PreechVis: Visual Profiling using Multiple-word Combinations

Seongmin Mun

, Gyeongcheol Choi

, Guillaume Desagulier

and Kyungwon Lee

Sciences du Langage (Computational Linguistic), Life Media Interdisciplinary Program, UMR 7114 MoDyCo - CNRS,

University Paris Nanterre, Ajou University, 200 Avenue de la Re

publique 92000, Paris, France

Life Media Interdisciplinary Program, Ajou University, 206, World cup-ro, Yeongtong-gu, Suwon, Republic of Korea

English Grammar and Linguistics, UMR 7114 MoDyCo - University Paris 8, CNRS, University Paris Nanterre,

200 Avenue de la Re

publique, 92000, Paris, France

Digital Media, Ajou University, 206, World cup-ro, Yeongtong-gu, Suwon, Republic of Korea

Keywords: Data Clustering, Data Filtering, Hierarchy Data, Text and Document Data, Interaction Design, Coordinated

and Multiple Views, Task and Requirement Analyses, Visualization in the Humanities.

Abstract: Words in the corpus include features and information, and the visualizing of such words can improve the

user’s understanding of them. Words in text corpus may be consist of one-word or they may be a combination

of words that together, constitute a word. The latter is referred to as a multiword expression. And if we analyze

both single word and multiword with visualization, we can get more accurate results and more information

than we analyze only single word from corpus. An interactive visualization can be useful for analyzing

multiword expressions, because the following features are of interest to linguistics scholars: (1) Showing the

combinatory POS pattern of a hierarchical form, (2) exploring results according to the POS pattern, and (3)

searching the source corpus for the analysis-result verification. Therefore, we propose PreechVis, an

interactive-visualization tool that includes all of the requisite functions for an analysis for which multiple

words (http://202.30.24.167:3010/PreechVisMWE) are utilized. For the present study, we used a total of 957

speeches of 43 U.S. Presidents from George Washington to Barack Obama as the corpus data. PreechVis is

divided into two views. In the first view, the system consists of a combination of Sunburst and RadVis.

Through the circular Sunburst, we present the POS and its combination patterns for each gram. In RadVis,

the Presidents were positioned according to their frequency value. In addition, when the President was

selected, the frequency value was displayed on Sunburst to improve the user’s understanding. In the second

view, the user can simultaneously confirm and verify the details of the result using the Wordcloud. The two

different views are synchronized with each other and are changed by the selected grams, issues, and

Presidents. In the experiments and case studies on the U.S.-President speech data, we verified the effectiveness

and usability of PreechVis.

1 INTRODUCTION

A visual analysis of textual data can support users in

the attainment of a general understanding of the

information of the text without the user having to

actually read through the text. This can be very

helpful for tasks with large volumes of text. Word-

based research is very common for visual corpus

analyses (Lu et al.,2016), (Wang et al., 2016),

(Heimerl et al., 2016).

The words in the corpus include features and

information, and visualizing a word helps users to

understand it easily. These words can be split into two

types. One type is a word with a one-word meaning,

whereas the second type is a word with a combined-

word meaning. The second word is called multiword

and is generated by a combination of different parts-

of-speech (POS) (Ramisch, 2015). Simply, the

multiword is a habitual recurrent word combination

of everyday language (JR, 1957).

For example, when people say that someone sets

the bar high, it is understood as a metaphor that his or

her competitors will find it difficult to win against

him or her. If we analyze both single word and

multiword with visualization, we can get more

accurate results and more information than we

analyze only single word from corpus.

An interactive visualization can be useful for

analyzing multiword expressions, because the

following features are of interest to linguistics

Mun, S., Choi, G., Desagulier, G. and Lee, K.

PreechVis: Visual Proﬁling using Multiple-word Combinations.

DOI: 10.5220/0006615500970107

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

97-107

ISBN: 978-989-758-289-9

scholars: (1) Showing the POS-combination pattern

of the hierarchical form, (2) exploring results

according to the POS pattern, and (3) searching the

source corpus for the analysis-result verification.

