Automated football match reports as models of textuality

2.1 What is automated journalism?

By definition, automated journalism is a form of automated text generation (also called natural language generation) aimed at the production and publication of journalistic content (Haim and Graefe 2018). The human part in this process is limited to the development of the algorithms (Carlson 2015: 416), after which the process of text production and sometimes even publication is fully automatic without human intervention.

In journalism studies, automated journalism is usually seen as part of a broader trend which is termed computational or algorithmic journalism (Thurman 2019). These terms cover the broad range of the use of algorithms for “the gathering, evaluation, composition, presentation, and distribution of news” (Thurman 2019: 180) and include things as diverse as data-driven graphics (Meier-Vieracker 2020) or personalization of news services based on usage metrics (Tandoc 2014). Although automation plays an important role in all these forms of computational journalism (Diakopoulos 2019), I will focus here on automated journalism in the narrow sense, i.e., the automated generation of journalistic texts.

The standard procedure of automated text generation in journalism is the so called template-based approach building on rule-based algorithms, which runs as follows (Haim and Graefe 2017: 2f., 2018: 150–152). In a first step, the software selects and evaluates structured data against the background of statistical trends based on a set of pre-defined rules. In the second step, texts are generated on the basis of pre-written templates. These templates range from phrasal patterns below the sentence level, which are selected, filled and syntactically adjusted up to textual macro structures. As a result, complete texts are generated, which include pre-fabricated parts, some randomly selected variants, and varying information taken or derived from the raw data (de Cesare 2021: 88; Diakopoulos 2019: 99).

Applications of such procedures, which in principle have been the same for about 30 years now (Schmitz 1994), are bound to areas where, firstly, enough structured data are available and, secondly, the composition of texts is so repetitive, formulaic and expectable that even rather strict rule-based approaches will lead to acceptable texts (Haim and Graefe 2017: 3). These include weather reports, (stock) market news and sport reports (see the overview in Dörr 2016). Especially in football, extensive databases are available that make it relatively easy to create match reports that not only reflect the pure facts of the reported match, but also contain statistically derived interpretations and evaluations, such as noting a team’s “strong initial phase”. These evaluations might be the reason why in journalism studies automated journalism is described as the generation of “narrative news texts” (Carlson 2015: 416; see also Diakopoulos 2019: 137), even though, from a linguistic point of view and as will be shown below, the narratives they present are limited to sequencing reported events rather than actually building a plot.

Numerous experimental reception studies in the field of communication and journalism studies have investigated if and to what extent automated texts are recognized as such by human readers (Graefe et al. 2018; also cf. the works cited in Thurman 2019: 185). Also, readers’ judgements of these texts according to approved categories of journalistic text quality like credibility have been investigated. Most studies show that in experimental settings automated texts are not or hardly distinguishable from human-written texts. The evaluations of readability, credibility and journalistic expertise do not differ significantly, either (Graefe et al. 2018; Lermann Henestrosa et al. 2023).

More fine-grained studies have suggested that automated texts are perceived as more objective and informative, but also as more boring, while human-written texts were judged as more pleasant to read (Clerwall 2014). However, the observed differences are hardly ever related to the concrete texts, and which textual elements might have led to the respective judgements is not further investigated. By the way, and this is rather irritating from a linguistic point of view, the texts used as stimuli in the experiments are usually not reported in the studies.

Moreover, there are studies on the reception of and opinions about automated journalism as a whole by professional journalists (Kunert 2020; Thurman et al. 2017). Many journalists emphasize the benefits of opinion-based and stylistically more individual texts by human writers compared to the most objective descriptions in automated texts (van Dalen 2012: 652). In these studies, however, the reference to texts (both automated and human-written) is all the more lacking, which could be used to show what might be a linguistic manifestation of, for example, individual style. It therefore seems instructive to subject the texts themselves to a detailed linguistic analysis. But, before doing so, we need a better theoretical understanding of text, texture and textuality.

