Evaluating the Inverted Pyramid Structure through Automatic
5W1H Extraction and Summarization
Brian Keith Norambuena
∗
Virginia Tech
Blacksburg, Virginia, United States
Michael Horning
Virginia Tech
Blacksburg, Virginia, United States
Tanushree Mitra
Virginia Tech
Blacksburg, Virginia, United States
ABSTRACT
The inverted pyramid is a basic structure of news reporting used
by journalists to convey information and it is considered a key
element of objectivity in news reporting. In this article, we propose
the Inverted Pyramid Scoring method to evaluate how well a news
article follows the inverted pyramid structure using main event
descriptors (5W1H) extraction and news summarization. We evalu-
ate our proposed method on a proprietary data set of Associated
Press news articles across breaking and non-breaking news span-
ning two topics—political and business. Our results show that the
method works at distinguishing the structural dierences between
breaking and non-breaking news. In particular, our results conrm
that breaking news articles are more likely to follow the inverted
pyramid structure.
CCS CONCEPTS
• Computing methodologies →
Information extraction;
• Ap-
plied computing → Publishing.
KEYWORDS
natural language processing, computational journalism, inverted
pyramid, 5W1H extraction
ACM Reference Format:
Brian Keith Norambuena, Michael Horning, and Tanushree Mitra. 2020.
Evaluating the Inverted Pyramid Structure through Automatic 5W1H Ex-
traction and Summarization. In Proceedings of Computation + Journalism
Symposium (C+J 2020). ACM, New York, NY, USA, 7 pages.
1 INTRODUCTION
The inverted pyramid structure—a system of news writing that
arranges facts in descending order of importance—has been a cor-
nerstone of journalism since the late 19th century [
8
]. This style
of structuring information emphasizes fact-based reporting and
neutrality—two of the key components asserting objectivity in jour-
nalistic writing [
16
]. These elements are also particularly important
for hard news reports that require timely reporting and are charac-
terized by high news value (e.g., breaking news stories on political
topics). Moreover, scholars have found that an inverted pyramid
information structure is a distinctive feature of real journalistic
reports; whereas fake news stories often rely on opinion-based
reporting and at times are written in structurally ambiguous ways
[
19
]. Thus, determining how well a news story ts the inverted
pyramid arrangement could be useful in determining whether the
report follows journalistic standards. This paper introduces one
∗
Also with Universidad Católica del Norte, Department of Computing & Systems
Engineering.
such measure for structural analysis of news writing—the Inverted
Pyramid Score (IPS).
Prior work on computationally analyzing the inverted pyramid
structure includes Zhang and Liu’s visual and statistical exploration
using rhetorical structure theory [
24
] and Dai et al.’s classier-
based approach to detect structure in news articles using various
lexical, syntactic and semantic features [
4
]. While these approaches
exploit the rhetorical and syntactic structures in news reporting,
they do not leverage the two distinctive elements of the inverted
pyramid reporting style—summarizing and compressing the most
newsworthy aspects of the story at the very start [
8
] Thus, we
design our IPS scoring method by leveraging these two key elements
and comprises of the following two components:
(1) Main event descriptor locations
: The most newsworthy
aspects of a story (i.e., the 5W1H questions: who, what, when,
where, why and how) are compressed in the opening para-
graphs (see Figure 1). The answers to the 5W1H questions
describe the main event of the news article.
(2) Summary similarity
: The opening paragraphs summarize
the story in such a way that it is possible to cut out the
last paragraphs without losing key information [
14
]. Thus,
any summary of an inverted pyramid news article should
be similar to the Opening Paragraphs (OP), composed of
the headline, lead, and second paragraph (see Figure 1). We
consider the 2nd paragraph in addition to the headline and
the lead because it may include some key details, such as the
answers to 5W1H questions [8].