Therefore, PreechVis, an interactive-visualization

tool that covers the necessary functions for the

exploration of larger amounts of multiword corpus

information, has been created for this study. This

work provides the following contributions: (1) The

two types of corpus-data words for the exploration of

more information and accurate results is presented,

(2) an interactive visualization and multifunctional

tool for the visual analysis is presented utilizing the

multiword type, and (3) the visualization tool is

assessed via case studies for which the U.S.

Presidential Address is used to verify the utility of

PreechVis.

2 RELATED WORKS

For this work, it is assumed that with a textual

multiword analysis, the information is easier to

discover and to analyze through an interactive

visualization. Supporting this hypothesis, Carlos

Ramisch summarized a textual-analysis technique for

multiword expressions in a recent book (Ramisch,

2015). Further, many studies on the visual tools for

textual analyses have been conducted (Koch et al.,

2014), (Sun et al.,2014).

Numerous visualizations have been created to

extract and explore more information in large corpus

data. For a number of visualizations such as EvoRiver

(Sun et al., 2014) and OpinionFlow (Wu et al., 2014),

a word-based analysis is employed regardless of the

multiword combinations. The focus here is the word-

based multiword analysis and the defining of how the

multiword result can be presented in an interactive

visualization. Much research has been conducted on

corpus-data visualization.

2.1 Word-based Corpus Visualization

Word-based corpus visualization, which aims to

understand and explore words-based text corpora, has

received considerable attention in recent years (Sun

et al., 2014), (Cui et al., 2014a), (Wu et al., 2014).

EvoRiver(Sun et al., 2014) is a time-based

visualization that allows users to explore

competition-related interactions and to detect

dynamically evolving patterns, as well as their major

causes. Cui et al. (Cui et al., 2014a) presented an

interactive visual textual-analysis approach that

allows users to progressively explore and analyze the

complex evolutionary patterns of hierarchical topics.

Wu et al. (Wu et al., 2014) introduced a visual-

analysis system called OpinionFlow to empower

analysts to detect opinion-propagation patterns and

glean insights. Further, for OpinionFlow, a Sankey

graph is combined with a tailored density map in one

view to visually convey the diffusion of opinions

among many users.

These related works focus on visual explorations

of words without a considering multiword

expressions. Whereas, our present work includes

multiword expressions in its word-based illustrations.

2.2 Visual Graph Comparison

A visual-graph comparison aims to analyze the

similarities and differences between variables. A

number of visual-graph- comparison methods have

been proposed by many studies (Andrews et al.,

2009), (Cui et al., 2014b), (Collins and Carpendale,

2007).

Andrews et al. (Andrews et al., 2009) presented a

technique and a prototype tool to support the visual

comparison of graphs and the interactive

reconciliation of candidate graphs into a single

reference graph. Cui et al. (Cui et al., 2014b)

introduced a novel flow-based visualization design

for the summarization of high-level evolution

patterns in a dynamic graph. Collins et al. (Collins

and Carpendale,2007) described VisLink, a

visualization environment in which one can display

multiple two-dimensional (2D) visualizations,

reposition and reorganize them in a three-dimensional

(3D) form, and display the relationships between

them by propagating the edges from one visualization

to another.

These works assume that visual graph

comparisons can help users understand the

similarities and differences between variables. In this

paper, the PreechVis allows for wider comparison,

across analysis results extracted from multiple

corpora based on user selections.

2.3 Verification of the Visual Findings

A visual analysis of the corpus data can help uses

understand the corpus without reading it. However, to

assess the utility of a visual-analysis tool, derived

insights must be verified through comparison with the

real corpus (Koch et al., 2014), (Stasko et al., 2007).

Koch et al. (Koch et al., 2014) presented a method

that supports visual-analytics tasks on large text

documents that is particularly useful in situations

where scrutiny is required and the textual source must

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

be used to verify the findings. Stasko et al. (Stasko et

al.,2007) developed a visual-analytics system called

Jigsaw that visually represents documents and their

entities to help analysts examine reports more

efficiently and to develop potential-action theories

more quickly. Further, this system provides multiple

coordinated views of the document entities with a

special emphasis on a visual illustration of the inter-

entity connections across different documents.