2.2 Texts, texture and (cues of) textuality

Any task of developing an algorithm for automated text generation draws on a theoretical concept of what constitutes a text, although this may not be explicitly reflected by the developers. An everyday notion of text that could serve as a reference point here aims at the production of grammatically well-formed sentences that are then combined into a larger but clearly delineated unit. Beyond mere concatenation, however, a text represents a coherent set of sentences that as a whole fulfills a recognizable communicative function (Brinker et al. 2018) – in our case the report of a football match.

In linguistics, this notion of text has been elaborated into a detailed text theory. In their groundbreaking work “Cohesion in English”, Halliday and Hasan (1976: 1) defined the term text as referring to “any passage, spoken or written, of whatever length, that does form a unified whole”. The basic unifying relation is referred to as cohesion, which can be further divided into grammatical cohesion and lexical cohesion (Halliday and Hasan 1976: 6). Linguistic means used as cohesive ties include pro-forms, syntactic constructions such as parallelisms, and connectives on the side of grammatical cohesion. Lexical cohesion can be established through semantic relations such as hyponymy as well as through expanding or condensing paraphrases (Schubert 2017: 319). While sentences are internally marked by structure, the patterns of cohesive ties, both grammatical and lexical ones, constitute the “texture” (Halliday and Hasan 1976: 2), which is defined as “the property of ‘being a text’” and of being distinguished from non-texts.

In a similar approach, De Beaugrande and Dressler (1981) in their foundation of text linguistics introduced the term textuality in order to capture constitutive properties of texts. Expanding the lexico-grammatical approach of Halliday and Hasan (1976), textuality is further differentiated into seven “standards of textuality” (De Beaugrande and Dressler 1981: 3). Cohesion as the set of mostly grammatical relations at the text surface is complemented by coherence as a continuity of sense building on (possibly implicit) conceptual relations like causality or temporal sequence used for explanations, narrations, etc. Moreover, De Beaugrande and Dressler (1981) determine intentionality, acceptability, situationality, informativity and intertextuality as standards of textuality. While intertextuality as the connectedness with other texts is rather intuitive, the other standards are more difficult to comprehend, as they seem to mix text structural properties on the one hand and producer- and recipient-sided properties as well as situational-contextual aspects on the other (Hausendorf et al. 2017: 2).

For this reason, Hausendorf et al. (2017: 8f.) have proposed to reframe the concept of textuality as a heuristic concept aimed at investigating linguistic means in texts which signal or indicate their textuality to their recipients. These means can be regarded as sources and cues of textuality, i.e., as features of texts which will lead the recipients to perceive them as texts. Specifically, Hausendorf et al. (2017) identify the following classes of sources of textuality associated with textuality cues in the texts: Delimitability concerns the various means for drawing the boundaries of a text, e.g., by headlines. Connectivity compromises the whole range of cohesive ties in the sense of Halliday and Hasan (1976), while thematic relatedness covers thematic coherence as well as patterns of thematic unfolding like narrative, explanatory or argumentative sequences. Pragmatic utility covers various features that signal the possible functions and uses of the text. Finally, intertextuality concerns the links to other texts through references, quotes and so on, while patternedness concerns patterns on different levels that a text shares with other instances of the same genre.

This approach provides a suitable theoretical basis for the research question on the textual properties of automated texts in comparison to human-written ones for three reasons. First, it allows to capture the textual features of (automated as well as human-written) football match reports in a consistent theoretical framework: match reports are delimited, grammatically cohesive and thematically coherent sequences of sentences, which are used to inform about the course and outcome of a football match. They may contain intertextual references to other texts and they are highly patterned texts both through a highly standardized text structure and through genre-specific formulaic sequences (Meier 2019). Second, the approach is particularly well applicable to the task of automated text generation. In the terms of this approach, this task can now be described as the task of automatic selection and use of suitable cues of textuality. This process works without intentions and understanding on the part of the text generating machine and therefore has to be boiled down to the processing of a sophisticated, but static set of rules which nevertheless succeeds in emulating textuality. Third, the focus on cues of textuality fits with the analytic interest identified above as a research gap, i.e., the investigation of textual elements which guide recipients’ interpretations and judgements on the texts.

