DoSVis: Document Stance Visualization

Kostiantyn Kucher

, Carita Paradis

and Andreas Kerren

Department of Computer Science, Linnaeus University, V

axj

o, Sweden

Centre for Languages and Literature, Lund University, Lund, Sweden

Keywords:

Stance Visualization, Sentiment Visualization, Text Visualization, Stance Analysis, Sentiment Analysis, Text

Analytics, Information Visualization, Interaction.

Abstract:

Text visualization techniques often make use of automatic text classiﬁcation methods. One of such methods

is stance analysis, which is concerned with detecting various aspects of the writer’s attitude towards utter-

ances expressed in the text. Existing text visualization approaches for stance classiﬁcation results are usually

adapted to textual data consisting of individual utterances or short messages, and they are often designed for

social media or debate monitoring tasks. In this paper, we propose a visualization approach called DoSVis

(Document Stance Visualization) that focuses instead on individual text documents of a larger length. DoSVis

provides an overview of multiple stance categories detected by our classiﬁer at the utterance level as well as

a detailed text view annotated with classiﬁcation results, thus supporting both distant and close reading tasks.

We describe our approach by discussing several application scenarios involving business reports and works of

literature.

1 INTRODUCTION

Textual data has been playing an increasingly impor-

tant role for various analytical tasks in academic re-

search, business intelligence, social media monitor-

ing, journalism, and other areas. In order to explore

and make sense of such data, a number of text vi-

sualization techniques have emerged during the last

20 years (J

anicke et al., 2015; Kucher and Kerren,

2015). The majority of text visualization techniques

rely on methods originating from computational lin-

guistics and natural language processing which ana-

lyze the speciﬁc aspects of texts, such as topic struc-

ture, presence of named entities, or expressions of

sentiments and emotions. The latter one, i.e., sen-

timent analysis / opinion mining, has usually been

associated with data domains such as customer re-

views, social media, and to a lesser degree, literature

and political texts (Pang and Lee, 2008; Mohammad,

2016). There is also research on sentiment analy-

sis of business reports and CEO letters which studies

the relation between the language and ﬁnancial indi-

cators (Kearney and Liu, 2014; Nopp and Hanbury,

2015). The existing sentiment visualization tech-

niques for textual data support a variety of data do-

mains, data source types, and user tasks (Kucher et al.,

2017a). At the same time, few existing visualiza-

tion techniques make use of another method related

to sentiment analysis—stance analysis (Mohammad

et al., 2016; Skeppstedt et al., 2016b; Simaki et al.,

2017b). Stance analysis of textual data is concerned

with detecting the attitude of the writer ranging from

the general agreement/disagreement with a certain ut-

terance or statement (e.g., “I hold the same position

as you on this subject”) to the more ﬁne-grained as-

pects such as certainty/uncertainty (e.g., “I am not

completely convinced that it really happened”). The

StaViCTA project

has taken the latter approach in

order to develop an automatic stance classiﬁer and

visualize stance detected in textual data. The exist-

ing stance visualization techniques have usually fo-

cused on political text data such as transcripts of de-

bates (El-Assady et al., 2016), blog posts and com-

ments (Kucher et al., 2016a; Kucher et al., 2016b),

and tweets (Mohammad et al., 2016; Martins et al.,

2017).

In this paper, we explore other possible applica-

tions of visual stance analysis and focus on data do-

mains and user tasks that are not addressed in the ex-

isting literature. In contrast to the techniques which

support visual analysis of multiple short documents

such as social media posts, we look into scenarios in-

Advances in the description and explanation of Stance

in discourse using Visual and Computational Text Analytics

(http://cs.lnu.se/stavicta/).

168

Kucher, K., Paradis, C. and Kerren, A.

DoSVis: Document Stance Visualization.

