Events are Not Simple: Identity, Non-Identity, and Quasi-Identity

Despite considerable theoretical and computational work on coreference, deciding when two entities or events are identical is very difficult. In a project to build corpora containing coreference links between events, we have identified three levels of event identity (full, partial, and none). Event coreference annotation on two corpora was performed to validate the findings. 1 The Problem of Identity Last year we had HLT in Montreal, and this year we did it in Atlanta. Does the “did it” refer to the same conference or a different one? The two conferences are not identical, of course, but they are also not totally unrelated—else the “did it” would not be interpretable. When creating text, we treat instances of entities and events as if they are fixed, well-described, and well-understood. When we say “that boat over there” or “Mary’s wedding next month”, we assume the reader creates a mental representation of the referent, and we proceed to refer to it without further thought. However, as has been often noted in theoretical studies of semantics, this assumption is very problematic (Mill, 1872; Frege 1892; Guarino, 1999). Entities and (even more so) events are complex composite phenomena in the world, and they undergo change. 1 This work was supported by grants from DARPA and NSF, as well as by funding that supported Prof. M. Felisa Vedejo from UNED Madrid. Since nobody has complete knowledge, the author’s mental image of the entity or event in question might differ from the reader’s, and from the truth. Specifically, the properties the author assumes for the event or entity might not be the ones the reader assumes. This difference has deep consequences for the treatment of the semantic meaning of a text. In particular, it fundamentally affects how one must perform coreference among entities or events. As discussed in Section 6, events have been the focus of study in both Linguistics and NLP (Chen and Ji, 2009; Bejan and Harabagiu, 2008, 2010; Humphreys et al., 1997). Determining when two event mentions in text corefer is, however, an unsolved problem. Past work in NLP has avoided some of the more complex problems by considering only certain types of coreference, or by simply ignoring the major problems. The results have been partial, or inconsistent, annotations. In this paper we describe our approach to the problem of coreference among events. In order to build a corpus containing event coreference links that is annotated with high enough inter-annotator agreement to be useful for machine learning, it has proven necessary to create a model of event identity that is more elaborate than is usually assumed in the NLP literature, and to formulate quite specific definitions for its central concepts. 2 In this work, we mean both events and states when we say “event”. A state refers to a fixed, or regularly changing, configuration of entities in the world, such as “it is hot” or “he is running”. An event occurs when there is a change of state in the world, such as “he stops running” or “the plane took off”. Event coreference is the problem of determining when two mentions in a text refer to the ‘same’ event. Whether or not the event actually occurred in reality is a separate issue; a text can describe people flying around on dragons or broomsticks. While the events might be actual occurrences, hypothesized or desired ones, etc., they exist in the text as Discourse Elements (DEs), and this is what we consider in this work. Each DE is referred to (explicitly or implicitly) in the text by a mention, for example “destroy”, “the attack”, “that event”, or “it”. But it is often unclear whether two mentions refer to the same DE or to closely related ones, or to something altogether different. The following example illustrates two principal problems of event coreference: While Turkish troops have been fighting_E.1 a Kurdish faction in northern Iraq, two other Kurdish groups have been battling_E.2 each other. A radio station operated_E.3 by the Kurdistan Democratic Party said_E.4 the party's forces attacked_E.5 positions of the Patriotic Union of Kurdistan on Monday in the Kurdish region's capital Irbil. The Voice of Iraqi Kurdistan radio, monitored_E.6 by the British Broadcasting Corp., said_E.7 more than 80 Patriotic Union fighters were killed_E.8 and at least 150 wounded_E.9. The fighting_E.10 was also reported_E.11 by a senior Patriotic Union official, Kusret Rasul Ali, who said_E.12 PUK forces repelled_E.13 a large KDP attack_E.14.