on the inspection of the scree plot. Such a plot has the shape of a scree, since just a few components have a high eigen value, and all the others a very low one of approximately zero (Field, 2009). Only the components with eigen values above the point at which the curve flattens out should be retained. Yet, the sample size was rather low for a principle component analysis. Thirty is far less than the prescribed 130 (number of variables * 10) (cf. Field, 2009). To determine the significance of the components, therefore, also the variance that a component explained and the loadings of the variables on the component were taken into consideration. I strived that at least four variables loaded higher than 0.6 on a component, since the component can then be regarded reliable regardless of any sample size (cf. Guadagnoli and Velicer, 1988).
The orthogonal rotation method, Varimax, was opted to rotate the extracted components such that they became independent components, on which some variables loaded highly and some lowly. To check whether it was appropriate to use orthogonal rotation, I verified whether the component transformation matrix was nearly symmetrical.
3. RESULTS
In this section, I first present the analyses into the pause, text production, and revision behavior of eleventh grade students writing a linear text. At the end of §3.1, the results are discussed in light of theories about text production (cf. Chenoweth & Hayes, 2003) and pause behavior (cf. Schilperoord, 2001a). Afterwards, it is established what the effect of hypertext writing compared to linear text writing is on the writing process. In §3.2, it is tested whether hypertext writing affects the pattern that is unraveled with the analyses into the pause, text production, and revision behavior of eleventh grade students writing a linear text. In §3.3, it is determined whether the bursts and pauses occurred in a changing frequency over the writing process, and whether the text type students wrote in influenced the distribution. Finally, in §3.4, the results of a principle component analysis are reported, with which I investigated whether the hypertext and linear text writers differed in the overall constellation of their writing process.
3.1. The text production process of eleventh grade students
I tested whether the burst types differed in probability of occurrence and in length, in order to assess the text production and revision behavior of eleventh grade high school students writing a linear text. Their pause behavior was determined by analyzing how frequently they paused at different textual locations, and how much pause time they spent at each level.
3.1.1. Burst frequency
A within-subjects ANOVA with main burst type as independent variable and probability of occurrence as dependent variable showed that the frequency with which a burst occurred significantly depended on its type (F(2, 28) = 51.02, p<.001). Planned pairwise comparisons revealed that each main burst type differed significantly from the other main types in occurrence (p<.01). The chance that a burst was an R-burst was highest (M=.524, SD=.024), successively followed by the chance that a burst was a P-burst (M=.324, SD=.023), or an I-burst (M=.141, SD=.019).
The considerable difference in occurrence of the main types influenced the occurrence of the subtypes. When the PP-, RP-, PR-, and RR-bursts made up the levels of the independent variable ‘burst type’, a within-subjects ANOVA disclosed a significant effect of burst type on probability of occurrence as well (F(1.35, 18.89) = 18.35, p<.001). Post-hoc test showed that the RR-bursts (M=.335, SD=.027) occurred significantly more frequently than the PP-burst (M=.137, SD=.019), RP-bursts (M=. 188, SD=.010), and PR-bursts (M=.188, SD=.010) (p<.01).
The insertion bursts also differed in frequency of occurrence. With a Friedman’s ANOVA, a significant overall effect of insertion burst type was found (χ2(2)=12.57, p<.
01). Using Wilcoxon tests and an adjusted α-level of .017, it was demonstrated that IB-bursts (Mdn=.085) occurred significantly more often than IR-IB-bursts (Mdn=.027) and IG-bursts (Mdn=.037) (p<0.01).
significantly in occurrence (χ2(3)=34.89, p<.001). After adjusting the level of
significance to .008, Wilcoxon tests showed that the PRL- and PRI-bursts as well as RRL- and RRI-bursts differed significantly from one another in occurrence (p<.001). The RL-bursts (PRL: Mdn=.133; RRL: Mdn=.250) appeared more often than the RI-bursts (PRI: Mdn=.027; RRI: Mdn=.060).
In conclusion, the results of the analyses of burst frequency show that the text production process was most often interrupted for revision. The bursts that were initiated after a revision frequently again ended in a revision. These revisions were mainly done at the leading edge of text production. When the writers did make a revision insertion, they generally made it in a previously produced sentence, and not in the sentence they were producing at that moment or just had produced.