DOI: 10.5220/0006539101680175

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 3: IVAPP, pages

168-175

ISBN: 978-989-758-289-9

Figure 1: Visualization of a 16th century political treatise “The Prince” by Niccol

o Machiavelli in our tool DoSVis: (a) sliders

used for global ﬁltering and toggling of warning symbols; (b) scatterplot-like overviews based on the detected occurrences of

stance categories in the text; (c) a viewport rectangle representing the currently visible area of the document; (d) overviews of

detected stance category combinations (created with drag’n’drop); (e) ﬁltering and navigation controls for category overviews;

(f) ﬁltering and navigation controls for category combination overviews; (g) the detailed text view including a sidebar mark

for the currently highlighted utterance (in yellow); (h) a rectangular glyph representing the stance categories detected in the

utterance; and (i) a warning symbol (exclamation mark) representing low classiﬁcation conﬁdence.

volving exploration of longer documents such as busi-

ness reports (Kearney and Liu, 2014) and works of

literature (Sinclair and Rockwell, 2016). Our visu-

alization approach, called DoSVis (Document Stance

Visualization), uses the output of the automatic stance

classiﬁer developed as part of the StaViCTA project to

provide the users with an environment for exploring

the individual documents’ contents, annotated with

the stance categories detected at the utterance or sen-

tence level (see Figure 1). The main contributions of

this paper are the following:

• a visualization approach for individual text docu-

ments that supports visual stance analysis; and

• a demonstration of application scenarios for vi-

sual stance analysis in several data domains.

The rest of this article is organized as follows. In the

next section, we shortly describe the background of

stance analysis and existing approaches for stance vi-

sualization as well as text document visualization. Af-

terwards, we discuss our visualization methodology

in Section 3. We illustrate the applicability of our ap-

proach with several use cases in Section 4 and discuss

some aspects of our ﬁndings in Section 5. Finally, we

conclude this article in Section 6.

2 RELATED WORK

2.1 Stance Analysis and Visualization

A more conservative approach to automatic stance

analysis of textual data focuses on the detection of

agreement/disagreement or pro/contra positions of

the author, typically towards the given topic or tar-

get (Skeppstedt et al., 2016b; Mohammad et al.,

2016). The latter work describes the results of a

stance analysis contest for a Twitter data set with

the majority of submissions using support vector

machines (SVM) or neural networks as classiﬁers

and n-grams, word embeddings, and sentiment lexi-

cons as features. The same authors also introduce a

dashboard-style visualization of their stance data set

that provides a general overview, but does not focus

on the contents of individual documents. Another vi-

sualization approach for the analysis of speakers’ po-

sitions towards corresponding topics is ConToVi (El-

Assady et al., 2016). This approach is designed for

monitoring of political debates, and it also focuses on

the overall trends and topics rather than the text con-

tent.

There also exist other approaches that focus on a

wider set of categories related to stance, such as cer-

tainty/uncertainty (Kucher et al., 2016b) or specula-

tion and condition (Skeppstedt et al., 2016a). Kucher

DoSVis: Document Stance Visualization

169

et al. describe a visualization of their stance data

set with a tool called ALVA (Kucher et al., 2016a;

Kucher et al., 2017b). Similar to the other stance vi-

sualizations discussed above, ALVA focuses on the

overview of a data set or corpus consisting of multi-

ple utterances or sentences from blog posts and com-

ments. Finally, StanceXplore (Martins et al., 2017)

provides multiple coordinated views for exploratory

visual analysis of a corpus of tweets labelled with

multiple stance categories by a stance classiﬁer. In

contrast to all these works, our contribution proposed

in this paper is designed for a detailed exploration of

individual documents which are much larger/longer

than social media posts.

2.2 Visualization of Individual Text

Documents

The existing taxonomies of text visualization tech-

niques recognize individual documents as one of the

options of data sources as opposed to corpora (J

anicke

et al., 2015) or text streams (Kucher and Kerren,

2015; Kucher et al., 2017a), for instance. A typical

example of such a document is a work of literature

which can be explored by a scholar in Digital Hu-

manities using a software tool with some form of sup-

port for visualization (Drucker, 2016). Providing an

overview of the content of individual documents dates

back to early techniques, such as SeeSoft (Eick et al.,

1992) and TileBars (Hearst, 1995). Both provide

pixel-based summaries for text segments constituting

the documents. Affect Color Bar (Liu et al., 2003) im-

plements a similar idea, but uses categories related to

emotions. The resulting visualization allows the user

to get an overview of the affective structure of a text,

such as a novel, and to navigate to the correspond-

ing segment for close reading. Ink Blots (Abbasi

and Chen, 2007) is a technique based on highlighting

regions of text documents with background bubble

plots. The resulting bubble plots can be used without

the actual text content for overview purposes. Keim

and Oelke describe a compact pixel-based technique

which can use various text features to represent visual

ﬁngerprints of text segments (Keim and Oelke, 2007).