4.1 Quantitative results

Before going into detail, some global observations on the texts will be helpful. Figure 1 shows the average text lengths and sentence lengths grouped by the three sources.

Figure 1:

Text lengths and sentence lengths.

Figure 1 indicates that automated texts are shorter and consist of shorter sentences. This may be a first hint to a greater complexity of the human-written texts. In addition, the size of the boxes shows that there is less scatter in the dataset of automated texts, which points to the rule-based generation of the texts. The same holds true for the average number of paragraphs as shown in Figure 2.

Figure 2:

No. of paragraphs.

The fact that the human texts are subdivided more finely may seem like a purely layout-related detail. But assuming that paragraphs are delimitable units of meaning, often with typographically marked subheadings, this finding suggests that human writers are able to subdivide the thematic macrostructure of the text into smaller parts. In contrast, the algorithm only distinguishes between the main parts of the report like initial situation, course of the game and outlook for the coming matches, and it does not set any subheadings. Thus, the textuality of delimitability is better exploited in human-written texts through “structuring signals” (Hausendorf et al. 2017: 150) such as subheadings. Moreover, the automated texts show a very static kind of patternedness through a most formulaic and recurrent structure of the texts.

4.2 Corpus linguistic results: Lexico-grammatical cues of connectivity and thematic relatedness

The results reported so far remain rather abstract and do not take into account the lexico-grammatical features of the texts. Therefore, I will now focus on more linguistically substantial categories, i.e., parts of speech and lemmas. I will continue to adopt a quantitative perspective and conduct a contrastive frequency analysis called keyword analysis. Keywords (or key items) “are those whose frequency (or infrequency) in a text or corpus is statistically significant, when compared to the standards set by a reference corpus” (Bondi 2010: 3). I will use the automated text corpus as the target corpus, which will be compared to the human-written texts using the Wilcoxon rank-sum test, which has been shown to be more adequate than the widely used keyword metric Log Likelihood Ratio as it also takes the dispersion of keywords into account (Sönning 2023). As I will show, the automated and human-written texts differ significantly in how cues of connectivity (Hausendorf et al. 2017: 161) and cues of thematic relatedness are being used.

4.2.1 Key parts of speech

The results of a key part of speech^[4] analysis is shown in Figure 3. All items displayed were found to be statistically significant with a significance threshold of p < 0.05. However, the bars show the difference in relative frequencies to make the actual difference easier to estimate. The left side shows the parts of speech used more frequently in the automated texts and the right side the parts of speech used less frequently.

Figure 3:

Significant frequency differences of POS tags.

Some results are expectable given the operating principle of the text generation algorithm. On the one hand, NE (proper names) and CARD (cardinal numbers) are massively overrepresented in the automated texts, since these items represent the information gathered in the databases by the algorithm. Even if there are little or no embellishing passages, the hard facts of the game still remain to be reported. On the other hand, ADJA (attributive adjectives as in schöne Flanke ‘beautiful cross’) and ADJD (adverbial or predicative adjectives as in perfekt zurückgelegt ‘perfectly laid off’) are strongly underrepresented in the automated texts. Such evaluative embellishments, which cannot be derived from the raw data alone, remain the preserve of human writers.