3.1.2. Burst length
I executed a repeated-measures ANOVA with main burst type as independent variable and burst length as dependent variable, as to establish whether the burst types differed in burst length. A significant main effect of main burst type on burst length was found (F(1.28, 4.98)=55.14, p<.001). Planned pairwise comparisons revealed that the P- (M=4.11, SD=1.08) and R-bursts (M=4.25, SD=.82) were comparable in length (p>. 05), and that the I-bursts (M=1.85, SD=.71) were significantly shorter (p<.001).
It did not affect the length of a P- or R-burst whether it was initiated after a pause or revision. Another two-way (2*2) within-subjects ANOVA was carried out with P and R as levels of the independent variable ‘initiation type’ and the independent variable ‘termination type’, and burst length as dependent variable. Neither a significant overall effect of initiation type on burst length was found (F(1, 14)=1.71, p>.05), nor a significant overall effect of termination type on burst length (F(1, 14)=.96, p>.05), nor an interaction between initiation and termination type (F(1, 14)=.05, p>.05).
Furthermore, no difference in length could be found between bursts that ended with a revision at the leading and bursts that ended with a revision insertion. A two-way (2*2) within-subjects ANOVA was carried out with P and R as levels of the independent variable ‘initiation type’, RL and RI as levels of the independent variable ‘termination type’, and burst length as dependent variable. The analysis showed no significant overall effect of initiation type on burst length (F(1, 13)=1.34, p>.05), no significant overall effect of termination type on burst length (F(1, 13)=1.96, p>.05), and no interaction between initiation and termination type (F(1, 13)=.48, p>.05)
Among the insertion bursts, a difference in length was observed. A repeated measures ANOVA with insertion type as independent variable and burst length as dependent variable revealed a significant main effect (F(2, 28)=4.92, p<.05) with IB-bursts (M=1.98, SD=.82) being significantly longer than IR-IB-bursts (M=1.19, SD=.90) (p<.05).
significantly shorter than P- and R-bursts. Insertions in previously produced sentences were longer than insertions in sentences that were just completed.
3.1.3. Pause location
The eleventh grade writers did not pause equally often on the distinguished textual locations: within words, between words, between subsentences, between sentences, and between paragraphs. A Friedman’s ANOVA with pause type as independent variable and probability of occurrence as dependent variable disclosed an overall difference in occurrence between the different pause types (χ2(4)=43.89, p<.001). Using Wilcoxon tests and an α-level of .005, it was found that the students paused more often between words (Mdn=.705) than within words (Mdn=.022), between subsentences (Mdn=.087), between sentences (Mdn=.129), and between paragraphs (Mdn=.024) (p<.001). Furthermore, pauses between sentences occurred more often than pauses within words, and between paragraphs (p<.001), and pauses between subsentences occurred more often than pauses within words (p<.005).
When repeating the analyses with percentage of pause time as dependent variable, comparable results were found. A Friedman’s ANOVA showed a significant main effect of pause type on percentage of pause time (χ2(4)=41.30, p<.001). Wilcoxon tests
revealed that the following differences were significant at an α-level of .005: (a) the
percentage of total pause time paused between words (Mdn=.600) was significantly higher than the percentage time paused within words (Mdn=.007), between subsentences (Mdn=.088), between sentences (Mdn=.201), and between paragraphs (Mdn=.042). (p<.001); and (b) the percentage of the total pause time paused between subsentences and sentences was significantly higher than the percentage time paused within words (p<.001).
These results thus show that writers most often paused between words, and that they spent most pause time at that location. Pauses between subsentences and sentences followed at considerable distance in probability of occurrence and time spent on it, and pauses within words and between paragraphs at considerable distance from pauses between subsentences and sentences. Whereas writers paused significantly more often between sentences than between paragraphs, they did not spent significantly more pause time between sentences than between paragraphs.
3.1.4. Discussion of the writing process of eleventh grade students In this study, the pause, translation, and revision behavior was determined of eleventh grade high-school students writing a linear text. I analyzed the frequency with which the eleventh grade students produced the different kinds of language bursts (cf. Baaijen et al., 2012), the length of the produced bursts, the frequency with which they paused at different text locations, and the amount of pause time they spent at these locations.