VarifocalReader (Koch et al., 2014) supports both dis-

tant and close reading (see (J

anicke et al., 2015), for

example) by using topic segmentation, overview of

text structure, and highlighting of automatically anno-

tated words or chunks. Lexical Episode Plots (Gold

et al., 2015) provide an overview of topics recur-

ring throughout a text (more speciﬁcally, a transcript

of political debates). uVSAT (Kucher et al., 2016b)

uses scatterplot-like representations for overviews of

stance markers detected in a text document. Finally,

Chandrasegaran et al. implement an interactive inter-

face for visual analysis and open coding annotation

of textual data, which includes structural overviews

for distant reading and colored text view for close

reading (Chandrasegaran et al., 2017). Our approach

adopts ideas similar to many of such visualization

techniques in order to provide an overview of stance

classiﬁcation results for an individual document at

the utterance level. In contrast to some of the tech-

niques discussed above, though, our goal is to pre-

serve the two-way mapping between utterances and

visual items used in the overview, so that the users

could refer to the overview while performing close

reading.

Many existing techniques which provide support

for close reading use a certain form of highlighting in-

dividual words or chunks of text (Strobelt et al., 2016)

to represent custom annotations or labels. For ex-

ample, Ink Blots (Abbasi and Chen, 2007) highlight

an approximate region based on the position of cer-

tain marker words or features. Serendip (Alexander

et al., 2014) highlights words relevant to speciﬁc top-

ics. uVSAT (Kucher et al., 2016b) highlights words

and n-grams from the lists of stance marker words

and topic terms. Chandrasegaran et al. provide the

user with controls for highlighting speciﬁc parts of

speech and information content in the detailed text

view of their interactive interface (Chandrasegaran

et al., 2017). As opposed to these approaches, our

goal for representing the textual content of documents

is to support the output of a stance classiﬁer with mul-

tiple non-exclusive categories. Therefore, we use a

strategy relying on non-intrusive glyphs rather than

direct highlighting of the text to represent the classiﬁ-

cation results.

3 VISUALIZATION

METHODOLOGY

The input data for our tool DoSVis is generated by

a stance classiﬁer pipeline currently developed by

our project members (Kucher et al., 2016a; Kucher

et al., 2017b; Simaki et al., 2017b; Skeppstedt et al.,

2017a). The pipeline (see an illustration in Figure 2)

divides the input text into utterances and then classi-

ﬁes each utterance with regard to a set of stance cat-

egories such as uncertainty, hypotheticals, and pre-

diction. The tasks related to the set of stance cate-

gories, the data annotation process, and the training

of the classiﬁer were carried out in collaboration with

our experts in linguistics and computational linguis-

tics. The stance categories used by the classiﬁer are

not mutually exclusive, i.e., several categories may be

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

170

Figure 2: The architecture of our approach. DoSVis uses the output of the stance classiﬁer for a text document divided into

utterances. Each utterance may be simultaneously labelled with multiple stance categories.

simultaneously detected in any given utterance. Our

approach can actually be generalized to any set of

categories or labels associated with utterances. We

have tested this by using two versions of the stance

classiﬁer: (1) an SVM-based classiﬁer with 10 stance

categories (Kucher et al., 2017b), and (2) a logistic

regression (LR)-based classiﬁer with 12 stance cate-

gories (Skeppstedt et al., 2017a). Both of these clas-

siﬁers also provide a form of conﬁdence estimates for

the classiﬁcation decisions based on (1) Platt scal-

ing (Platt, 1999) and (2) probability estimates (Hos-

mer et al., 2013), respectively. After the initial pre-

processing and classiﬁcation stages, the input data for

the visualization module consists of a JSON ﬁle with

an array of utterances labelled with classiﬁcation re-

sults.