PDAT (attributive demonstrative pronouns) are overrepresented in the automated texts. From a text-analytical perspective, this is an interesting finding. Demonstrative pronouns are often used as coreferential, anaphoric means which refer back to a referent already introduced in or derivable from prior discourse (Diessel 1999: 6; Schwarz-Friesel and Consten 2011: 357), even across sentence boundaries. Thus, they can be regarded as a cohesive tie par excellence. Example (1) contains a complex anaphora diese Niederlage (‘this defeat’), the referent of which is the whole event expressed in the preceding sentence musste … Punkte abgeben (‘lost the points’):

(1)

Gegen Bayer 04 Leverkusen musste man zum zweiten Mal in Folge die Punkte abgeben. Durch diese Niederlage fiel Wolfsburg in der Tabelle auf Platz acht zurück. (retresco)

‘Against Bayer 04 Leverkusen, Wolfsburg lost the points for the second time in a row. This defeat dropped Wolfsburg down to eighth place in the table.’

A look at the corpus data, however, shows that the actual complex-anaphoric use of demonstratives is restricted to a small range of expressions like diese Niederlage (‘this defeat’), dieser Sieg (‘this win’) and dieses Remis (‘this draw’). Different from prototypical complex anaphora, they reformulate and condense the preceding proposition, but do not add much information or evaluation (Schwarz-Friesel and Consten 2011: 358). Also, the entire cross-sentence construction seems to be part of a pre-established pattern which is selected as a whole.

While attributive demonstrative pronouns are overrepresented in the automated texts, substituting demonstratives pronouns (PDS) are underrepresented. In particular, the sentence-initial demonstrative pronoun das (‘that’) distinguishes human-written from automated texts. This pronoun is used as a complex anaphora, too. Its antecedent (Schwarz-Friesel and Consten 2011: 355) is a whole proposition rather than a single expression, and their referents are complex bundles of events, as in (2).

(2)

Danach machte der Club ein wenig auf, ohne zunächst gefährlich zu werden. Das aber änderte sich in der 73. Minute als Behrens nach einem Pereira-Abschluss im Nachschuss aus kurzer Distanz am Tor vorbeizielte. (kicker)

‘After that, the club opened up a bit without becoming dangerous at first. But that changed in the 73rd minute when Behrens missed the goal from close range after a Pereira shot.’

As can be concluded from this example, the complex-anaphoric use of das presupposes a conceptual representation of a complex state of affairs, which then can be referred to as a whole. This seems to go beyond the capabilities of the text generation algorithm. In most cases of substituting demonstrative pronouns in the automated texts, they refer back to a concrete noun phrase as in (3):

(3)

Am Zwischenstand änderte sich weiter nichts, sodass dieser auch gleichzeitig der Pausenstand war. (retresco)

‘The intermediate score remained unchanged, so that this was also the score at the break.’

Again, this construction turns out to be a pre-fixed pattern, which the algorithm seems to select as a whole instead of linking smaller components into a larger unit as human writers would do.

Among the parts of speech overrepresented in the automated texts, we also find subordinating conjunctions (KOUS), which connect two propositions and are thus prototypical cohesive ties, too (Stede 2018: 27). Table 2 shows the five most frequent conjunctions tagged as KOUS in automated and human-written texts.

Table 2:

Subordinating conjunctions.

automated (n = 973; 6,586 pMW)		human-written (n = 2034; 4,541 pMW)
als (‘when’)	36 %	dass (‘that’)	16 %
sodass (‘so that’)	16 %	weil (‘because’)	13 %
während (‘while’)	14 %	ehe (‘before’)	12 %
dass (‘that’)	12 %	als (‘as’ or ‘when’)	11 %
da (‘because’)	5 %	während (‘while’)	8 %

The most frequent conjunction in the automated texts is the conjunction als (‘when’), which usually establishes a temporal relation of coincidence between two events (Breindl et al. 2014: 300). However, the corpus data show that it is used in a quite particular way in the automated texts as in (4):

(4)

Der FC Bayern München verpasste den Ausgleich, als ein Kopfball von James Rodriguez das Tor verfehlte (73.). Eine gute Chance für den FC Bayern vergab Lewandowski, als sein Kopfball das Tor verfehlte (77.). (retresco)

‘FC Bayern München missed the equaliser when a header by James Rodriguez missed the goal (73.). A good chance for FC Bayern was missed by Lewandowski when his header missed the goal (77.).’

