Version classiqueVersion mobile
OpenEdition Books

Proceedings of the Second Italian Conference on Computational Linguistics CLiC-it 2015

 | 
Cristina Bosco
, 
Sara Tonelli
, 
Fabio Massimo Zanzotto

Cross-language projection of multilayer semantic annotation in the NewsReader Wikinews Italian Corpus (WItaC)

Manuela Speranza et Anne-Lyse Minard

Résumé

In this paper we present the annotation of events, entities, relations and coreference chains performed on Italian translations of English annotated texts. As manual annotation is a very expensive and time-consuming task, we devised a crosslingual projection procedure based on the manual alignment of annotated elements.

In questo articolo descriviamo l’annotazione degli eventi, delle entità, delle relazioni e delle catene di coreferenza realizzata su traduzioni in italiano di testi inglesi già annotati. Essendo l’annotazione manuale un compito molto dispendioso, abbiamo ideato una procedura di proiezione interlinguale basata sull’allineamento degli elementi annotati.

Texte intégral

This research was partially funded the EU NewsReader project (FP7-ICT-2011-8 grant 316404).

1. Introduction

  • 1 Wikinews (http://en.wikinews.org) is a collection of multilingual online news articles written coll (...)

1The NewsReader Wikinews Italian Corpus (WItaC) is a new Italian annotated corpus consisting of English articles taken from Wikinews1 and translated into Italian by professional translators.

2The English corpus was created and annotated manually within the NewsReader project,2 whose goal is to build a multilingual system able to reconstruct storylines across news articles in order to provide policy and decision makers with an overview of what happened, to whom, when, and where. Semantic annotations in the NewsReader English Wikinews corpus span over multiple levels, including both intra-document annotation (entities, events, temporal information, semantic roles, and event and entity coreference) and cross-document annotation (event and entity coreference). As manual annotation is a very expensive and timeconsuming task, we devised a procedure to automatically project the annotations already available in the English texts onto the Italian translations, based on the manual alignment of the annotated elements in the two languages.

3The English corpus, taken directly from Wikinews, together with WItaC, being its translation, ensures access to non-copyrighted articles for the evaluation of the NewsReader system and the possibility of comparing results in the two languages at a finegrained level.

  • 3 The NewsReader consortium has annotated also the Spanish and Dutch translations of the same Wikinew (...)

4WItaC aims at being a reference for the evaluation of storylines reconstruction, a task requiring several subtasks, e.g. semantic role labeling (SRL) and event coreference. In addition, it is part of a cross-lingually annotated corpus,3thus enabling for experiments across different languages.

5The remainder of this article is organized as follows. We review related work in Section 2. In Section 3 we present the annotations available in the English corpus used as the source for the projection of the annotation. In Section 4 we detail some adaptations of the guidelines specific for Italian. In Sections 5 and 6 we describe the annotation process and the resulting WItaC corpus. Finally, we conclude presenting some future work.

Related work

6A number of semantically annotated corpora are available for English, whereas most other languages are under-resourced. As far as Italian is concerned, WItaC is the first corpus offering annotations of entities, events, and event factuality, together with semantic role labeling and crossdocument coreference annotation.

7For entities and entity coreference, the reference Italian corpus is I-CAB (Magnini et al., 2006), which is annotated with entities, time expressions (following the TIMEX2 standard), and intradocument entity coreference; for cross-document person entity coreference, we refer to CRIPCO (Bentivogli et al., 2008). Regarding temporal information and event factuality, two annotated corpora are available: respectively, the EVENTI corpus (Caselli et al., 2014), used as the evaluation dataset for the EVENTI task at Evalita 2014 and annotated with events, time expressions (TIMEX3), temporal signals, and temporal relations, and FactIta Bank (Minard et al., 2014), a subsection of EVENTI annotated with event factuality.

8To the best of our knowledge there exist no other Italian corpora with semantic role labeling and event cross-document coreference annotation. The reference corpus for SRL in English is the CoNLL-2008 corpus (Surdeanu et al., 2008). For cross-document coreference, the ECB+ corpus (Cybulska and Vossen, 2014) has recently been created extending the ECB corpus.