Our approach is based on a rather straightforward

visual design in order to be intuitive to the users with-

out prior training in visualization. DoSVis is imple-

mented as a web-based system using JavaScript and

D3 (D3, 2011). Its user interface depicted in Figure 1

provides an overview and a detailed text view for the

selected document. The users can control the inter-

pretation of line break symbols to adjust the document

layout, which can be preferable in case of some docu-

ments converted from the PDF format (see Section 5).

The sliders located at the top right (see Figure 1(a))

specify the classiﬁcation conﬁdence thresholds for

displaying the classiﬁcation results at all and display-

ing warning symbols (exclamation marks within the

glyphs, see Figure 1(i)), respectively, in order to help

the users focus on more reliable results.

The overview of stance classiﬁcation results con-

sists of scatterplot-like representations for individual

stance categories displayed in Figure 1(b). We have

decided to follow this design with separate represen-

tations for categories due to the data considerations

described above. Any utterance in our data can po-

tentially be labelled with up to 10 or 12 stance cat-

egories simultaneously, therefore, alternative designs

would have to use overly complex glyphs or ignore

the resulting categories to some extent (Kucher et al.,

2017b; Martins et al., 2017). Each utterance with

a detected stance category is represented by a dot

marker in the corresponding overview plot. The dot

position itself reﬂects the position of the utterance in

the text. More speciﬁcally, the position is based on

the coordinates of the HTML element representing

the utterance relative to the overall text view HTML

container. Each stance category is associated with a

certain color based on the color maps from Color-

Brewer (ColorBrewer, 2009). The opacity of the dot

is based on the classiﬁcation conﬁdence value. Visual

items with conﬁdence values below the global thresh-

old are hidden. The overview plots support pan &

zoom for the vertical axis, and the default zoom level

is set to ﬁt the complete document text. The area cur-

rently visible in the main text view is represented by

a viewport rectangle in each plot (see Figure 1(c)).

Each overview supports details on demand and navi-

gation over the text by hovering and clicking, respec-

tively. The users can also hide the overview plots and

navigate to the previous/next occurrence of the corre-

sponding stance category by using the buttons located

under each plot (see Figure 1(e)).

Besides the interactions with a single overview

plot, the users can drag-and-drop the plots onto each

other. This results in a new plot providing the

overview of utterances which are labelled with the

corresponding combination of categories. Such plots

for the combinations of two and three categories, re-

spectively, are displayed in Figure 1(d). In order to

distinguish such combination plots from regular cate-

gory overview plots, we have used rectangular mark-

ers with a dark grey color. The opacity mapping

and global ﬁltering behaviour for the visual items

are based on the lowest conﬁdence value with re-

gard to the category combination. Such combination

overview plots support the same interactions as regu-

lar category overview plots, except for the “hide” but-

ton being replaced by the “remove” button (cf. Fig-

ure 1(e+f)).

DoSVis also provides a detailed text view (dis-

played in Figure 1(g)) with stance category labels and

details on demand, thus supporting both distant and

close reading approaches (J

anicke et al., 2015). We

use sets of non-intrusive rectangular glyphs located

above utterances to represent the categories detected

DoSVis: Document Stance Visualization

171

(a) Tableau Software 2015 annual report.

(b) Yahoo Inc. 2015 annual report.

Figure 3: Overviews of stance categories detected in several documents with the LR classiﬁer at 66% classiﬁcation conﬁdence.

by the classiﬁer (see Figure 1(h)). These glyphs share

the color coding, opacity mapping, and ﬁltering be-

haviour with the overview plots. They are also con-

nected with linking and brushing—see the elements

highlighted in yellow in Figure 1(b+d+g). One addi-

tional design element used for the glyphs in the main

text view is a low conﬁdence warning represented

by an exclamation mark, as depicted in Figure 1(i).

Such marks are displayed for the classiﬁcation results

with conﬁdence values lower than the global thresh-

old controlled by the corresponding slider.