In this example, the conjunction als connects two propositions that describe the same event in two different ways, which therefore are necessarily coincident. In a way, such a construction allows to build a syntactically complex sentence, which, however, does not correspond to a semantically complex proposition. In contrast, the same scenes are described in the human-written text as follows (5):

(5)

Weiser blockte einen Kopfball von James aus vier Metern (73.), Lewandowski nickte vorbei (77.). (kicker)

‘Weiser blocked a header by James from four metres (73.), Lewandowski nodded past (77.).’

The event is reported in a much more condensed manner without unfolding it into syntactically complex sentences. In the human-written texts, we find uses of als that actually establish a more substantial temporal relation between two events as in (6).

(6)

Die reguläre Spielzeit war längst abgelaufen, als die Frankfurter noch einen Schreckmoment überstehen mussten. (kicker)

‘Regular playing time had long since expired when the Frankfurt team had to survive a moment of shock.’

Here, the regular time being expired sets the ground on which the actual event, the moment of shock, appears. This pattern, prototypical for narratives (Gülich and Hausendorf 2000: 374), seems to go beyond what the algorithm’s set of rules can achieve.

The second most frequent conjunction in the human written text is weil (‘because’), which points to the fundamental conceptual relation of causality as a basis for text coherence. Match reports must not only report the mere facts, but also provide explanations for what happened on the pitch (Schütte 2006). The construction of cause-effect relations is therefore crucial. Typical examples from a human-written texts are (7) and (8):

(7)

In der Schlussphase wurde es noch einmal turbulent, weil beide Teams den Sieg wollten. (kicker)

‘In the final phase, things got turbulent again because both teams wanted the victory.’

(8)

Augsburg hielt nun besser mit, auch weil Werder zeitweise den Fuß vom Gaspedal nahm. (sky)

‘Augsburg now kept up better, also because Werder slowed down at times.’

In both examples, the facts reported in the subordinate clauses provide causal explanations of what happened on the pitch. In the automated texts, however, weil is used only 11 times and makes up only 1 % of the subordinating conjunctions. Comparable with als discussed above, weil is used in a peculiar way as in (9):

(9)

Die verbleibende Zeit der ersten Halbzeit blieb ohne weitere Treffer, weil die Chancen durch Fink, Gießelmann und Usami ohne Erfolg blieben. (retresco)

‘The remaining time of the first half remained without further goals because chances by Fink, Gießelmann and Usami were unsuccessful.’

In this example, the facts reported in the subordinate clauses do not provide a causal explanation in the proper sense. They rather give a further specification or elaboration of the fact reported in the main clause by increasing its level of detail (Scheffler and Stede 2016). The two propositions connected by the conjunction seem to describe the same fact, just in different granularity. The syntactic complexity and the conceptual relation evoked by weil does not adequately correspond to a causal relation on the propositional level.

Heavily underrepresented in automated texts are adverbs (ADV). Again, not only the frequency of adverbs in general differs between automated and human-written texts, but also which adverbs are used. Table 3 shows the five most frequent adverbs in both sub-corpora.

Table 3:

Adverbs.

automated (n = 5,538; 37,488 pMW)		human-written (n = 24,647; 55,022 pMW)
nur (‘only’)	12 %	auch (‘also’)	8 %
noch (‘still)	6 %	aber (‘but’)	7 %
nun (‘now)	5 %	noch (‘still’)	6 %
bisher (‘so far’)	5 %	nur (‘only’)	5 %
auch (‘also)	5 %	wieder (‘again’)	4 %

A most noticeable difference can be observed with aber (‘but’), which is – in its adverbial sense^[5] – used more than twenty times as often in human-written texts (4056 pMW vs. 176 pMW) (cf. Juknevičienė and Viluckas 2019: 68f. for a similar observation on generated football match reports in a video game). As an adversative connective, aber signals or even builds up a contrast between two propositions (Breindl et al. 2014: 516). As the examples (10) and (11), taken from human-written texts, show, aber can indicate a contrast between certain circumstances on the one hand and an actual event on the other:

(10)

Der FSV dominierte, lief aber in einen Konter. (kicker)

‘The FSV dominated but ran into a counterattack.’

