English. This paper presents a methodology for the annotation of the semantic and functional components of news articles (Content Zones, henceforth CZs). We distinguish between narrative and descriptive zones and, within them, among finergrained units contributing to the overall communicative purpose of the text. Furthermore, we show that the segmentation in CZs could provide valuable cues for the recognition of time relations between events.
Italiano. In questo lavoro viene presentata una metodologia per l’annotazione delle componenti semantiche e funzionali del testo giornalistico (Zone di Contenuto). Distinguiamo tra zone narrative e descrittive e, al loro interno, tra ulteriori unità che contribuiscono al dispiegamento dello scopo comunicativo del testo. Inoltre, mostriamo che la segmentazione in Zone di Contenuto offre preziosi indizi per il riconoscimento delle relazioni temporali tra eventi.
The logical structure of a document, i.e. its hierarchical arrangement in sections, paragraphs, sentences and the like, reflects a functional organization of the information flow and creates expectations on where the desired information may be located. As it is often the case, however, breakups in sections and paragraphs are motivated by style or even arbitrary choices.
The segmentation of the text in Content Zones
(CZs, henceforth), i.e. functional categories
contributing to the overall message or purpose, as
induced by the genre of the text1, provides more
reliable and fine-grained cues to access the
structure of its types of functional content. Previous
attempts to annotate CZs have mainly focused
on highly standardized texts like scientific articles
To the best of our knowledge, no efforts have
been undertaken to devise an annotation scheme
targeting the functional structure of news articles
in terms of their content: the inverted pyramid
structure, i.e. the gathering of key details at the
beginning, followed by supporting information in
order of diminishing importance, is too
coarsegrained to be effectively used for information
extraction purposes. Our hypothesis is that modeling
the document’s content via CZs could yield
benefits for high-level NLP tasks such as Temporal
Processing, Summarization, Question-Answering,
among others. In addition to this, CZs qualify as
a higher-level analysis of a text/discourse which
captures different information with respect to
Discourse Relations. The remainder of the paper is
structured as follows: in Section 2 the motivations
of this work are presented, together with related
studies. Section 3 reports on our inventory of CZs,
used to annotate a corpus of English news
articles. Details on the corpus are provided in
Section 4. In Section 5, we describe a case-study
on the correlation between CZs and temporal
re1We adopt Systemic Functional Linguistics’ view of
genre as “a staged, goal oriented, purposeful activity in
which speakers engage as members of our culture"
The class NARRATION (NARR) includes the following zones: 2
The bulk of the work on discourse structures has
focused on low-level structures corresponding to
Discourse Relations holding between textual
segments pairs. CZs take a different view on texts,
as they perform a function towards the text as
a whole. As an instance of a particular genre,
every text is meant to accomplish a
culturallyestablished communicative purpose, e.g. a news
article reports on events happening in the world.
This goal is not accomplished all at once:
separate functional stages (i.e. CZs) convey fragments
of its overall meaning
In the elaboration of news articles’ CZs, we
were mostly inspired by
The opposition between dynamicity and staticity,
mainly realized by grammatical and lexical aspect,
is adopted as the basic parameter for
differentiating between two macro CZs: NARRATION and
DESCRIPTION. The former is aimed at reporting
temporally interrelated (dynamic) events, the latter
is used to comment by focusing on selected
entities, properties, and states of affairs. Each of these
Foreground (FGR): text span containing
the most salient events, i.e. those in the
focus of attention
Follow-up (FUP): reactions and consequences to FGR events (to whom they’re related through cause-effect and temporal succession relations), i.e. relationally new information moving the discourse forward.
Expectation (EXP): assumptions and probable or possible outcomes, i.e. non factual information (e.g. conditionals, modality).
The class DESCRIPTION (DSCR) includes the following zones:
Description (DES): characteristics of a person or an object, customary circumstances, or states of affairs.
Evaluation (EVL): subjective descriptions, explicit judgements showing the author or some other agent’s attitude towards a target.
In addition, a third macro-class is posited, OTHER (OTHR), containing categories performing auxiliary functions towards the other CZs:
Attribution (ATT): text span containing
the source and, if present, the cue of an
attribution
Metatext (MTX): text span guiding the reader’s attention towards metatextual elements like figures or tables.
Interrogative (INT): questions directly addressed to the reader, e.g. to introduce a new topic or to prompt a reaction.
Tense PRESENT PAST FUTURE PRESPART PASTPART NONE 46 21 6 2 6
Major approaches to functional discourse structuring adopt the sentence or the paragraph as unit of annotation. On the other hand, we have opted for a clause level annotation as this allows us to better deal with news articles’ high level of information density. Although CZs are conceptually nonoverlapping, empirical analysis indicates that an annotation unit may fit into more than one category, that is a clause may represent complex contents. Cases as such suggest that the more informative content should be preferred. In the example below, the tag ATT is assigned, even though a descriptive content may as well be recognized. 1. [On an office wall of the Senate intelligence committee hangs a quote from Chairman David Boren,]AT T {PDTB2, wsj_0771} The annotation of CZs is further complicated by the fact the distribution of the zones does not follow the linear order of the text. In most cases, CZs are discontinuous, that is either their contiguity may be “broken” by the presence of other CZs or the same CZ may appear in different sentences along the entire document (see example ?? for the FGR zone).