We validate and show the eectiveness of our IPS method on a
proprietary data set of Associated Press’ News articles, comprising
news from December 2016 to December 2017. Our experiments
demonstrate that our proposed method is capable of evaluating the
inverted pyramid structure and, on average, can distinguish the
structural dierences between breaking and non-breaking news.
This is the rst step towards evaluating objectivity by the use of
the inverted pyramid structure.
2 BACKGROUND AND RELATED WORK
We divide this section into three parts. First, we delve into the
dierent news structures used in written journalism. Next, we focus
on the two components that form the core of our IPS calculation,
namely 5W1H extraction and text summarization.
2.1 Structure of News Articles
There are many ways to structure a news article. We describe the
four most common structures found in the literature [
4
,
8
,
18
].
In the Inverted Pyramid structure, the article presents content in
descending order of importance with key events placed rst and
C+J 2020, March 20 – 21, 2020, Boston, MA, USA Brian Keith Norambuena, Michael Horning, and Tanushree Mitra
Figure 1: News article [13] showing the Opening Paragraphs
(OP) and the rest of the article. The highlighted phrases
show the answers to the 5W1H questions that dene a news
event: What happened? Who is involved? When and Where
did it happen? Why and How did it happen?
additional details discussed later. In the Kabob, the article starts
with an anecdote to capture the reader’s attention, then introduces
the key events and main story, followed by a general discussion
with more details. In the Martini Glass, the article relies on narrative
chronology beginning with a summary of the main event. The arti-
cle starts with the inverted pyramid structure but then transitions
into a narrative story following a chronological order. Lastly, in the
Narrative structure, the article presents a chronological sequence
of events with more details than usual news articles.
Each one of these structures has distinctive features that dis-
tinguish them from one another, in particular, writing style and
presentation order [
4
]. While the last three structures are useful for
some news, none enjoy the popularity of the inverted pyramid, the
most common structure in written journalism. Furthermore, the
inverted pyramid is considered a fundamental pillar of objectivity
[
16
] as well as a key feature of professional journalistic news re-
porting as opposed to fake news [
20
]. Hence, for the purposes of
this study, we focus on analyzing the inverted pyramid structure,
instead of attempting to classify among all possible journalistic
structures.
2.2 Main Event Descriptor Extraction
There are many methods to extract main event descriptors (i.e., the
5W1H answers). We present some approaches in this subsection.
Most works are purely based on rules that leverage lexical, syn-
tactic and semantic information to obtain answer candidates. Verb-
based approaches work by identifying the main action in a sentence
or text, represented by a verb [
17
]. Once the method has identied
the main verb, it extracts the arguments associated with it to nd
the main event descriptors (e.g., the subject of the verb). Semantic
role labeling identies semantic predicates at a sentence level [
3
].
Then, it identies syntactic components through shallow parsing
and assigns them a semantic role in the predicate. This method
leverages syntactic relationships to identify text semantics. Machine
learning classiers can be trained and used to extract main event
descriptors [
23
]. In particular, they can predict the arguments of
the main predicate of a sentence (i.e., the 5W1H answers). However,
due to the scarcity of annotated data sets [
6
] there is less work on
methods that exploit annotations to improve their results.
Finally, we highlight the recent works of Hamborg et al. [
6
,
7
]
which present the development of an open-source system for 5W1H
extraction (
Giveme5W1H
) along with a gold-standard data set. We
have used their extraction and scoring methods as a guide for our
own 5W1H extraction system.
2.3 News Summarization
News summarization is an extensively studied application of nat-
ural language processing. It comprises of two main approaches:
abstractive and extractive. While abstractive methods rephrase and
compress the original text to create the summary, extractive meth-
ods select key sentences from the text to build the summary [10].
For our work, we consider using an extractive summarization al-
gorithm. In particular, we use TextRank, an algorithm that has
been successfully used in multiple applications [
2
], is domain-
independent and does not require deep linguistic knowledge [15].