(11)

Die Eintracht schien in dieser Phase deutlich unterlegen, schwamm sich aber in der Schlussphase des ersten Durchgangs noch einmal frei. (kicker)

‘Eintracht seemed to be clearly outclassed at this stage, but broke free once again in the final phase of the first half.’

In both examples, expectations derived from apparent dominance or inferiority are thwarted by the events introduced with aber. This contrast is further emphasized by the connective aber, which thus takes a concessive interpretation (Blakemore 1989; Stede 2004: 276).^[6] It goes with the conditional presupposition that a dominant team will not concede a goal under normal conditions as in (10) (Breindl et al. 2014: 520) so that the counterattack comes as a surprise. In (11), the verb scheinen (‘seem’) indicates the mere appearance of being outclassed, but still aber marks the team’s breaking free as a non-expected event. As both examples show, the connective aber plays an important role in dramatic embellishments, since it allows to highlight or even to construct unexpected turning points as indispensable features of tellable stories (Baroni 2014). All this seems to presuppose abstract and at the same time evaluative representations of game situations which cannot be derived directly from the raw data alone and thus remain a preserve of human writers. Accordingly, the overall few uses of adverbial aber in the automated texts fall into a small set of pre-fixed patterns, which the template-based algorithm selects as a whole according to quantifiable rules as in (12).

(12)

Mit bisher nur vier Treffern zeigte der Sturm des 1. FC Nürnberg erhebliche Defizite. Dafür stand die Defensive aber ziemlich sicher. (retresco)

‘With only four goals so far, the offense of 1. FC Nürnberg showed considerable deficits. But the defense was pretty solid.’

It is revealing that the passages with aber in the automated texts do not describe dynamic game scenes, but relations as they can be derived from the mere numbers of goal ratios.

4.2.2 Key lemmas

The observation that human writers seem to have other options for creating contrasts and staging moments of surprise in their narrative representations of the games can also be further supported by keyword analysis at the lemma level. Results are shown in Figure 4. Again, all lemmas displayed are highly significant with p < 0.001, but the bars show the difference of relative frequencies. Only lemmas overrepresented in human-written texts are shown here.

Figure 4:

Significant frequency differences of lemmas, human versus automated.

It shows again that adversative connectives such as aber and doch (both meaning ‘but’) are typical for the human-written texts. The temporal adverb zunächst (‘initially’) also functions as a connective, since it necessarily requires a continuation and thus indicates a chronological sequence of events. Moreover, the corpus data show that it is typically combined with adversative connectives as in (13):

(13)

Die Borussen hatten zunächst alles im Griff und sahen nach einer 3:0-Führung wie der sichere Sieger aus. Aber die TSG glaubte immer an sich, spielte mutig nach vorne und erzielte ab der 75. Minute durch eine große Moral noch drei Treffer. (kicker)

‘Borussia initially had everything under control and looked like a sure winner after taking a 3-0 lead. But TSG always believed in itself, played courageously forward and scored three more goals from the 75th minute onwards thanks to great morale.’

Similarly, the temporal adverb dann (‘then’) is often combined with aber as in (14):

(14)

Die Münchner gingen früh in Front, wurden dann aber zu passiv und kassierten den Ausgleich. (kicker)

‘Munich took an early lead, but then became too passive and conceded the equalizer.’

Such representations of a chronological sequence of events that highlight a surprising change of state seem to be a fundamental narrative principle in human-written match reports. Linguistic constructions of contrast, e.g., by adversative conjunctions and adverbs are a simple but effectful means for this. However, they presuppose conceptual representations of events that can be meaningfully contrasted in the first place, and this goes beyond the capacity of a text generation algorithm.