2. [South Korea registered a trade deficit of $101 million in October,]F GR [reflecting the country’s economic sluggishness,]EV L [according to government figures released Wednesday.]AT T [Preliminary tallies by the Trade and Industry Ministry showed another trade deficit in October, the fifth monthly setback this year,]F GR [casting a cloud on South Korea’s export-oriented economy.]EV L {PDTB, wsj_0011}
In other cases, due to the use of the clause as minimal annotation span of a CZ, nested CZs may occurr (see example ??).
3. [South Korea’s economic boom, [which began in 1986,]BGR stopped this year because of prolonged labor disputes, trade conflicts and sluggish exports.]BGR {PDTB, wsj_0011} 4
We used the CZs annotation schema and the
annotation tool CAT
(20 from the test section of TempEval-3
To test our hypotheses about some formal properties of CZs, we carried out a corpus study. The results are reported below.
Position in the text. 71.7% of FGRs are located in the opening sections of the articles and their occurrence decreases towards the central (18.4%) and closing sections (9.8%). BGRs show a fairly complementary distribution to FGRs, as they mostly occupy the central (31.6%) and closing sections (27.3%) of the articles. As expected, ATTs are quite evenly distributed among the three sections. The remaining CZs do not show any clear-cut tendencies.
Verbal tenses. Table ?? shows the distribution target of verbal tenses, as annotated in the TimeBank corpus, among CZs. BGRs and ATTs are dominated by the past tense, this is in accordance with our expectations as the former is characterized by temporal precedence relations to FGR events and the second mostly contains events of saying. CZs belonging to the DSCR class are significantly dominated by the present tense, usually associated with imperfective aspect and staticity. The high frequency of present tenses in FGRs and BGRs doesn’t necessarily defy our expectations, since FGRs contain both dynamic and static events and the tag PRESENT is also used to refer to instances of present perfect.
Modality markers. The majority of modality markers is located in EXPs and, more broadly, in the narrative CZs, as shown in Figure ??. In the TimeBank corpus, the MODALITY tag is mostly assigned to modal auxiliaries, we believe that the annotation of modal adverbs would further raise the percentages observed in EXPs and in the NARR class.
Pronouns. Looking at Figure ?? we can see that almost 50% of all pronouns is located in BGRs. The percentages are consistent with our expectations as BGRs convey referentially old information and, although FUPs and EXPs elaborate on FGR events, they often introduce new referents. Note that the distribution of pronouns is not, alone, a sufficient indicator of referential oldness since also lexical and zero anaphoras should be taken into account.
In news articles events are not iconically presented in the linear order of their real succession, this poses a challenge to systems aimed at uncover43:93 47:32 0 20 30 40 50
Percentage (%) Figure 1: Distribution of modality markers and pronouns among CZs. ing their temporal event structure. Therefore, we used the annotations available for the TimeBank section of the corpus to check whether some connections between CZs and temporal relations between event pairs exist. The full set of temporal relations specified in TimeML contains 14 types of relations: BEFORE, AFTER, IBEFORE, IAFTER, BEGINS, BEGUN_BY, ENDS, ENDED_BY, DURING, DURING_INV, INCLUDES, IS_INCLUDED, SIMULTANEOUS and IDENTITY. We simplified the set as follows: the relation types that invert each other were collapsed into a single type; given the low frequency of the relation type IBEFORE, it was mapped to the corresponding more coarsegrained type BEFORE.
Given the narrative shape of news articles, the corpus is considerably dominated by precedence source - target (BEFORE) and succession (AFTER) relations. Table ?? shows that the majority of temporal relations holds between events belonging to the same CZ types: events tend to precede, include, occur during, begin, end, be simultaneous to and anaphorically evoke (through TimeML IDENTITY temporal relations) other events belonging to the same zone.
FGR events precede rather than follow ATT, FUP and EXP events. BGR events, the most involved in BEFORE relations, tend to precede other events, especially if located in ATTs and FGRs. Unexpected outcomes mostly occur in cases like the following, where the FGR event precedes the BGR one. This is because conflicting contents may be expressed in the same unit (in this case a reaction to the FGR event and the list of its premises): 4. [Delta Air Lines earnings soared to 33% to a record in the fiscal first quarter,]F GR [bucking the industry trend toward declining profits.]F UP [The Atlanta-based airline, the third largest in the U.S., attributed the increase to higher passenger traffic, new international routes and reduced service by Rival Eastern Airlines...]BGR {PDTB, wsj_1011} As highlighted in Table ??, NARR events begin or end other NARR or DSCR events (more specifically, these relations hold between events belonging to instances of the same CZ) and DSCR events include (rather than being included in) other events.
IDENTITY relations mostly involve FGRs: as 6
We have developed an inventory of zone labels for the genre news article and shown that the so-generated content structure could help narrowing down the range of time relations connecting events.
Future work would involve testing the
stability and reproducibility of the annotation scheme
through the measurement of inter-annotator
agreement and elaborating a separate annotation
scheme for editorials, whose argumentative style
reflects different structuring principles than those
acting in news reports. Finally, we would like
to automatize the process of annotation and test
the effectiveness of the approach in texts
belonging to different genres, e.g. novels
This has been partially supported by the Erasmus + Traineeship Program 2015/2016 from University of Pavia and the NWO Spinoza Prize project Understanding Language by Machines (sub-track 3).
3https://github.com/cltl/ContentZones. git