Finally, news summarization is closely related to 5W1H extrac-
tion, since answers to the main event descriptor questions can
be used to provide an explicit summary of the main event [
7
]. In
essence, both summaries and main event descriptors are performing
the same task: they distill the article into a simpler representation.
In the context of the inverted pyramid structure, the results from
both tasks should always be related to the beginning of the article.
3 DATA COLLECTION AND ANNOTATION
We start by describing the dataset used, present the subset employed
for the IPS evaluation and describe the annotated sample used for
evaluating the main event descriptor extraction (5W1H).
3.1 AP News Data
Our work is based on a proprietary data set from the Associated
Press News (AP News) spanning a full year of news articles from
December 2016 to December 2017, a total of 65,535 articles. The
Associated Press was instrumental in creating the inverted pyramid
structure [
22
] and continues to use it for reporting. Hence, the AP
news data is ideal for testing our inverted pyramid scoring method.
Each news article in our dataset includes information about
the topics of the news, referred to as subject tags or categories,
spanning from serious topics like Business to lightweight topics,
such as Entertainment. Following Bakshy et al’s [
1
] hard-soft news
classication scheme, the following AP News categories are likely
to be hard news: Science, Politics, Business, Health, and Weather.
Hard news is characterized by a high level of newsworthiness or
news value and require timely publication (e.g., politics or business
news). Whereas, soft news has a low level of informational value
(e.g., entertainment news) and does not need immediate publication
[
11
]. The AP data also contains a label indicating whether a news
article is Breaking or Non-Breaking news. Breaking news articles
usually follow the inverted pyramid structure and we expect them
to have a higher IPS compared to non-breaking news articles. Thus,
if our scoring method works, it should be able to score breaking
news with signicantly higher IPS value than non-breaking new
articles.
Evaluating the Inverted Pyramid Structure through Automatic 5W1H Extraction and Summarization C+J 2020, March 20 – 21, 2020, Boston, MA, USA
What genre of news should we evaluate with our scoring method?
We decided to test our scoring method on “hard breaking news”
since the inverted pyramid is seen as the distinctive feature of hard
news reports [
21
]. Furthermore, we are interested in applying IPS
as a means of establishing journalistic standards and as a method
to contrast it with fake news. Hence, focusing on hard news makes
sense as they are often subjected to misinformation.
We rst ltered articles with missing elds (e.g., countries, tags),
non-English articles, and retained those that were tagged as “United
States”. We focused on a single month to reduce variability in the
news articles and to minimize computational costs. In particular,
we chose November 2017 as it is the most recent month with a
high number of articles (5,045 articles, compared to December 2017
which only has 945 articles). For this rst phase of our study, we
focus on the two most frequent hard news categories in our data
set: Politics and Business. Our nal sample comprised 1,529 articles.
3.2 Annotations
One of the key steps in building our scoring method is extracting
answers to the 5W1H main event descriptor questions from an
article. How well does our 5W1H extraction work? To answer, we
extracted a random sample of 30 breaking news from our data (15
political and 15 business articles) and obtained annotations for the
5W1H answers from experts trained in writing journalistic articles
following the inverted pyramid. Specically, we asked senior jour-
nalism students to assess our 5W1H answers who are extensively
trained in using the inverted pyramid and writing 5W1H answers.
Journalism students were provided a questionnaire with a 3-point
Likert scale to evaluate the descriptors. Additionally, if a descriptor
was not present in the article, students could mark it as N/A. Figure
5 in the appendix shows an example question. We received six
evaluations for each article, totaling 180 annotations for our sample
of 30 articles. For each descriptor per article, we averaged the expert-
assigned scores. Next, we averaged these results again over all
articles to get the nal evaluation for each question.
4 INVERTED PYRAMID SCORE
Here we present the Inverted Pyramid Score and its two main
components: 1) main descriptor locations and 2) summary similarity.
The IPS answers the following question: how well does a news
article follow journalistic standards? And, in particular, how well
does an article t the inverted pyramid structure?