4 USE CASES

As mentioned in Section 1, we focus on use cases be-

yond social media monitoring. One of them is the

exploration of business reports: an analyst or an in-

vestor may be interested not only in the reported ﬁ-

nancial results, but also in the language used through-

out the report. Our tool DoSVis could be used in

this case to explore the results of automatic stance

analysis similar to the existing application of senti-

ment analysis (Kearney and Liu, 2014; Nopp and

Hanbury, 2015). The users would beneﬁt from the

opportunity to get an overview for the complete text

and to navigate between stance occurrences to explore

such longer texts in detail and verify the classiﬁca-

tion results. For example, the PDF versions of the

2015 annual reports from Tableau Software and Ya-

hoo Inc. contain 98 and 180 pages, respectively. Their

overviews in DoSVis are displayed in Figure 3(a+b) at

the selected classiﬁcation conﬁdence level of 66%. It

is interesting to note that both reports contain a rather

large number of expressions of uncertainty which is

detected in approximately 8% of utterances in both

cases. The density of such expressions is particu-

larly high in the early sections of the reports where

forward-looking statements are located. The occur-

rences of uncertainty combined with hypotheticals

or prediction are mainly found in the same regions

of the text. The comparison between the two docu-

ments with regard to speciﬁc categories reveals that

the Tableau Software report has a larger proportion

of detected hypotheticals (3.79% vs 2.67% of utter-

ances) and need & requirement (5.01% vs 3.08%)

than the Yahoo Inc. report, and a lower proportion

of prediction (1.00% vs 3.91%). It is also interesting

to note that categories such as agreement, disagree-

ment, tact, and rudeness are almost absent in the re-

sults, which can be explained by the genre of these

documents.

Another application of our approach is related to

the exploration of works of literature. Scholars in

digital humanities (Schreibman et al., 2016) could

make use of the support for distant and close read-

ing provided by DoSVis. Figure 3(c) displays an

overview of Arthur Conan Doyle’s “The Hound of

the Baskervilles” and provides the user with a gen-

eral impression of the stance category occurrences in

the text. In contrast to the ﬁnancial reports described

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

172

above, it is easy to notice that the novel contains much

more occurrences of categories such as certainty, dis-

agreement, and tact. Our approach could, therefore,

be interesting to the scholars in digital humanities and

linguistics with regard to the analysis of differences

between genres of text by using category overviews

as sort of a ﬁngerprint (Keim and Oelke, 2007). Fur-

thermore, the scholars could make use of the opportu-

nity to analyze occurrences of stance category com-

binations by drag-and-dropping the overview plots.

Several recent papers on stance analysis (Simaki

et al., 2017a; Skeppstedt et al., 2017b) discuss co-

occurrences of such stance categories as prediction

with uncertainty and hypotheticals with uncertainty,

respectively, in political blog data. Figure 4 provides

an overview of corresponding category combinations

in “The Hound of the Baskervilles”, which can be

interesting to the researchers in Digital Humanities.

The user can immediately get insights about the dis-

tribution of these stance category combinations, e.g.,

there are just two instances of prediction with un-

certainty, and no occurrences of combinations of all

three categories are detected at the current classiﬁ-

cation conﬁdence level. By clicking visual items or

using the navigation buttons, the user can then navi-

gate to the corresponding utterances for close reading.

In this case, exploratory analysis with DoSVis would

allow the user to identify concrete interesting cases

as opposed to interpreting overall category statistics

computed with non-interactive analyses.

Figure 4: Overviews of several stance category combina-

tions detected for the data in Figure 3(c).

5 DISCUSSION

Stance Classiﬁcation. The existing methods of

automatic stance classiﬁcation do not reach the same

levels of precision/accuracy (Mohammad et al., 2016)

as, for instance, sentiment classiﬁcation methods,

especially for topic-independent tasks (Skeppstedt

et al., 2016a). This raises concerns related to the

users’ trust in classiﬁcation results and the corre-

sponding visualization, especially when low conﬁ-

dence values are reported by the classiﬁer. Neverthe-

less, our proposed visualization approach allows the

users to explore the classiﬁcation results in detail and

make the ﬁnal judgment themselves. DoSVis can also

easily make use of improved classiﬁers available in

the future.

Preprocessing. In order to apply our approach to

the analysis of various reports and books available as

PDF documents, text data must be extracted and clas-

siﬁed utterance after utterance. For longer documents,

manual preprocessing is not feasible, and automatic

conversion of PDF to plain text often results in noisy

or almost unusable data (Constantin et al., 2013). It

would also be desirable to preserve the original lay-

out of document pages in many cases. We consider

this as part of the future work which could be based

on the previously described approaches (Mao et al.,

2003; Strobelt et al., 2009).