For the verification of visual findings, PreechVis

offers visual results with a real corpus to

simultaneously verify visual-analytics insights.

3 DESIGN CONSIDERATION

3.1 User Tasks

To understand the needs of the multiword analysis,

regular cooperation was sought from computational-

linguistic researchers to learn about their hypotheses

and goals in the study of the multiword. Also, they

were presented with a variety of designs, and

important feedback was collected from them to

iteratively refine the disambiguation results. The

experience revealed that researchers typically need to

accomplish the following tasks in their explorations

of corpus data:

▪ Task1: Exploration of the Word-

combinations Patterns for Each Gram.

Researchers are interested in this task to

discover the different word-combinations

patterns for each gram. N-gram is a contiguous

sequence of n items from a given sequence of

text or speech (Sidorov et al., 2013). In this

paper, the word-combination patterns from the

unigram to the trigram were analyzed.

▪ Task 2: Identification of the Groups of U.S.

Presidents that Exhibit Similar Usage

Patterns in the Multiword. The researchers

look for typical usage patterns in the multiword

for the purpose of generalization. To find such

multiword patterns, a group of Presidents that

exhibit similar POS usages were identified.

▪ Task 3: Verification of Visual Findings with

Real Corpus Data. A visual analysis of corpus

data can support users in their attainment of an

understanding of the corpus information.

However, visual-analysis results are required to

verify and prove the findings (Koch et al.,

2014).

3.2 Design Objectives

In response to the previously mentioned tasks, the

following Design objectives were built to guide the

proposed approach:

▪ Design Objective 1: Identify the Word-

combinations Patterns of the Multiword for

Each Gram. Computational-linguistic

researchers need to know what kind of

combinatory patterns they have and how

different those of each gram are.

▪ Design Objective 2: Easily Compare the

Multiword Usage Patterns Among the

Presidents. To support the interactive

comparison of the multiword usage patterns

among the Presidents (T.2), a multi-view

system needs to be included in the main view

for the visualization of the usage patterns for

each President and the side view for the

selection of the Presidents. Through this

interaction, the user should be able to select

each President, and the usage patterns will be

changed according to the selected President.

▪ Design Objective 3: Verification of the

Visual Findings Using the Real Corpus

Data. The interactive visualization provides

various supports for the exploration, analysis,

and understanding regarding the corpus data.

However, it difficult to obtain all of the

information in the corpus by using the

visualization analysis; for this reason, a view is

created for a corpus-based verification.

4 DATA PROCESSING

In this section, a data-processing structure for the

extraction of the information from corpus data is

presented. The datasets used in this paper are taken

from the Miller Center (https://millercenter.org/), one

of the representative databases of U.S. history and

civil discourse. During the data processing, each

piece of information in the database is extracted into

several descriptive attributes including the personal

information of each President, the public speeches of

the President, and pictures of the President.

PreechVis: Visual Proﬁling using Multiple-word Combinations

Figure 1: Structure of the Data Processing. Framework for

the Word Acquisition from the Corpus Data.

Figure1 summarizes the architecture of the

proposed data processing, which will be subsequently

described in detail.

4.1 Processing

This section is divided into two parts. The first part is

the pre-processing part, wherein Cleaning with

RegExp, Lemmatization, Tokenization, and

Lowercasing were conducted. Then, the N-gram

analysis and the POS tagging for the candidate

extraction were conducted regarding the word.

4.2 Candidate Extraction and

Filtration

Using the previously mentioned procedure, the N-

gram results were obtained with the POS-tagging

result. These results were counted according to the

frequency value, and the data were filtered through

the application of a threshold (frequency value greater

than or equal to 10). In addition, word candidates

were extracted without the stop-words for each gram.

For instance, in the case of the bigram, data like

“house i” and “power we” became stop words and

were removed from the word candidates.

4.3 Word Validation

In this section, the filtered word candidates are

verified using several English dictionaries.

The output of the word-candidate filtration is

required for the verification. Also for the verification,

an algorithmic working base was developed on

several English dictionaries, as shown in figure 2. The

proposed algorithm automatically compares the

results to several English dictionaries, and if the

dictionary shows a result, it returns the word

candidate. The primary validated-word candidate was

manually verified by the computational-linguistic

researchers.

