English. This paper reports on an annotation experiment to enrich an existing temporally annotated corpus of Italian news articles with Narrative Containers, annotation devices representing temporal windows in text and marking up very informative temporal relations between temporal entities. The annotation has shown that the distribution of Narrative Containers is sensitive to the text genre and may be used to facilitate the creation of informative timelines.
Italiano. Questo lavoro illustra i risultati di un esperimento di annotazione per l’identificazione di Contenitori Narrativi, ovvero marcatori di “finestre” temporali in un testo, come strategia per arricchire un corpus di articoli di quotidiano in lingua italiana, gia` annotato con informazioni temporali. L’annotazione ha mostrato che la distribuzione dei Contenitori Narrativi e` legata al genere testuale e puo` essere usata per facilitare la creazione di linee temporali di eventi piu` informative.
Research in Temporal Processing has seen an
increasing interest thanks to the availability of
annotation schemes and corpora in multiple
languages
The remainder of this paper is structured as follows: the EVENTI corpus will be shortly introduced in Section 2, with a particular emphasis on the available temporal relations. Section 3 will present the notion of Narrative Container and the proposed annotation scheme. In Section 4 the results of a pilot annotation on the EVENTI dataset will be reported. Finally, conclusion, future work, and a pointer to the annotated data and guidelines will be reported in Section 5. 2
The EVENTI corpus, released in the context of the
EVALITA 20143 workshop, consists of 3 datasets:
the Main task training data, the Main task test
data, and the Pilot task test data. The corpus
has been annotated with a simplified version of
the It-TimeML Annotation Guidelines
The EVENT tag is used to annotate all lexical items which may realize an event mention. It includes verbs, nouns, adjectives, and prepositional phrases. The tag is enriched with 8 attributes expressing tense, (grammatical) aspect, part-ofspeech, mood, modality, verb form, TimeML class, and polarity.
2https://sites.google.com/site/ eventievalita2014/
3http://www.evalita.it/2014/tasks/ eventi
The TIMEX3 tag is used for the annotation of temporal expressions (timexes), expressing the type, the value and whether the timex is absolute or relative (e.g. “2015-05-18” vs. “yesterday”[ieri]).
The SIGNAL tag is employed to mark any linguistic elements, such as prepositions (e.g. in [in]), adverbs (e.g. before [prima]), or conjunctions (e.g. when [quando]), which support the identification and classification of a temporal relation between target entities (e.g. events and timexes).
Finally, the TLINK tag is used to annotate temporal relations. In the EVENTI task, the subset of possible temporal relations has been restricted to three subtypes of intra-sentence relations, namely: i.) pairs of syntactic main events in the same sentence; ii.) pairs of syntactic main event and subordinate event in the same sentence; and iii.) pairs of event and timexes. All 13 temporal relation values from It-TimeML (BEFORE, AFTER, IS INCLUDED, INCLUDES, SIMULTANEOUS, I(MMEDIATELY) AFTER, I(MMEDIATELY) BEFORE, IDENTITY, MEASURE, BEGINS, ENDS, BEGUN BY and ENDED BY) have been used.
The Main task datasets, which have been enriched with Narrative Containers, add up to 130,279 tokens, divided into 103,593 tokens for training and 26,686 for test. They contain 21,633 EVENTs (17,835 in training and 3,798 in test), 3,359 TIMEX3 (2,753 in training and 624 in test), 1,163 SIGNALs (923 in training and 231 in test), and 4,561 TLINKs (3,500 in training and 1,061 in test). 3
The notion of Narrative Container (NC) was first introduced by Pustejovsky and Stubbs (2011) to deal with some aspects of Temporal Processing, such as sensitivity to the text genre and interaction with discourse relations, not addressed in the TimeML Guidelines nor in the TimeBank corpus. NCs were proposed as a temporal window, providing left and right boundaries, to when events not anchored to timexes could have happened, thus overcoming issues related to linking of events with the Document Creation Time (DCT), i.e. when a text was written or published. In particular, standard TimeML markup imposes that all events have a link with the DCT but fail to specify that each event should also be annotating to its actual temporal anchor, i.e. to its moment of occurrence. As reported in Pustejovsky and Stubbs (2011), in example 1, TimeML guidelines will order both event mentions, e1 and e2, to the DCT with a BEFORE relation, anchor e1 to the timex “yesterday” (t) but will fail to provide the anchoring of e2: 1. The bomb explodede1 yesterdayt2011 09 09 and killede2 three people. [DCT=2011-0910]
A further justification to the introduction of NCs is related to the different informational status of temporal relations. Assuming the informativeness of a temporal link as a function of the information contained in the individual links and their closure, an anchoring relation, that is a relation between a timex and an event explicitly stating when the event occurred as the one between e1 and timex “yesterday” in example 1 (i.e. a temporal value of INCLUDES or IS INCLUDED), is assumed to be more informative than an ordering relations, i.e. a precedence relation between two events.
To the best of our knowledge, the only corpus
which extensively adopts the notion of NC and has
available annotated data is the THYME corpus of
clinical narratives
A NC enables an accurate reproduction of the way events in text cluster around temporal reference points, explicitly or implicitly realized in the document, as the narration unfolds. NC relations are thus anchoring relations between pairs of events or events and temporal expressions. They are marked with an additional link tag, i.e. CONTAINS, to distinguish them from standard TLINKs. Each NC relation admits two components: i.) the narrative anchor, i.e. an element pointing to a specific temporal dimension shared by other events or timexes within the text; and ii.) the anchored element(s), i.e. events which satisfy the anchorability requirements (see Section 3.1 for details) and participate in an NC relation. Timex anchors are chosen on a transparency basis (i.e. granularity and nature of the timex), whereas Event anchors are chosen according to their relevance and salience for the timeline.