9The method we propose for cross-lingual annotation projection taking advantage of the alignment between texts in two different languages is similar to other methods used, for example, to build annotated corpora with semantic roles (Pado and Lapata,´ 2009), temporal information (Spreyer and Frank, 2008; Forascu and Tufi, 2012), and coreference chains (Postolache et al., 2006). However, previous work is based on the use of corpora aligned at the word level either manually, which is very timeconsuming, or automatically, which is error prone. On the other hand, our method envisages a manual alignment at the markable level, where the extent of each element is annotated on the translated text and then aligned to the English annotated element on a semantic rather than syntactic basis.

3. Annotation available in the English source corpus

10The NewsReader Wikinews English corpus contains intra-document semantic annotation and crossdocument coreference annotation.

3.1 Annotation at document level

11The annotation is based on the NewsReader guidelines (Tonelli et al., 2014) and was performed using the CAT tool (Bartalesi Lenzi et al., 2012). The first five sentences (including the headline) of each document contain the following annotations: markables, relations, and intra-document coreference.

Markable annotation. Textual realizations of entity instances, referred to as entity mentions, are the portions of text in which entity instances of different types (people, organizations, locations, financial entities, and products) are referenced within a text. Each entity mention is described through that portion of text (extent) and two optional attributes, i.e. syntactic head and syntactic type.

The textual realization of an event, the event mention, can be a verb, a noun, a pronoun, an adjective, or a prepositional construction. It is annotated through its extent and a number of attributes, e.g. predicate (lemma), part-of-speech, and factuality. Factuality attributes (van Son et al., 2014) of an event include time, certainty and polarity.

The annotation of temporal expressions is based on the ISO-TimeML guidelines (ISO, 2012) , and thus includes durations, dates (e.g. the document creation time), times, and sets of times, with the following attributes : type, normalized value, anchorTimeID (for anchored temporal expressions), and beginPoint and endPoint (for durations).

Numerical expressions include percentages, amounts described in terms of currencies, and general amounts. Temporal signals, inherited from ISO-TimeML, make explicit a temporal relation. Similarly, causal signals (C-SIGNALS) indicate the presence of a causal relation between two events (e.g. because of, since, as a result, and the reason why).

Relation annotation. Based on the TimeML approach (Pustejovsky et al., 2003), temporal relations (e.g. ‘before’, ‘after’, ‘includes’, and ‘ends’) are used to link two event mentions, two temporal expressions or an event mention and a temporal expression. The annotation of subordinating relations also leans on TimeML, although its scope was reduced to the annotation of reported speech.

In addition, explicit causal relations between causes and effects denoted by event mentions have been annotated taking into consideration the cause, enable, and prevent categories of causation, and grammatical relations have been created for events that are semantically dependent on another event, to link them to their governing content verb/noun. Semantic role labeling is modeled through the HAS_PARTICIPANT relation, a one-to-one relation linking an event mention to an entity mention playing a role in the event. PropBank (Bonial et al., 2010) is used as the reference framework for the assignment of the semantic role to each relation.

Intra-document event and entity coreference. The annotation of coreference chains that link different mentions to the same instance is based on the REFERS TO relation.

Entity instances are described through the non text-consuming ENTITY tag and the two attributes entity type and tag descriptor; similarly, event instances are described through the non textconsuming EVENT tag and the two attributes event class and tag descriptor.

3.2 Annotation at corpus level

12Annotation at the corpus level (Speranza and Minard, 2014), performed using the CROMER tool (Girardi et al., 2014), relies on the creation of corpus instances (both entities and events) and on links holding between each mention and the corpus instance it refers to. Corpus instances are described through a unique instance ID and the DBpedia URI (when available). Annotation consists of:

  • cross-document entity coreference in the firstfive sentences;

  • cross-document entity and event coreference inthe whole document for a subset of 44 seed entities (i.e., annotation and coreference of all mentions referring to the seed entities and of the events of which the entities are participants).

4. Italian language specific annotations

13We adopted the NewsReader guidelines already available for English with some minor language specific adaptations, as described in detail in Speranza et al. (2014). For this reason the data on inter-annotator agreement provided for English by van Erp et al. (2015) can be used as a reference.