Figure 4 in the appendix presents a general overview of our
method. To nd the IPS, we rst apply standard preprocessing
steps on our dataset. Next, we compute a score for the
main de-
scriptor locations
and the
summary similarity
. Both scores are
computed with respect to the opening paragraphs comprising the
headline, lead, and 2nd paragraph. Finally, we compute the nal
IPS as the weighted average of the two component scores. We as-
signed a higher weight to main descriptor locations since the 5W1H
questions guide the writing of the inverted pyramid [5].
Data Preprocessing. We start by applying standard preprocessing
techniques, such as tokenization, sentence splitting, part-of-speech
tagging, dependency parsing and named entity recognition. Addi-
tionally, we use
neuralcoref
to handle coreference resolution. As
an example of coreference resolution, consider Figure 1. “Hawaii
group” is the answer to WHO, but this entity is also mentioned in
the text as a “group representing Hawaii commercial shermen.”
This is a coreference because they refer to the same entity.
4.1 Main Descriptor Locations
The rst component of the IPS models how well the article captures
the main event descriptors. In particular, we rst check whether
all the main event descriptors (i.e., 5W1H answers) are present
in the article. Our 5W1H main event extraction comprises two
steps: extracting all possible candidates for 5W1H and scoring
candidates to nd the best match. Next, we ensure that the main
event descriptors appear early in the text. We assign a score based on
the descriptor’s location in the article, penalizing those appearing
below the headline sentence.
4.1.1 Extracting main event descriptors. We build our main event
descriptor extraction module by following the generic 5W1H archi-
tecture described by Hamborg et al., specically their implementa-
tion of
Giveme5W1H
. We extend their implementation by including
additional rules and rening the candidate scoring mechanism.
Extracting possible candidates for 5W1H
What & Who. We nd all sentences of the form NP-VP-NP (e.g.,
[The cat]
NP
[quickly climbed]
VP
[the apple tree]
NP
). Usually, the
answer to WHO is contained in the rst NP of these structures
and the answer to WHAT is contained in the VP-NP part. We also
exclude candidates that contain attribution verbs in the VP (e.g., said,
told). Since attributions usually only oer supporting information,
they are unlikely to contain the answer to WHAT and WHO.
When. To extract WHEN candidates, we parse regular dates
and relative dates. We also check for dates that escaped auto-
matic date parsing (e.g., “Christmas weekend”). We handle them
by adding manual rules. Furthermore, we identify additional time
noun phrases using a dictionary of time nouns [12].
Where. For WHERE candidates, we nd all named entities that
are tagged as location. We geolocate them using the OpenCage API.
Why. We search for three elements to extract WHY candidates:
adverbs that express causal relationships, causal conjunctions, NP-
VP-NP structures with causal verbs and auxiliary verbs that can be
used for showing causes but aren’t specic enough (e.g., “to be,” as
in “the airplane failure was a mechanical issue.”).
How. We extract sentences that use one of the copulative con-
junctions (usually the phrase after the conjunction is the HOW).
We also nd NP-VP-NP phrases that have adverbs and adjectives
(since these modiers can reect the answer to HOW).
Candidate scoring to nd the best match
After extracting the potential candidates, we score and rank them
to get the nal answer for each 5W1H. We designed all scores to be
between 0 and 1. We score candidates based on a combination of
the following criteria: position, type, frequency, precision, length,
and other candidate-specic scoring criteria.
Position score. For all 5W1H questions, we assign a high score
when candidates are found early in the text. For occurrences in the
rst sentence of the document (or headline), we assign a score of
1. For occurrences in subsequent sentences, the score follows an
exponential decay, decreasing with an increase in position,
p
. Specif-
ically,
S
p
(C) = e
(−dp)
with decay coecient,
d >
0. To illustrate,
C+J 2020, March 20 – 21, 2020, Boston, MA, USA Brian Keith Norambuena, Michael Horning, and Tanushree Mitra
let’s refer to Figure 1. Considering logarithmic decay,
d = log(
2
)
,
we divide the score by half whenever we move farther away from
the headline. Thus, the WHO candidate (
p =
0) will be scored with
1 and the WHEN candidate with 0.25 (p = 2).