Scalability. We have tested DoSVis with docu-

ments of several sizes/lengths, the longest being the

2017 Economic Report of the President of the US

(599 pages). Our tool is able to display the corre-

sponding classiﬁcation results, albeit the performance

of some interactions is rather low. The largest delays

are caused by the web browser’s layout events for the

main text view. The potential solution is to avoid dis-

playing the complete document text in such cases and

use some form of sectioning instead—for instance,

Asokarajan et al. propose a visualization strategy re-

lying on multiple text scales (Asokarajan et al., 2016;

Asokarajan et al., 2017). As for the other scalabil-

ity concerns, the overviews for such large documents

are affected by overplotting. Our current implementa-

tion relies on pan & zoom to allow the users focus on

shorter text segments and avoid this effect. Alterna-

tive solutions could involve some forms of semantic

zooming, although it could potentially affect other in-

teractions.

6 CONCLUSIONS AND FUTURE

WORK

In this paper, we have demonstrated how stance clas-

siﬁcation results can be used for visual exploration of

a text document such as a business report or a novel.

We have described our tool DoSVis which provides an

interactive visualization of multiple stance categories

detected in the text. DoSVis can be used to estimate

the number of utterances with detected stance in a

given text, compare the results for several stance cat-

egories, and explore the text in detail. With the stance

classiﬁcation accuracy improving over time, we be-

lieve that such an approach will be useful for scholars

DoSVis: Document Stance Visualization

173

and practitioners, as illustrated by our potential use

cases. We plan to provide our prototype to the expert

users in order to get their feedback and reﬁne our im-

plementation. Our plans for further development of

DoSVis also include a user study in order to evaluate

some of our design decisions.

While DoSVis focuses on individual text docu-

ments, our future work includes the development of

novel visual representations for stance detected in

text corpora, temporal and streaming text data, and

text data associated with geospatial and relational at-

tributes.

ACKNOWLEDGEMENTS

This research was funded by the framework grant

“The Digitized Society—Past, Present, and Future”

with No. 2012-5659 from the Swedish Research

Council.

REFERENCES

Abbasi, A. and Chen, H. (2007). Categorization and anal-

ysis of text in computer mediated communication

archives using visualization. In Proceedings of the 7th

ACM/IEEE-CS Joint Conference on Digital Libraries,

JCDL ’07, pages 11–18. ACM.

Alexander, E., Kohlmann, J., Valenza, R., Witmore, M., and

Gleicher, M. (2014). Serendip: Topic model-driven

visual exploration of text corpora. In Proceedings of

the IEEE Conference on Visual Analytics Science and

Technology, VAST ’14, pages 173–182.

Asokarajan, B., Etemadpour, R., Abbas, J., Huskey, S., and

Weaver, C. (2016). Visualization of Latin textual vari-

ants using a pixel-based text analysis tool. In Proceed-

ings of the EuroVis Workshop on Visual Analytics, Eu-

roVA ’16. The Eurographics Association.

Asokarajan, B., Etemadpour, R., Abbas, J., Huskey, S.,

and Weaver, C. (2017). TexTile: A pixel-based fo-

cus+context tool for analyzing variants across multi-

ple text scales. In Short Papers of the EG/VGTC Con-

ference on Visualization, EuroVis ’17. The Eurograph-

ics Association.

Chandrasegaran, S., Badam, S. K., Kisselburgh, L., Ra-

mani, K., and Elmqvist, N. (2017). Integrating vi-

sual analytics support for grounded theory practice in

qualitative text analysis. Computer Graphics Forum,

36(3):201–212.

ColorBrewer (2009). ColorBrewer 2.0 — color advice for

cartography. http://colorbrewer2.org/. Accessed Octo-

ber 31, 2017.

Constantin, A., Pettifer, S., and Voronkov, A. (2013).

PDFX: Fully-automated PDF-to-XML conversion of

scientiﬁc literature. In Proceedings of the ACM

Symposium on Document Engineering, DocEng ’13,

pages 177–180, New York, NY, USA. ACM.