Figure 2: Process of the Algorithmic Working Base on

Several English Dictionaries.

The word candidate was extracted from single

words, bi-grams, and tri-grams according to the

previously described procedure; that is, 45995

candidates were extracted from the uni-gram, 729552

candidates were extracted from the bi-gram, and

2089617 candidates were extracted from the tri-gram.

Among them, 8910 in the uni-gram, 901 in the bi-

gram, and 301 in the tri-gram were extracted as

validated words with meaning and were analyzed.

4.4 Word-combination Patterns

The computational-linguistic researchers need to find

out the different word-combination patterns for each

gram. For this purpose, the word results with the POS

tagging were produced in the processing section.

Each of the validated words comprise a different POS

combination. There are 8 large and 36 detailed word

combinations in the uni-gram, 50 large and 200

detailed word combinations in the bi-gram, and 65

large and 97 detailed word combinations in the tri-

gram.

4.5 Applying Term-weighting

There are many ways to calculate weights, including

the Local mutual-information function, the

Logarithm function, the Entropy function, and the

term-frequency-inverse–document-frequency (TF-

IDF) function. The TF-IDF function was used to

identify the word combinations that are important to

a document in the corpus for each gram (Qu et

al.,2008), (Zhang et al., 2008). The TF-IDF formula

is as follows:

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

100

(1)

The main idea of the TF-IDF algorithm is as

follows: In a case where the words describe the

meaning of the sentence, the more times that a word

appears in a sentence, the greater the contribution;

furthermore, the greater the number of documents

wherein the words appear, the smaller that the word

result for a document contribution should be. By

applying the TF-IDF algorithm, users can quickly

locate a content of a document.

5 VISUALIZATION DESIGN

From the previously mentioned user tasks and design

objectives, an interactive visualization was designed

to extract further information from the U.S. President

Speech corpus. For this visualization, the display of

the POS-combination patterns and the extraction of

more accurate results were considered, as this is

sufficient for the answering of research questions in a

multiword study. The design of each visualization

component is introduced as follows.

5.1 Emphasis of the Word-combination

Pattern

A multiword is generated by a combination of the

different POS patterns that computational-linguistic

researchers search for. Therefore, it is important to

emphasize the multiword word-combination patterns.

In the example of figure4, Sunburst (Rodden, 2014)

is used to demonstrate this clearly.

Figure 3: Emphasis of the word-combination patterns with

Sunburst: (a) Presenting words in the uni-gram with the first

layer, (b) showing the multiword in the bi-gram, and (c) the

multiword in the tri-layer tri-gram.

Sunburst is a carrier of a spatial-information

visualization in a circular layout, and it will extend

outward with the increasing of the number of layers

(Liu and Wang, 2015). Further, its value is high in

terms of the exploration and analysis of the public

information for large data amounts as a typical

method for the visualization of hierarchical data. As

shown in figure3, a glyph was designed to represent

the word-combination patterns of words with larger

cost values.

5.2 Representation of the Usage

Patterns

Computational-linguistic researchers look for typical

multiword usage patterns. The combination pattern of

the POS is the hierarchical data, and the President’s

POS Usage Patterns are multivariate data.

Figure 4: PreechVis Visualization Design. Displaying the

information of Barack Obama by the usage frequency of the

POS in the uni-gram.

It is a very experimental task to present both of

them in one view, so for the design of the first view,

RadVis (Sharko et al., 2008), (Rubio-Snchez et al.,

2016) and Sunburst were combined to determine the

usage pattern of each President and their grouping.

For example, in figure4, the frequency bar that is

located on the above Sunburst shows the information

from Barack Obama’s speech regarding the POS type

that he used in his speech and his pattern (Mahyar and

Tory, 2014). And the nodes of the President that are

inside the circle show their groups according to the

usage frequency of the POS.

5.3 Verification for a Comparison with

the Real Corpus

Previous research has shown that the visual-analysis

result for corpus data is required to verify and prove

findings in a comparison with real documents. In the

example of figure5, this visual tool is synchronized

with the above first view so that it changes according

to the user selection, and so that it is simultaneously

verified using the corpus data.