Type score. Scoring based on candidate type, such as proper or
common noun, date or time, etc., depends on the 5W1H question
being answered. For WHO, it refers to whether the candidate is
a named entity (i.e., a proper noun). For example, if the extracted
candidate for WHO is a named entity, we score it as 1, otherwise it
is scored as 0. For WHEN, type refers to whether the candidate is
a proper date or a vague expression. For WHERE, it refers to the
type of location (e.g., geopolitical entities, geographical locations,
man-made structures, or organizations which can be used to refer
to places in some cases). For WHY and HOW, we score candidates
based on whether it is expressed through an NP-VP-NP pattern or
conjunction or a combination of both.
Frequency Score. For all questions, except WHY and HOW, we
rank-score candidates by their frequency of occurrence in the article.
The highest frequency candidate is scored as 1. If the candidate is a
named entity, we count all its coreferences, otherwise, we simply
count the raw occurrences. For example, consider “Hawaii” and
“United States” as WHERE candidates for the article in Figure 1.
If we only consider the parts shown in Figure 1, then the article
mentions the rst candidate four times and the second candidate
only once. We normalize the counts by the highest frequency and
assign a score of 1 to “Hawaii” and 1/4 to “United States.”
Precision and Length Score. For WHERE and WHEN, we consider
the
Precision
of the candidate. For example, a date with an exact
time is ranked higher than a vague phrase like “election time” and
“London” is ranked higher than “UK” because it is a more precise
location. For WHY and HOW, we consider the Length of the can-
didate. We prefer longer explanations for the cause and method. To
implement this, we count the number of words in the candidate
and divide by the maximum count in all candidates. Moreover, we
add a redundancy penalty if the candidate repeats the answer to
WHAT or if we get the same answer for WHY and HOW.
Other Scoring Criteria. For WHEN, we also score candidates by
distance to publication date, preferring dates closer to the publica-
tion date. For WHERE, we score candidates by clustering. We assign
a higher score if a candidate is close to the other candidates. For
example, if most locations are in Germany, then we would assign
less score to a random location in Japan. For HOW, we score candi-
dates by modier frequency, which counts the number of adverbs
and adjectives used by the candidate.
4.1.2 Location Scoring of Main Event Descriptors. We assign the
location scores for each main event descriptor using the following
criteria: if an article follows an inverted pyramid structure, it should
provide answers to the 5W1H questions in the OP (see Figure 1).
Thus, if we nd the answers there, we assign a high IPS. While
the headline and lead are usually one sentence long each [
8
], the
2nd paragraph can have at most three sentences. We found this
maximum length by analyzing breaking news articles in the data
set. Hence, for the purposes of our estimation, we consider the OP
to be the rst 5 sentences of an article. We give a full score if all
5W1H descriptors are contained in the OP. Otherwise, we apply an
exponential penalty by location of each descriptor. More formally,
considering the headline index to be 0, for each descriptor D,
LocScore(D) =
(
2
4−max
(
4, Location(D)
)
if answer found
0 if answer not found.
Finally, we obtain a weighted average of all the location scores.
Since HOW and WHY are not necessarily present, and even humans
may have problems extracting them, we assign them a lower weight
than the other descriptors.
4.2 Summarization
The second component of the IPS models how well an article is sum-
marized by the OP. By denition, an article following an inverted
pyramid structure must be summarizable by removing everything
except the OP—the headline, lead, and 2nd paragraph. Note how in
Figure 1 the OP contains all relevant information about the news
story. Hence, our generated summary should be similar to the OP.