14For the annotation of clitics, which do not exist in English, we decided to leave the annotation at the word level, rather than split it into smaller units, so as to be consistent with annotations on existing corpora, e.g. I-CAB (Magnini et al., 2006). So in the case of a token composed of a verb (i.e. an event mention) and a clitic corresponding to a pronominal mention of a markable entity, the whole token was annotated both as an entity and as an event. The syntactic head attribute of the entity mention, having as value the clitic, and the predicate attribute of the event mention, having as value the verbal root, contribute to distinguish the two annotated elements (see [1]).

  1. Aveva già deciso di dargli un aiuto (‘He had already decided to give him some help’)
    EVENT_MENTION: [dargli], pred “dare”
    ENTITYMENTION: [dargli], head “gli”

15As Italian, unlike English, is a null-subject language where clauses lacking an explicit subject are permitted, we devised specific guidelines that allowed us to straightforwardly align English pronouns to Italian null subjects. In particular, null subjects having finite verb forms as predicates and referring to existing entity instances (see [2]) were marked through the creation of an empty (i.e. non text-consuming) ENTITY_MENTION tag, which was then linked to other markables following the guidelines for regular text consuming entity mentions; in addition, annotators filled the tag descriptor attribute with a human friendly name and the sentence number (e.g. “He-LuiS2” for the null subject in [2]).

  1. Obama fece un discorso. [Ø] Disse che [...] (‘Obama gave a speech. [He] said that [...]’)

  • 4 In WItaC modals are also linked to their governing verb through a grammatical link.

16The annotation of modals for Italian is based on It-TimeML, where they are marked as events like all other verbs.4

5. Annotation of WItaC

17The method we propose for the annotation of the Italian corpus consists of cross-lingual projection of annotation from a source corpus to a target corpus; it enabled us to reduce the effort by approximately three times. The annotation was performed in five steps starting with a file containing the source English annotated text and the Italian translation aligned at the sentence level.

  1. Mention annotation. The first step of the annotation, performed using the CAT tool, consisted of the identification and annotation of all markable extents.

    • 5 No exceptions were needed for aligning null subjects.

    Alignment. The use of CAT, which is highly customizable, enabled us to set up the alignment between Italian and English markables by simply adding to the Italian markables a new attribute which takes as value the ID of a different markable. Annotators filled this attribute with the corresponding English markable by using drag-and-drop. In some cases it was also necessary to mark the attributes and/or relations that should not be imported (by writing a note in the comment attribute), or to create extra relations.5 If a mention had no equivalent, annotators filled in the values of the attributes and created the relations in which it was involved and, if it did not already exist, the instance to which it referred.

  2. Automatic projection. The automatic projection was performed using a Python script working on the XML files produced by the CAT annotation tool. For each article, the script takes as input the file containing both the English fully annotated text and the Italian text on which the annotated markables have been aligned. It produces as output a file in which the Italian text has been enriched with the annotations imported from English, i.e. the event instances, the entity instances, the relations (including the REFERS_TO relation which models intra-document coreference), and the values of the non-language-specific attributes (unless a specific comment is present).

  3. Manual revision. Manual revision consists of an overall check of the annotations imported automatically; in particular, it involves the annotation of the language specific attributes and the deletion of the relations that had been marked as non-importable (using the CAT tool).

  4. Projection of cross document coreference. The projection of the cross-document annotation consists of importing coreference from the English corpus taking advantage of the alignment performed in the second step and extending the entity and event instances by importing the IDs of the English instances and their DBpedia URIs.

6. Dataset Description

18WItaC is composed of 120 articles. In Table 1 we give the size of the whole corpus and the size of the “first 5 sentences” section, i.e. the subsection annotated with markables, relations, intra-document coreference and cross-document entity coreference. In total 6,127 markables have been annotated in Italian; of these, 5,580 are aligned to English markables while 547 have no English correspondent.

Table 1: Italian and English corpus size

Whole corpus

First 5 sentences

Ita.

Eng.

Ita.

Eng.