PreechVis: Visual Proﬁling using Multiple-word Combinations

101

Figure 6: PreechVis Visualization Interface. The interface of the proposed visual system representing the corpus data about

the speeches of 43 U.S. Presidents from George Washington to Barack Obama.

Figure 5: PreechVis Visualization Design: (a) Showing the

real corpus and coloring by tying words, the (b)

visualization-analysis result as a Word Cloud, and the (c)

color legend and sorted results.

PreechVis uses a Word Cloud (b) to provide a

visually distinguishable overview of the President’s

speech. This visual method is useful for learning

about the number and kinds of words that are present

in the corpus data. In the side view (c), the user can

easily search the color legend about the used Word

Cloud and the sorted word results. Additionally,

PreechVis can also display the view of the real speech

for the verification (a) that is on the left side of

figure5.

5.4 PreechVis

Figure6 describes the main workspace of the

proposed visual system after the loading of all of the

Presidential speeches. The three buttons in the layer

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

102

headers (Figure6 (b)) provide the option to change the

word combinations of the word for each gram.

Additionally, the user can change the visual result by

selecting the right-side option (Figure6 (c)). This

makes the PreechVis approach very flexible because

the different visual results of any of the Presidents’

speeches can be viewed easily. The analyst can

therefore determine the answer of their research

question quickly.

Personal Information. Figure6 (a) presents the

information about the selected president, wherein the

information of George Washington, including the

picture, term, birth details, and political party, are

displayed. The user can easily change this

information using the several options on the right side

(Figure6 (c)).

Visual Result. This graph pane (Figure6 (b)) was

developed through a combining of Sunburst with

RadVis. In the circular Sunburst, the POS and its

combination patterns are presented for each gram.

And the Presidents that are located inside were

positioned according to their frequency value by the

POS or the POS-combination patterns of the word. In

addition, when the user selects a President, the

frequency value is displayed on Sunburst.

Selection. Figure6 (c) helps the user to change the

visual result according to their research question. This

view is divided into the following three views:

Political-party color, Select issues, and Select

president. The buttons in the layer headers (Figure6

(c)) provide the option to change the node color

according to the President’s political party. It is also

possible to remove the node color according to the

President’s political party in the visual result. The

President’s speech can be categorized into the

following 11 issues (Hughps, 2009), (Andrade and

Young, 1996): Economic growth, social welfare, U.S.

policies (war), health care, immigration,

humanitarian aid, protection of the U.S. (terror),

establishment of democracy, promotion of U.S.

strength, U.S. priorities, and all of the speeches.

Through this selection view, the user can derive a

visual result of his or her issue of interest.

Speech. Figure6 (d) presents the real speech

according to the user selection, and this view can

display the word through a highlighting of it in the

speech. A user can also highlight words or an

highlighting word and turn on or off the highlight

words in the speech (Figure5 (a)).

Word Cloud. Figure6 (e) presents the result words

using the Word Cloud (b) function to support the user

in his or her attainment of a better understanding of

the analysis results of the President’s speech. This can

be very helpful for the user’s learning of the number

and kinds of words that are present in the speech. And

figure6 (f) provides more information such as a color

legend about the words in Word Cloud and the sorted

results. Additionally, when the user hovers over each

word, it shows the support view including the POS

type and the word count.

6 EVALUATION AND CASE

STUDIES

Case studies were conducted to evaluate the

effectiveness of the proposed interactive visualization

and its usability. The authors worked with

computational-linguistic researchers who study the

multiword and possess expert knowledge on the

subject. They used PreechVis to find information

regarding their research questions and compared the

details from the U.S. President Speech corpus.

6.1 Case Studies

This section demonstrates the exploratory use of the

system regarding several research questions.

▪ Q1: What kind of multi-words exist and how

different are the POS-combination patterns

in each gram?