Thus, we implement our summary similarity module by comparing
the summary of the full article with the OP. First, we summarize the
full article using an extractive summarization algorithm—TextRank.
TextRank ranks an article by the most important sentences and then
uses those to build the summary. Next, we compare the full article
summary and the OP by comparing the language representations of
the two. In particular, we do this using
Spacy
and their pre-trained
en_core_web_lg
model. This model uses GloVe vectors and it was
trained with a multi-task CNN on blogs, news, and comments [
9
].
We average all the word vectors contained in a text to get its nal
representation. Finally, we compute the summarization score using
the Cosine similarity distance between the vector representations
of the OP and the summary.
5 RESULTS AND DISCUSSION
Here we show our main ndings and discussions. We begin by
presenting the evaluation of our main event descriptors extractor.
Next, we report the results on the November 2017 AP News articles,
showing the IPS distributions for breaking and non-breaking news.
5.1 5W1H Extraction
Table 1 shows the evaluation results of our 5W1H method. We
nd that our extractor is capable of obtaining the right answers for
the basic 4W with 78% accuracy on average. Out of the four basic
descriptors, our method systematically extracted better results for
WHERE in this data set. This could be attributed in turn to the
date-line being explicitly included in AP News articles.
However, for the full main event descriptors we only achieve
67% average accuracy. This reduction in accuracy makes sense
considering the inherent diculty of extracting the causes and
methods from news articles. Even though the accuracy for WHY
and HOW is still low compared to the other questions, our method
is on par with the state-of-the-art.
As a baseline for comparison,
Giveme5W1H
gets 0.73 accuracy
for all descriptors and 0.82 for the basic 4W on a BBC news data
set [
6
]. However, it is hard to draw a direct comparison because
of dierences in the background of the annotators (journalism
students vs IT students) and of data sets (AP News vs BBC).
Evaluating the Inverted Pyramid Structure through Automatic 5W1H Extraction and Summarization C+J 2020, March 20 – 21, 2020, Boston, MA, USA
Question Business Politics Total
Who 0.74 ± 0.06 0.77 ± 0.07 0.76 ± 0.04
What 0.79 ± 0.05 0.73 ± 0.06 0.76 ± 0.04
When 0.71 ± 0.05 0.83 ± 0.05 0.77 ± 0.04
Where 0.87 ± 0.04 0.81 ± 0.04 0.84 ± 0.03
Why 0.42 ± 0.08 0.51 ± 0.06 0.46 ± 0.05
How 0.46 ± 0.07 0.42 ± 0.07 0.44 ± 0.05
Avg (Total) 0.66 ± 0.08 0.68 ± 0.07 0.67 ± 0.07
Avg (4W) 0.78 ± 0.04 0.78 ± 0.02 0.78 ± 0.02
Table 1: 5W1H evaluation results for the breaking news in
the AP data set by subject categor y (± standard errors).
5.2 Inverted Pyramid Score
After testing the main event descriptor extractor on the previous
sample we turn to the main task. Using our full data set, we compute
the IPS of each article and show its distribution and basic statistics in
Figure 2. In general, a higher IPS means that the articles adhere more
to the inverted pyramid structure. Thus, these results match our
intuition that breaking news usually follows the inverted pyramid
structure. Non-breaking news shows more structural variety, as
evidenced by their higher standard deviation and lower IPS.
While our IPS method gets intuitively correct results on breaking
and non-breaking news, there might be other factors that aect
whether a news article is written using an inverted pyramid struc-
ture or something else. In particular, an important element is the
writing style used in the article. Inverted pyramid news will likely
follow an expository writing style rather than a narrative writing
style [
4
]. Consequently, we could add a new component to our
scoring method that accounted for writing style dierences.
Min: 0.18
Max: 0.99
Avg: 0.71
Std: 0.14
Min: 0.25
Max: 0.99
Avg: 0.87
Std: 0.12
Statistics
Statistics
Figure 2: Plot of the IPS distribution and basic statistics
for breaking and non-breaking news. On average, breaking
news has a higher IPS than non-breaking news.
