Effective Headlines of Newspaper Articles in a Digital Environment

Figures & data

Table 1 A sample of four original–rewritten headline pairs (both English translations and the original headlines are given for each pair)

Example	Original headline	Rewritten headline
1	“Historically, more things did work out than did not”	Boyan (21) developed from student with an idea to executive of a big enterprise
	“In de geschiedenis zijn meer dingen wel gelukt dan niet”	Van scholier met een idee groeide Boyan (21) uit tot topman van een miljoenenbedrijf
2	Donald Trump still can learn quite a lot from Frank Underwood	What Donald Trump still can learn from Frank Underwood
	Donald Trump kan nog heel wat leren van Frank Underwood	Wat Donald Trump nog kan leren van Frank Underwood
3	How predictable is the weather?	Why is the weather forecast so often wrong?
	Hoe voorspelbaar is het weer?	Waarom zit het weerbericht er zo vaak naast?
4	A hero does not think, he does	Why two Dutch marines received high awards for their exceptional courage
	Een held denkt niet, die doet	Waarom twee Nederlandse mariniers hoge onderscheidingen kregen voor uitzonderlijke moed

Table 2 The features that were extracted from headlines

Feature	Description
Number of characters	The number of characters in the headline, including spaces and punctuation
Number of words	The number of words in the headline
Average word length	The length of all the words divided by the number of words
Number of sentences	The number of sentences in the headline
Number of sentimental words	The number of sentimental words in the headline, split up by positive and negative word counts. This uses the Dutch language subjectivity lexicon by Jijkoun and Hofmann (2008)
Readability score	The Flesch Reading Ease score of the headline. This is a number between 0.0 and 1.0, where a higher score means easier to understand (Kincaid et al. 1975)
Containing question	Whether or not the headline contains a question
Containing quote	Whether or not the headline contains a quote. A difference is being made between a full quote where the quote spans the entire headline, and a partial quote where the quote only spans part of the headline.
Containing signal words	The number of occurrences for each of the following signal words: daarom, deze, dit, hierdoor, hierom, hoe, waarom, wanneer, welke, wie, and zo [directly translated as respectively hence, this, this, therefore, therefore, how, why, when, which, who, and like that]
Containing pronouns	The number of personal and possessive pronouns that are used. This is split up into first, second, or third person, and into singular and plural
Containing number	Whether or not the headline contains a number. Only numerical representations (like “52”) are counted, written out numbers (like “fifty-two”) were not
First word type	The word type of the first word in the headline. The Pattern Python library (De Smedt and Daelemans 2012) was used for word tagging. Pattern uses the Page Treebank tagset (Marcus, Marcinkiewicz, and Santorini 1993)

Figure 1 Distribution of the performance scores of all headlines in the dataset

Table 3 Mean, median, and standard deviation for each feature, for both the rewritten and original headlines

Feature	Mean		Median		SD
	Original	Rewritten	Original	Rewritten	Original	Rewritten
Number of characters	35.57	74.81	34.00	73.00	15.36	24.96
Number of words	6.11	12.39	6.00	12.00	2.84	4.18
Average word length	5.19	5.07	4.80	5.00	1.73	0.89
Number of sentences	1.05	1.21	1.00	1.00	0.26	0.46
Number of sentimental words	0.57	0.94	0.00	1.00	0.75	0.96
Readability score	0.48	0.40	0.45	0.40	0.32	0.23

Table 4 The number of headlines with a given feature

Feature	Number of headlines
	Original	Rewritten
Containing question	181	248
Containing quote	302 full, 55 partial	205 full, 299 partial
Containing signal words ^a	184	805
Containing pronouns ^a	621	1168
Containing number	12	246

^a The total number of signal words/pronouns is given.

Table 5 The list of hypotheses that were formed as a result of the comparison between the original and rewritten headlines

Hypothesis		p	∆Performance
H1	Longer headlines (>50 characters) are preferred over shorter headlines (≤50 characters)	0.297	–
H2	Headlines with shorter words (≤7 characters per word) are preferred over headlines with longer words (>7 characters per word)	0.024	+33
H3	Headlines that contain a question are preferred over headlines that do not	0.019	–17
H4	Headlines that contain a partial quote are preferred over headlines that do not contain any quote	0.239	–
H5	Headlines that do not contain any quote are preferred over headlines that contain a full quote	0.030	+17
H6	Headlines that contain one or more signal words are preferred over headlines that do not	0.002	+16
H7	Headlines that contain one or more personal or possessive pronouns are preferred over headlines that do not	0.000	+25
H8	Headlines that contain one or more sentimental words are preferred over headlines that do not	0.018	+12
H9	Headlines that contain one or more negative sentimental words are preferred over headlines that do not	0.001	+17
H10	Headlines that contain a number are preferred over headlines that do not	0.202	–
H11	Headlines that start with a personal or possessive pronoun are preferred over headlines that do not	0.002	+20

In the third column the p value is given, showing whether the hypothesis can be accepted at α = 0.05. If so, the hypothesis number is bold, and the change in performance in percentage points is given in the last column.

Jijkoun, Valentin , and Katja Hofmann . 2008. “Task-Based Evaluation Report: Building a Dutch Subjectivity Lexicon.” ILPS-ISLA, University of Amsterdam 2–11.

 Google Scholar

Kincaid, J. Peter , Robert P. Fishburne, Jr , Richard L. Rogers , and Brad S. Chissom . 1975. Derivation of New Readability Formulas (Automated Readability Index, Fog Count and Flesch Reading Ease Formula) for Navy Enlisted Personnel . DTIC Document. Technical report.

 Google Scholar

De Smedt, Tom , and Walter Daelemans . 2012. “Pattern for Python.” The Journal of Machine Learning Research 13 (1): 2063–2067.

 Google Scholar

Marcus, Mitchell P. , Mary Ann Marcinkiewicz , and Beatrice Santorini . 1993. “Building a Large Annotated Corpus of English: The Penn Treebank.” Computational Linguistics 19 (2): 313–330.

 Google Scholar