This question was answered using a proposal of

an interactive visual system that facilitates the display

of the POS-combination patterns of each gram. In the

example of figure7, each of the multi-words comprise

a different POS combination. Further, 8 large and 36

specific word combinations are evident in the uni-

gram, 50 large and 200 specific word combinations

are evident in the bi-gram, and 65 large and 97

specific word combinations are evident in the tri-

gram.

Figure 7: Word-combination Patterns by Part-of-Speech

(POS). (a) POS combination in the uni-gram, (b) POS

combination in the bi-gram, and (c) POS combination in the

tri-gram.

PreechVis: Visual Proﬁling using Multiple-word Combinations

103

▪ Q2: Can we find groups that are grouped or

divided by the POS and the POS

combination?

PreechVis presents the visual result from two

different data formats based on the previously

mentioned visual techniques.

Figure 8: Visual Result in the tri-gram. (a) Displaying the

President’s group by the usage pattern and (b) changing the

node color by the political party.

Figure8 represents all of the Presidents that are

inside the circle as nodes, where they are grouped in

a single formation. This indicates that the patterns of

the POS and the POS combination according to the

Presidents are very similar. From the colors according

to the political party, “Republican” is located in the

middle of the group and “Democrat” is located both

above and below; however, they are so close to each

other, a significant difference is not evident.

▪ Q3: Can we get more accurate results from

the U.S. President Speech corpus?

Figure 9: Analysis result of Harry Truman’s speech by: (a)

uni-gram and (b) bi-gram.

A serious error will occur in the analysis result if

a researcher uses words with only a one-word

meaning; for example, the word “United States,” and

this word frequently appears in the speeches.

However, if a multiword analysis is not used, the

words “United” and “States” will account for a large

proportion of the analysis results. The proposed

visual tool, however, has addressed this problem, as

shown in figure9.

6.2 Usage Scenario

In the following subsection, a usage scenario that

demonstrates the suitability of PreechVis for analysis

tasks is presented. Additionally, a usage case that

shows how the U.S. President’s public speeches can

be analyzed with the proposed visual system is

described. The analysis of the usage case illustrates

the effectiveness regarding a comparison of the

Presidents’ speeches.

For this part, the two U.S. Presidents George W.

Bush and Barack Obama were compared in terms of

two main issues. The first issue is “Health Care.”

Figure10 shows the visual result of the two Presidents

for Health Care. In the case of Bush, he used words

like “Medicare (32),” “coverage (20),” “legislation

(14),” and “help (13)” frequently in his Health Care

speech. Alternatively, Obama commonly used words

like “insurance (86),” “people (67),” “going (51),”

“plan (44),” and “president (43)” in his speech.

Figure 10: Comparison of Two Presidents for “Health

Care” in the uni-gram.

Figure11 represents the visual results of the

bigram and the trigram. The users can change the

visual result by using either of the three gram buttons

in the top of the main view. In the bigram result, Bush

and Obama frequently used “health care (20, 71)” in

their speeches. However, Obama used more

multiword (“health insurance (31),” “insurance

company (9),” “right thing (7),” etc.) regarding

Health Care than Bush. In the trigram, the proposed

system shows the multiword consisting of three

words of each President. In the case of Obama, “to

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

104

make sure (4)” and “in the way (3)” are frequently

evident, as well as his passion on the word of Health

Care.

Figure 11: Comparison of Two Presidents for “Health

Care” in the bi-gram and the tri-gram.

To summarize, Obama is evidently more

interested in health care than Bush according to a tally

of the words and the size of the Word Cloud.

The second issue for the usage scenario is

“Protection of the U.S. (terror).” The United States

has previously received many terrorist attacks,

including the “September 11 attacks” and the

“Oklahoma City bombings”. The protection of the

U.S. from terrorism is therefore a very important

issue for the country’s Presidents.

Figure 12: Comparison of Two Presidents for “Protection

of the U.S. (terror)” in the uni-gram.

PreechVis was used to compare the two

Presidents’ opinions on the issue “Protection of the

U.S. (terror).” Figure12 represents the comparative

analytic results between Bush and Obama on the

Protection of the U.S. (terror) issue. In the results of

Bush, words like “America (287),” “Iraq (285),”

“People (222),” and “Iraqi (142)” feature many times

in his speech. Alternatively, Obama commonly used

words such as “people (231),” “security (103),”

“America (98),” and “Israel (93)” in his speech. In

these results, the different countries that were

discussed in the issue Protection of the U.S. (terror)

are evident.