6 CONCLUSION
We have presented our work on evaluating the inverted pyramid
structure using 5W1H extraction and summarization. Our anal-
yses of results show that the method works well, allowing us to
distinguish between breaking and non-breaking news articles.
In terms of improving our method, future work includes making
improvements to the 5W1H extractor and using state-of-the-art
summarization schemes tailored for news articles. In terms of po-
tential applications, we plan on using this work to evaluate the
dierent structures of news articles, not only restricting ourselves
to the inverted pyramid. The current implementation could also be
used, with additional features and descriptors, to provide a classi-
cation tool for breaking and non-breaking news.
Finally, our long-term goal is to evaluate how fake news sources
structure their articles, as well as comparing them to mainstream
outlets. By nding these structural dierences we hope to elucidate
how fake news articles dier from regular news.
ACKNOWLEDGMENTS
This work was partially funded by CONICYT PCFHA / DOCTOR-
ADO EXTRANJERO BECAS CHILE/2019 - 72200105.
REFERENCES
[1]
Eytan Bakshy, Solomon Messing, and Lada A. Adamic. 2015. Exposure to ide-
ologically diverse news and opinion on Facebook. Science 348, 6239 (2015),
1130–1132.
[2]
Federico Barrios, Federico López, Luis Argerich, and Rosita Wachenchauzer. 2015.
Variations of the Similarity Function of TextRank for Automated Summarization.
In Argentine Symposium on Articial Intelligence (ASAI 2015) (Rosario, 2015).
[3]
Kunal Chakma and Amitava Das. 2018. A 5w1h based annotation scheme for
semantic role labeling of English tweets. Computación y Sistemas 22, 3 (2018),
747—-755.
[4]
Zeyu Dai, Himanshu Taneja, and Ruihong Huang. 2018. Fine-grained Structure-
based News Genre Categorization. In Proc. of the Workshop Events and Stories in
the News 2018. Association for Computational Linguistics, Santa Fe, New Mexico,
U.S.A, 61–67.
[5]
Delia Gavriliu. 2012. From the Print Press to Online Press: Constraints and
Liberties of the Journalistic Discourse. Procedia - Social and Behavioral Sciences
63 (2012), 263 – 270. The 4th Edition of the International Conf.: Paradigms of
the Ideological Discourse 2012.
[6]
Felix Hamborg, Corinna Breitinger, and Bela Gipp. 2019. Giveme5W1H:
A Universal System for Extracting Main Events from News Articles.
arXiv:cs.CL/1909.02766
[7]
Felix Hamborg, Soeren Lachnit, Moritz Schubotz, Thomas Hepp, and Bela Gipp.
2018. Giveme5W: Main Event Retrieval from News Articles by Extraction of
the Five Journalistic W Questions. In Transforming Digital Worlds. Springer
International Publishing, Cham, 356–366.
[8]
Tim Harrower. 2010. Inside reporting. Vol. 310. McGraw-Hill Education, 1221
Avenue of the Americas, New York, NY 10020.
[9]
Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language under-
standing with Bloom embeddings, convolutional neural networks and incremen-
tal parsing. (2017).
[10]
Jagadish S Kallimani, KG Srinivasa, and B Eswara Reddy. 2012. Summarizing
news paper articles: experiments with ontology-based, customized, extractive
text summary and word scoring. Cybernetics and Information Technologies 12, 2
(2012), 34–50.
[11]
Sam N. Lehman-Wilzig and Michal Seletzky. 2010. Hard news, soft news, ‘general’
news: The necessity and utility of an intermediate classication. Journalism 11,
1 (2010), 37–56.
[12]
Michaela Mahlberg. 2005. English general nouns: A corpus theoretical approach.
Vol. 20. John Benjamins Publishing.