The same applies to the adverb auch (‘also’), overrepresented in human-written texts. It requires two comparable events, states or the like, which can then be connected. In (15), auch ties the proposition of Stindl’s miss back to the Swiss player’s miss mentioned before.

(15)

Der 22-jährige Schweizer zögerte allerdings zu lange, Verteidiger Joe Scally (18) sprintete rechtzeitig zur Hilfe. Die Gladbacher dominierten weiter, doch auch ein Kopfball von Lars Stindl (67.) aus kurzer Distanz ging knapp über das Tor. (sky)

‘However, the 22-year-old Swiss hesitated too long, defender Joe Scally (18) sprinted to the rescue in time. Gladbach continued to dominate, but a header by Lars Stindl (67) from close range also went just over the goal.’

In the automated texts, on the contrary, the use of auch is limited to a small set of prefixed patterns like womit es auch weiterhin beim Gleichstand blieb (literally, ‘which kept the score tied also further on’).

To sum up, the key item analyses presented in this section show that both the automated and the human-written texts make use of a wide set of textuality cues. These range from grammatical cohesive ties like conjunctions, adverbial connectives and even (complex) anaphora, which establish a high grade of connectivity in the texts, to thematic relations of contrast, temporal sequence or causality. However, a closer look shows that the actual use of these cues differs greatly between the two sorts of texts. In the automated texts, they are often part of prefixed patterns and seem to work without prior conceptual representations, which then would have to be connected in the text production process. In the human-written texts, on the contrary, these textuality cues are used more flexibly, especially for the purpose of narrative elaboration. If the task is not only to describe the course of the game but to present a suspenseful narrative, human writers are clearly superior.

4.3 Qualitative analysis: chains of thematic development

In order to further enrich the corpus linguistic results, a look at a larger snippet is helpful. The following passage (16) is taken from an automated match report:

Apart from the grammatical cohesive ties discussed above, this snippet shows various means that further establish thematic coherence. For example, the first goal reported in the first sentence is taken up and specified by the phrase to be on target in the second sentence, just as the vague time information only a few moments old is further specified by in the second minute. The phrases extension of lead and the next goal subtly build on aforementioned information, and the fact that Leipzig missed out is specified by the information that the Düsseldorf keeper saved a shot. The antonomasia 24-year-old striker, which reformulates the proper name Yussuf Poulsen, instantiates a figure typical of the genre of sports reporting (Burkhardt 2006: 63) and also establishes connectivity while preventing repetitiveness. The phrase last goal of the turbulent start phase finally bundles all the events reported so far into an evaluative account. The algorithm is obviously capable of weighting individual pieces of information against the background of certain statistical regularities in such a way that a certain exceptionality of the event is made clear. The phrase as early as the 9 ^th minute also points in this direction. All this forms the basis for a coherent thematic development of the report, which goes beyond a mere listing of single events and has narrative qualities.

However, a comparison of this text with its human-written counterpart shows that they differ precisely in their narrative qualities. The report of kicker.de reads as follows (17):

While there are some parallels like the antonomasia Konaté > Frenchman > defender and the emphasis on the extraordinariness of the 0:3 early in the game, the differences are even more noticeable. First, this report is much richer in evaluative descriptions and sayings like absolute dream start or caught completely on the wrong foot. Second, the accounts of single events are much more detailed like stumbled the ball into the goal with the support of the inside post – information which is not delivered through the match data the algorithm relies on. Finally, a variety of adverbs already discussed in Section 4.2.1 creates additional weightings and contrasts. Twice a figure is used which binds two consecutive events together and creates a contrast between them. Poulsen initially failed, then finished again, this time giving an early lead. Similarly, Werner still failed to beat Rensing, but a short time later Poulsen then scored. In both cases, the success of the goal is described in the light of the previous failure, giving them additional meaning and presenting them as surprising events. In this manner, human writers can structure their narratives even more variably and suspensefully.