Figure 13: Comparison of Two Presidents for “Protection

of the U.S. (terror)” in the bi-gram and the tri-gram.

Figure13 shows the visual results of the bigram

and the trigram. In the visual results of Bush, terms

like “United States (77),” “Al Qaeda (61),” “Saddam

Hussein (50),” and “Middle East (48)” often appear

in his speech regarding the Protection of the U.S.

(terror). In particular, “Al Qaeda (61)” and “Saddam

Hussein (50)” appear frequently in the results.

Therefore, the user can easily identify the

organization and the person that are involved in this

case. In the visual results of Obama, terms like

“United States (76),” “Al Qaeda (50),” “Bin Laden

(21),” and “Middle East (17)” frequently appear. It is

also possible to recognize the change of the person

who is involved in the case depending on the

President. In the trigram result, Bush and Obama used

two of the same terms (“in the middle (22,10)” and

“Osama bin Laden (3,8)”) in their speeches.

However, “in the middle (22,10)” may be the result

of “in the Middle East,” which is a four-word

combination. Therefore, this matter needs to be

supplemented in a future work.

In summary, more accurate information can be

recognized, and the information in the corpus data can

be understood quickly through the use of PreechVis.

7 DISCUSSION

Several interviews were conducted with domain

PreechVis: Visual Proﬁling using Multiple-word Combinations

105

experts who study computational-linguistic and

natural-language processing, and these researchers

agreed that the exploration of the multiword for each

gram is a major strength of PreechVis. Further, the

proposed visual tool can facilitate a quick exploration

of the corpus information and more accurate results

can be obtained, as shown in the presented case

studies. The case studies three important implications

and usage scenarios confirm the usability and

effectiveness of the system. This study presents two

important implications.

First, the study reveals a data-processing method

that can obtain more accurate word results than when

analyzing without multiword from the corpus data.

This study uses a linguistic approach to obtain more

accurate word-combination words, and this was

explained in the data-processing without manual

work. As a result of this study, the proposed visual

system shows more accurate word-combination word

results for each gram.

Second, the visual results of the POS-combination

pattern make it possible to generalize the usage

pattern of the POS and the POS combination. The

computational-linguistic and natural-language-

processing researchers plan to develop a system that

automatically recognizes the multiword without

manual works; this is the beginning of the study of the

generalization of usage patterns according to the POS

and the POS combination of each gram.

And the proposed system can be made more

generic by finding other usage cases of this tool; for

example, authors instead of U.S. Presidents, and

writings instead of speeches.

The present work, however, is hampered by

limitation that this study covers the expressions of

words consisting of three-word combinations.

However, the absence of words consisting of four-

word combinations is problematic; for example, “in

the Middle East,” etc. Therefore, this matter needs to

be supplemented in a future work.

Overall, though, the feedback is positive, and the

experts extracted new findings and gained more

information from the corpus.

8 CONCLUSIONS

For this work, a new interactive-visualization

approach, PreechVis, was designed and demonstrated

regarding the analysis of corpus data. This visual tool

can help users to understand the corpus. The proposed

system was developed using the multi-view and a

novel technique to show the analysis result of the

corpus with the multiword. PreechVis supports a

flexible exploration of the multiword in the corpus,

and the combination patterns of the POS for each

gram can be identified. The case studies and the usage

scenario demonstrate how this tool can be used.

In a future work, the methods for the display of

more visual results for which the linguistic approach

is utilized will be improved, and more multi-words

without the limit of the N-gram will be incorporated.

The authors also hope to support the exploration of

tools that use other computational-linguistic

techniques.

ACKNOWLEDGEMENTS

This work was supported by the 2017 BK21 Program,

Ajou University and National Research Foundation

of Korea (NRF2015S1A5B6037107).

REFERENCES

Andrade, L. and Young, G. (1996). Presidential agenda

setting: Influences on the emphasis of foreign policy.