Two sub-types of NCs can be identified: Temporal Containers (TCs): they correspond to the timexes in the text which clearly anchor the events in analysis on a timeline; the relation can hold both at intra- and intersentence level. Example 2 from our annotated corpus shows a timex (2001) and the events it anchors (e1–e4): 2. [...] la Sonata composta[e1] nel 2001[TCanchor], il cui primo esecutore fu[e2] lo stesso Lucchesini. In questa esecuzione[e3] si ritrovavano[e4] gia` tutte le doti musicali di Lucchesini [...].
Event Containers (ECs): they correspond to event mentions which function as a temporal anchor for other event mentions. ECs can be useful in cases where no anchoring timex is available or to model event-subevent relations. Example 3 shows a sentence with no explicit temporal expression, where the anchoring of events (e1–e3) is possible only with respect to the event (ricognizione).
il tenente ha dato disposizioni[e1] per il presidio, e nella fase[e2] iniziale ha ordinato[e3] ai sottoposti di fare rapporto al campo base.
Naturally, the NC represented here is only a
visual aid picturing the conceptual outcome of
applying CONTAINS relations between the
anchor (here, the TIMEX3 2001) and anchored
elements (here, EVENTs composta, fu, esecuzione,
and ritrovavano)
The set of events which can be anchored has been
restricted to factual events. The identification of
eligible anchorable events has been manually
conducted at this stage of the annotation. We adopted
the definition of factuality as proposed in the
FactBank
We excluded all events which are presented as subjective (i.e. judgements, opinions). In example 4, esplosione is a factual event and was anchored as such, whereas sbagliato describes it through the grid of the writer’s judgement, who states that the explosion was a mistake, and thus not anchored.
4. L’esplosionee1 e` avvenuta a mezzanotte di luned`ı [...]. Insomma, gli attentatori hanno sbagliatoe2 obiettivo.
Finally, generic events, i.e. events which acquire some kind of attributive value towards discourse participants, expressing persistent properties or reiterated, habitual activities, were not anchored. 4
The EVENTI-NC corpus includes documents from both the training and the test sections of the Main task of the EVENTI corpus. It includes 58 annotated articles, for a total of 24.259 tokens, covering roughly 11% of the EVENTI corpus; Table 1 shows the number of EVENTs and TIMEX3 involved in our annotation.
Table 2 reports the number of annotated containers in our corpus, and their distribution according to their type. TCs make up for almost 63% of the total number of NCs, against the 37% of ECs. General EVENTI-NC statistics
Annotated articles
TIMEX3 markables
It is interesting to notice that 11.8% of TCs is realized by empty TIMEX3s, i.e. temporal expressions which do not correspond to lexical items but can be inferred and which are necessary to for assigning a correct value to a timex. We conducted an in-depth analysis of the NC anchors following two parameters: i.) the properties of NC anchors on their own; and ii.) the sensitivity to the document genre, i.e. the news domain, on the line of Pustejovsky and Stubbs (2011).
Concerning the first parameter, we first investigated the incidence of verbal anchors as opposed to nominal anchors. Whereas there appear to be no tendency towards verb or nouns being more likely to anchor other events, it is interesting to take a look within these categories. Out of all the verbal anchors, 42.9% are reporting verbs or verbs employed in a declarative context. We observed that there is a preference for ECs to correspond to the the event with the highest degree of topicality in the article, or the most important event (climax event). For example, one article4 reports on President Clinton’s surgery in 2004: the largest EC in the document is anchored by intervento (surgery), with a total of 12 anchored items.
Sensitivity to text genre can be easily observed with TC anchors. 25% of them anchor events in a timespan that can be measured as 1 day with respect to the DCT. Anchors for these containers are mostly represented by non absolute temporal expressions, such as temporal adverbs (e.g. “ieri” [yesterday], “domani” [tomorrow], among others) and by the DCT itself, which represents 11% of the TC anchors.
Genre-sensitivity might also be the factor behind the average number of NCs in the corpus. The documents have an average of 5.17 NCs, and even for more lengthy articles, the textual anchors were rarely more than 7. The average of 5.17 NCs/article might be due to the fact that newspaper articles usually refer to a limited number of facts, whose core is usually made of a handful of recent happenings; whereas the fluctuating relationship between length and NC number usually depends both on the content of the article and on the granularity of the selected NC. 5
This paper reports on a first proposal of an
annotation scheme and accompanying annotated data for
NCs in Italian news articles. The NC annotation is
an additional layer on top of already available data
for Temporal Processing in Italian. It addresses
pending issues (e.g. the annotation of the
temporal relations between event and the Document
Creation Time) and increases the informativeness
of the document timelines. Overall, we observed
that there is a preference for NC to be realized by
timexes in a limited time span. However, NCs may
also be realised by events. In this case, nouns and
verbs have a similar distribution with a preference
for events which have a central role in the news or
facilitate the clustering of the information (e.g.
reporting events). Such a behaviour is different with
respect to clinical narratives where nominal events
are more frequently selected as NC
Future work will aim at assessing the reliability of the proposed scheme via an inter-annotator agreement study and at completing the annotation of the entire EVENTI corpus. Finally, the annotated data and guidelines are publicly available 5 to encourage additional testing and experiments.
5https://sites.google.com/ site/ittimeml/documents/ narrative-container-data.zip? attredirects=0&d=1