D3 (2011). D3 — data-driven documents. http://d3js.org/.

Accessed October 31, 2017.

Drucker, J. (2016). Graphical approaches to the digital

humanities. In Schreibman, S., Siemens, R., and

Unsworth, J., editors, A New Companion to Digital

Humanities, pages 238–250. John Wiley & Sons.

Eick, S. G., Steffen, J. L., and Sumner, E. E. (1992).

Seesoft—A tool for visualizing line oriented software

statistics. IEEE Transactions on Software Engineer-

ing, 18(11):957–968.

El-Assady, M., Gold, V., Acevedo, C., Collins, C., and

Keim, D. (2016). ConToVi: Multi-party conversa-

tion exploration using topic-space views. Computer

Graphics Forum, 35(3):431–440.

Gold, V., Rohrdantz, C., and El-Assady, M. (2015). Ex-

ploratory text analysis using lexical episode plots. In

Short Papers of the EG/VGTC Conference on Visual-

ization, EuroVis ’15. The Eurographics Association.

Hearst, M. A. (1995). TileBars: Visualization of term dis-

tribution information in full text information access.

In Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, CHI ’95, pages 59–66.

ACM Press/Addison-Wesley Publishing Co.

Hosmer, Jr., D. W., Lemeshow, S., and Sturdivant, R. X.

(2013). Applied Logistic Regression. John Wiley &

Sons, Inc.

anicke, S., Franzini, G., Cheema, M. F., and Scheuermann,

G. (2015). On close and distant reading in digital hu-

manities: A survey and future challenges. In Proceed-

ings of the EG/VGTC Conference on Visualization —

STARs, EuroVis ’15. The Eurographics Association.

Kearney, C. and Liu, S. (2014). Textual sentiment in ﬁ-

nance: A survey of methods and models. Interna-

tional Review of Financial Analysis, 33:171–185.

Keim, D. A. and Oelke, D. (2007). Literature ﬁngerprint-

ing: A new method for visual literary analysis. In Pro-

ceedings of the IEEE Symposium on Visual Analytics

Science and Technology, VAST ’07, pages 115–122.

Koch, S., John, M., W

orner, M., M

uller, A., and Ertl, T.

(2014). VarifocalReader — In-depth visual analysis

of large text documents. IEEE Transactions on Visu-

alization and Computer Graphics, 20(12):1723–1732.

Kucher, K. and Kerren, A. (2015). Text visualization tech-

niques: Taxonomy, visual survey, and community in-

sights. In Proceedings of the 8th IEEE Paciﬁc Visu-

alization Symposium, PaciﬁcVis ’15, pages 117–121.

IEEE.

Kucher, K., Kerren, A., Paradis, C., and Sahlgren, M.

(2016a). Visual analysis of text annotations for stance

classiﬁcation with ALVA. In Poster Abstracts of the

EG/VGTC Conference on Visualization, EuroVis ’16,

pages 49–51. The Eurographics Association.

Kucher, K., Paradis, C., and Kerren, A. (2017a). The state of

the art in sentiment visualization. Computer Graphics

Forum.

Kucher, K., Paradis, C., Sahlgren, M., and Kerren, A.

(2017b). Active learning and visual analytics for

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

174

stance classiﬁcation with ALVA. ACM Transactions

on Interactive Intelligent Systems, 7(3):14:1–14:31.

Kucher, K., Schamp-Bjerede, T., Kerren, A., Paradis, C.,

and Sahlgren, M. (2016b). Visual analysis of online

social media to open up the investigation of stance

phenomena. Information Visualization, 15(2):93–116.

Liu, H., Selker, T., and Lieberman, H. (2003). Visualizing

the affective structure of a text document. In CHI ’03

Extended Abstracts on Human Factors in Computing

Systems, CHI EA ’03, pages 740–741. ACM.

Mao, S., Rosenfeld, A., and Kanungo, T. (2003). Docu-

ment structure analysis algorithms: A literature sur-

vey. SPIE Proceedings, 5010.

Martins, R. M., Simaki, V., Kucher, K., Paradis, C., and

Kerren, A. (2017). StanceXplore: Visualization for

the interactive exploration of stance in social media.