# files

120

120

120

120

# sentences

1,845

1,797

597

597

# tokens

44,540

40,231

15,676

13,981

19Exploiting the alignment, relations and attributes have been imported automatically. For only 5.7% of the markables the attributes could not be projected (e.g. two events with different PoS). In Table 2 we present the number of markables and relations annotated in the Italian corpus. Out of the total 2,709 entity mentions, 56 are null subjects aligned with English pronominal entity mentions.

Table 2: Annotations in the first five sentences

Markables

Relations

EVENT_MENTION

2,208

SLINK

220

ENTITY_MENTION

2,709

TLINK

1,711

TIMEX3

507

CLINK

61

VALUE

415

GLINK

300

SIGNAL

253

HAS_PART

1,865

C-SIGNAL

35

Total

6,127

Total

4,157

Instances

Coreference chains

EVENT INSTANCE

1,773

REFERS_TO

3,054

ENTITY INSTANCE

1,281

Total

3,054

20As a result of the projection of event and entity cross-document coreference chains from English, WItaC contains 740 entity instances and 887 event instances annotated at the corpus level. Annotation by projection enables us to also have crosslingual annotation, which means that the instances are shared between English and Italian.

7. Conclusions and future work

21We have presented WItaC, a new corpus consisting of Italian translations of English texts annotated using a cross-lingual projection method. We acknowledge some influence of English in the translated texts (for instance, we noticed an above-average occurrence of noun modifiers, as in “dipendenti Airbus”) and in the annotation (for instance, annotators might have been influenced by English in the identification of light verb constructions in the Italian corpus). On the other hand, this method enabled us not only to considerably reduce the annotation effort, but also to add a new cross-lingual level to the NewsReader corpus; in fact, we now have two annotated corpora, in English and Italian, in which entity and event instances (in total, over 1,600) are shared.

22In the short-term we plan to manually revise the projected relations and add the language-specific attributes. We also plan to use the corpus as a dataset for a shared evaluation task and afterwards we will make it freely available from the website of the HLT-NLP group at FBK6 and from the website of the NewsReader project.

Bibliographie

Valentina Bartalesi Lenzi, Giovanni Moretti, and Rachele Sprugnoli. 2012. CAT: the CELCT Annotation Tool. In Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC’12), pages 333–338, Istanbul, Turkey, May. European Language Resources Association (ELRA).

Luisa Bentivogli, Christian Girardi, and Emanuele Pianta. 2008. Creating a gold standard for person cross-document coreference resolution in italian news. In LREC Workshop on Resources and Evaluation for Identity Matching, Entity Resolution and Entity Management.

Claire Bonial, Olga Babko-Malaya, Jinho D. Choi, Jena Hwang, and Martha Palmer. 2010. PropBank Annotation Guidelines, Version 3.0. Technical report, Center for Computational Language and Education Research, Institute of Cognitive Science, University of Colorado at Boulder. http://clear.colorado.edu/compsem/documents/propbank_guidelines.pdf.

Tommaso Caselli, Rachele Sprugnoli, Manuela Speranza, and Monica Monachini. 2014. EVENTI EValuation of Events and Temporal INformation at Evalita. In Proceedings of the First Italian Conference on Computational Linguistic CLiC-it 2014 & the Fourth International Workshop EVALITA 2014 Vol. II: Fourth International Workshop EVALITA 2014, pages 27–34.

Agata Cybulska and Piek Vossen. 2014. Using a sledgehammer to crack a nut? Lexical diversity and event coreference resolution. In Proceedings of the 9th International Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland, May. European Language Resources Association (ELRA).

Corina Forascu and Dan Tufi. 2012. Romanian Time-Bank: An Annotated Parallel Corpus for Temporal Information. In Proceedings of the 8th Language Resources and Evaluation Conference (LREC2012), Istanbul, Turkey, May. European Language Resources Association (ELRA).

Christian Girardi, Manuela Speranza, Rachele Sprugnoli, and Sara Tonelli. 2014. CROMER: A Tool for Cross-Document Event and Entity Coreference. In Proceedings of the 9th Internation Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland, May. European Language Resources Association (ELRA).

ISO. 2012. ISO 24617-1: Language Resource Management. Semantic Annotation Framework (SemAF). Time and Events (SemAF-Time, ISO-TimeML). ISO International Standard.