Political Research Quarterly, 49(3):591–605.

Andrews, K., Wohlfahrt, M., et al. (2009). Visual graph

comparison. In 2009 13th International Conference

Information Visualisation.

Collins, C. and Carpendale, S. (2007). Vislink: Revealing

relationships amongst visualizations. IEEE

Transactions on Visualization and Computer Graphics,

13(6):1192–1199.

Cui, W., Liu, S., et al. (2014a). How hierarchical topics

evolve in large text corpora. IEEE Transactions on

Visualization and Computer Graphics, 20(12):2281–

2290.

Cui, W., Wang, X., et al. (2014b). Let it flow: a static

method for exploring dynamic graphs. In 2014 IEEE

Pacific Visualization Symposium. IEEE.

Heimerl, F., Han, Q., et al. (2016). Citerivers: Visual

analytics of citation patterns. IEEE Transactions on

Visualization and Computer Graphics, 22(1):190–199.

Hughps, J. (2009). Analyzing policy issues in presidential

speeches and the media: An agenda-setting study. In

University of Nevada, Las Vegas, USA.

JR, F. (1957). Papers in linguistics 1934-1951. Oxford

University Press.

Koch, S., John, M., et al. (2014). Varifocal reader in depth

visual analysis of large text documents. IEEE

Transactions on Visualization and Computer Graphics,

20(12):1723–1732.

Liu, C. and Wang, P. (2015). A sunburst-based hierarchical

information visualization method and its application in

public opinion analysis. In 2015 8th International

Conference on BioMedical Engineering and Informatics

(BMEI 2015).

IVAPP 2018 - International Conference on Information Visualization Theory and Applications

106

Lu, Y., Steptoe, M., et al. (2016). Exploring evolving media

discourse through event cueing. IEEE Transactions on

Visualization and Computer Graphics, 22(1):220– 229.

Mahyar, N. and Tory, M. (2014). Supporting

communication and coordination in collaborative sense

making. IEEE Transactions on Visualization and

Computer Graphics, 20(12):1633–1642.

Qu, S., Wang, S., et al. (2008). Improvement of text feature

selection method based on tfidf. In 2008 International

Seminar on Future Information Technology and

Management Engineering.

Ramisch, C. (2015). Multiword Expressions Acquisition.

Springer.

Rodden, K. (2014). Applying a sunburst visualization to

summarize user navigation sequences. In IEEE

Computer Graphics and Applications. IEEE.

Rubio-Snchez, M., Raya, L., et al. (2016). A comparative

study between radviz and star coordinates. IEEE

Transactions on Visualization and Computer Graphics,

22(1):619–628.

Sharko, J., Grinstein, G., et al. (2008). Vectorized radviz and

its application to multiple cluster datasets. IEEE

Transactions on Visualization and Computer Graphics,

14(6):1444–1451.

Sidorov, G., Velasquez, F., et al. (2013). Syntactic

dependency based n-grams: More evidence of

usefulness in classification. In Springer Verlag Berlin

Heidelberg 2013. Springer.

Stasko, J., Gorg, C., et al. (2007). Jigsaw: Supporting

investigative analysis through interactive visualization.

In IEEE Symposium on Visual Analytics Science and

Technology 2007. IEEE.

Sun, G., Wu, Y., et al. (2014). Evoriver: Visual analysis of

topic coopetition on social media. IEEE Transactions

on Visualization and Computer Graphics, 20(12):1753–

1762.

Wang, X., Liu, S., et al. (2016). Topic panorama: A full

picture of relevant topics. IEEE Transactions on

Visualization and Computer Graphics, 22(12):2508–

2521.

Wu, Y., Liu, S., et al. (2014). Opinion flow: Visual analysis

of opinion diffusion on social media. IEEE Transactions

on Visualization and Computer Graphics, 20(12):1763–

1772.

Zhang, W., Yoshida, T., et al. (2008). Tfidf, lsi and multi-

word in information retrieval and text categorization.

In 2008 IEEE International Conference on Systems,

Man and Cybernetics. IEEE.

PreechVis: Visual Proﬁling using Multiple-word Combinations

107