In Proceedings of the 2nd Workshop on Visualization

for the Digital Humanities, VIS4DH ’17.

Mohammad, S. M. (2016). Sentiment analysis: Detecting

valence, emotions, and other affectual states from text.

In Meiselman, H. L., editor, Emotion Measurement,

pages 201–237. Woodhead Publishing.

Mohammad, S. M., Kiritchenko, S., Sobhani, P., Zhu, X.,

and Cherry, C. (2016). SemEval-2016 task 6: De-

tecting stance in tweets. In Proceedings of the In-

ternational Workshop on Semantic Evaluation, Se-

mEval ’16.

Nopp, C. and Hanbury, A. (2015). Detecting risks in the

banking system by sentiment analysis. In Proceed-

ings of the 2015 Conference on Empirical Methods

in Natural Language Processing, EMNLP ’15, pages

591–600. Association for Computational Linguistics.

Pang, B. and Lee, L. (2008). Opinion mining and sentiment

analysis. Foundations and Trends in Information Re-

trieval, 2(1–2):1–135.

Platt, J. C. (1999). Probabilistic outputs for support vector

machines and comparisons to regularized likelihood

methods. In Smola, A. J., Bartlett, P. L., Sch

olkopf,

B., and Schuurmans, D., editors, Advances in Large

Margin Classiﬁers, pages 61–74. MIT Press.

Schreibman, S., Siemens, R., and Unsworth, J. (2016). A

New Companion to Digital Humanities. John Wiley

& Sons.

Simaki, V., Paradis, C., and Kerren, A. (2017a). Stance clas-

siﬁcation in texts from blogs on the 2016 British ref-

erendum. In Proceedings of the International Confer-

ence on Speech and Computer, SPECOM ’17, pages

700–709. Springer.

Simaki, V., Paradis, C., Skeppstedt, M., Sahlgren, M.,

Kucher, K., and Kerren, A. (2017b). Annotating

speaker stance in discourse: The Brexit Blog Corpus.

Corpus Linguistics and Linguistic Theory.

Sinclair, S. and Rockwell, G. (2016). Text analysis and

visualization. In Schreibman, S., Siemens, R., and

Unsworth, J., editors, A New Companion to Digital

Humanities, pages 274–290. John Wiley & Sons.

Skeppstedt, M., Kucher, K., Paradis, C., and Kerren, A.

(2017a). Language processing components of the

StaViCTA project. In Proceedings of the Workshop

on Logic and Algorithms in Computational Linguis-

tics, LACompLing ’17, pages 137–138.

Skeppstedt, M., Sahlgren, M., Paradis, C., and Kerren, A.

(2016a). Active learning for detection of stance com-

ponents. In Proceedings of the Workshop on Com-

putational Modeling of People’s Opinions, Personal-

ity, and Emotions in Social Media at COLING ’16,

PEOPLES ’16, pages 50–59. Association for Compu-

tational Linguistics.

Skeppstedt, M., Sahlgren, M., Paradis, C., and Kerren, A.

(2016b). Unshared task: (Dis)agreement in online de-

bates. In Proceedings of the 3rd Workshop on Argu-

ment Mining at ACL ’16, short papers track, ArgMin-

ing ’16, pages 154–159. Association for Computa-

tional Linguistics.

Skeppstedt, M., Simaki, V., Paradis, C., and Kerren, A.

(2017b). Detection of stance and sentiment modiﬁers

in political blogs. In Proceedings of the International

Conference on Speech and Computer, SPECOM ’17,

pages 302–311. Springer.

Strobelt, H., Oelke, D., Kwon, B. C., Schreck, T., and Pﬁs-

ter, H. (2016). Guidelines for effective usage of text

highlighting techniques. IEEE Transactions on Visu-

alization and Computer Graphics, 22(1):489–498.

Strobelt, H., Oelke, D., Rohrdantz, C., Stoffel, A., Keim,

D. A., and Deussen, O. (2009). Document Cards:

A top trumps visualization for documents. IEEE

Transactions on Visualization and Computer Graph-

ics, 15(6):1145–1152.

DoSVis: Document Stance Visualization

175