Bernardo Magnini, Emanuele Pianta, Christian Girardi, Matteo Negri, Lorenza Romano, Manuela Speranza, Valentina Bartalesi Lenzi, and Rachele Sprugnoli. 2006. I-CAB: the Italian Content Annotation Bank. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy, May. European Language Resources Association (ELRA).

Anne-Lyse Minard, Alessandro Marchetti, and Manuela Speranza. 2014. Event Factuality in Italian: Annotation of News Stories from the Ita-TimeBank. In Proceedings of the First Italian Conference on Computational Linguistic CLiC-it 2014, pages 260–264.

Sebastian Padó and Mirella Lapata. 2009. Crosslingual Annotation Projection of Semantic Roles. Journal of Artificial Intelligence Research, 36(1):307–340, September.

Oana Postolache, Dan Cristea, and Constantin Orasan. 2006. Transferring Coreference Chains through Word Alignment. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy, May. European Language Resources Association (ELRA).

James Pustejovsky, José M. Castaño, Robert Ingria, Roser Saurí, Robert J. Gaizauskas, Andrea Setzer, Graham Katz, and Dragomir R. Radev. 2003. TimeML: Robust Specification of Event and Temporal Expressions in Text. In New Directions in Question Answering, pages 28–34.

Manuela Speranza and Anne-Lyse Minard. 2014. NewsReader Guidelines for Cross-Document Annotation. Technical Report NWR2014-9, Fondazione Bruno Kessler. http://www.newsreader-project.eu/files/2014/12/NWR-2014-9.pdf.

Manuela Speranza, Ruben Urizar, and AnneLyse Minard. 2014. NewsReader Italian and Spanish specific Guidelines for Annotation at Document Level. Technical Report NWR2014-6, Fondazione Bruno Kessler. http://www.newsreader-project.eu/files/2014/02/NWR-2014-61.pdf.

Kathrin Spreyer and Anette Frank. 2008. Projectionbased Acquisition of a Temporal Labeller. In Proceedings of IJCNLP, pages 489–496, Hyderabad, India, January.

Mihai Surdeanu, Richard Johansson, Adam Meyers, Lluís Màrquez, and Joakim Nivre. 2008. The CoNLL-2008 Shared Task on Joint Parsing of Syntactic and Semantic Dependencies. In Proceedings of the Twelfth Conference on Computational Natural Language Learning, CoNLL ’08, pages 159–177, Stroudsburg, PA, USA. Association for Computational Linguistics.

Sara Tonelli, Rachele Sprugnoli, Manuela Speranza, and Anne-Lyse Minard. 2014. NewsReader Guidelines for Annotation at Document Level. Technical Report NWR2014-2-2, Fondazione Bruno Kessler. http://www.newsreader-project.eu/files/2014/12/NWR-2014-2-2.pdf.

Marieke van Erp, Piek Vossen, Rodrigo Agerri, Anne-Lyse Minard, Manuela Speranza, Ruben Urizar, Egoitz Laparra, Itziar Aldabe, and German Rigau. 2015. Annotated Data, version 2. Technical Report D3-3-2, VU Amsterdam. http://www.newsreader-project.eu/files/2012/12/NWR-D3-3-2.pdf.

Chantal van Son, Marieke van Erp, Antske Fokkens, and Piek Vossen. 2014. Hope and Fear: Interpreting Perspectives by Integrating Sentiment and Event Factuality. In Proceedings of the 9th International Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland, May. European Language Resources Association (ELRA).

Notes

1 Wikinews (http://en.wikinews.org) is a collection of multilingual online news articles written collaboratively in a wiki-like manner.

2 http://www.newsreader-project.eu/

3 The NewsReader consortium has annotated also the Spanish and Dutch translations of the same Wikinews articles.

4 In WItaC modals are also linked to their governing verb through a grammatical link.

5 No exceptions were needed for aligning null subjects.

6 https://hlt-nlp.fbk.eu/technologies

Auteurs

Fondazione Bruno Kessler, Trento - manspera@fbk.eu

Fondazione Bruno Kessler, Trento - minard@fbk.eu

Lire

Open access

Acheter