[13]
Audrey McAvoy. [n.d.]. Hawaii group wants to defend licenses for for-
eign shermen. Associated Press News ([n. d.]). https://apnews.com/
00962efef9e941ee8563cd24855c1fb8
[14]
Melvin Mencher and Wendy P Shilton. 2011. News reporting and writing. Brown
& Benchmark Publishers.
[15]
Rada Mihalcea. 2004. Graph-based ranking algorithms for sentence extraction,
applied to text summarization. In Proc. of the ACL Interactive Poster and Demon-
stration Sessions. 170–173.
[16]
David TZ Mindich. 2000. Just the facts: How "objectivity" came to dene American
journalism. NYU Press.
[17]
Kristen Parton, Kathleen R McKeown, Bob Coyne, Mona T Diab, Ralph Grishman,
Dilek Hakkani-Tür, Mary Harper, Heng Ji, Wei Yun Ma, Adam Meyers, et al
.
2009. Who, What, When, Where, Why? Comparing Multiple Approaches to the
Cross-Lingual 5W Task. In Proc. of the Joint Conf. of the 47th Annual Meeting of
the ACL and the 4th International Joint Conf. on Natural Language Processing of the
AFNLP. Association for Computational Linguistics, Suntec, Singapore, 423–431.
[18]
Horst Pöttker. 2003. News and its communicative quality: the inverted pyra-
mid—when and why did it appear? Journalism Studies 4, 4 (2003), 501–511.
[19]
Jaakko Salo. 2019. A Genre Analytical Comparison of Real and Fake News. (2019).
C+J 2020, March 20 – 21, 2020, Boston, MA, USA Brian Keith Norambuena, Michael Horning, and Tanushree Mitra
[20]
Pamela J. Shoemaker. 2017. News Values: Reciprocal Eects on Journalists and
Journalism. American Cancer Society, 1–9.
[21]
Elizabeth A. Thomson, Peter R. R. White, and Philip Kitley. 2008. “Objectivity” and
“Hard News” reporting across cultures. Journalism Studies 9, 2 (2008), 212–228.
[22]
Patrick Walters. 2017. Beyond the inverted pyramid: Teaching the writing and all-
formats coverage of planned and unplanned breaking news. Teaching Journalism
& Mass Communication 7, 2 (2017), 9–22.
[23]
Sibel Yaman, Dilek Hakkani-Tür, Gokhan Tur, Ralph Grishman, Mary Harper,
Kathleen R McKeown, Adam Meyers, and Kartavya Sharma. 2009. Classication-
based strategies for combining multiple 5-w question answering systems. In
Tenth Annual Conf. of the International Speech Communication Association.
INTERSPEECH-2009, 2703–2706.
[24]
Hongxin Zhang and Haitao Liu. 2016. Visualizing structural “inverted pyramids”
in English news discourse across levels. Text & Talk 36, 1 (2016), 89–110.
APPENDIX: FIGURES
Figure 3: Diagram describing the inverted pyramid structure. The information is shown in descending order of importance,
with the key details at the top.
Figure 4: Diagram showing the computation of the Inverted Pyramid Score (IPS) of a news article. The rst step is preprocessing,
then we compute the main event descriptor locations score and the summary similarity score. For the main descriptor locations
score we extract the position of the main event descriptors (5W1H answers), then we get the weighted average of these scores.
For the summar y similarity score, we get the similarity between an extractive summary of the article and the key details. We
get the nal IPS using the weighted average of the previous scores.
Evaluating the Inverted Pyramid Structure through Automatic 5W1H Extraction and Summarization C+J 2020, March 20 – 21, 2020, Boston, MA, USA
Figure 5: Example of a question from the evaluation questionnaire. The questionnaire showed the full article with highlighte d
answers, some additional information about the article on the top left, and the answers b elow that. The students had to select
one answer for each 5W1H question after reading the article and the proposed answers.
