Event Coreference Resolution Task and Features Analysis
Event Coreference Resolution Task involves analyzing various events, such as explosions and bombings, to determine connections and resolve references. Features include event type, triggers, similarities, distances between events, argument overlaps, and more. Incorporating event attributes as features enhances the analysis process, enabling the classification of event mention pairs. The content also delves into specific event mentions, like the explosion in a cafe, and attributes such as modality, polarity, genericity, tense, and conflict resolution.
Download Presentation
Please find below an Image/Link to download the presentation.
The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. Download presentation by click this link. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.
E N D
Presentation Transcript
Event Coreference Resolution Heng Ji (UIUC)
Event Coreference Resolution: Task 1. An explosion in a cafe at one of the capital's busiest intersections killed one woman and injured another Tuesday 4. Ankara police chief Ercument Yilmaz visited the site of the morning blast 5. The explosion comes a month after 2. Police were investigating the cause of the explosion in the restroom of the multistory Crocodile Cafe in the commercial district of Kizilay during the morning rush hour 6. a bomb exploded at a McDonald's restaurant in Istanbul, causing damage but no injuries 7. Radical leftist, Kurdish and Islamic groups are active in the country and have carried out the bombing in the past 3. The blast shattered walls and windows in the building
Typical Event Mention Pair Classification Features Category Feature Description Event type type_subtype pair of event type and subtype Trigger trigger_pair trigger pairs pos_pair part-of-speech pair of triggers nominal if the trigger of EM2 is nominal exact_match if the triggers exactly match stem_match if the stems of triggers match trigger_sim trigger similarity based on WordNet Distance token_dist the number of tokens between triggers sentence_dist the number of sentences between event mentions event_dist the number of event mentions between EM1and EM2 the number of arguments with entity and role match Argument overlap_arg unique_arg the number of arguments only in one event mention diffrole_arg The number of coreferential arguments but role mismatch
Incorporating Event Attribute as Features Event Attributes Event Mentions Attribute Value Toyota Motor Corp. said Tuesday it will promote Akio Toyoda, a grandson of the company's founder who is widely viewed as a candidate to some day head Japan's largest automaker. Other Modality Managing director Toyoda, 46, grandson of Kiichiro Toyoda and the eldest son of Toyota honorary chairman Shoichiro Toyoda, became one of 14 senior managing directors under a streamlined management system set to be At least 19 people were killed in the first blast Asserted Polarity Positive There were no reports of deaths in the blast Negative An explosion in a cafe at one of the capital's busiest intersections killed one woman and injured another Tuesday Specific Genericity Roh has said any pre-emptive strike against the North's nuclear facilities could prove disastrous Generic Tense Israel holds the Palestinian leader responsible for the latest violence, even though the recent attacks were carried out by Islamic militants Past We are warning Israel not to exploit this war against Iraq to carry out more attacks against the Palestinian people in the Gaza Strip and destroy the Palestinian Authority and the peace process. Future Attribute values as features: Whether the attributes of an event mention and its candidate antecedent event conflict or not; 6% absolute gain (Chen et al., 2009)
Clustering Method 1: Agglomerative Clustering Basic idea: Start with singleton event mentions, sort them according to the occurrence in the document Traverse through each event mention (from left to right), iteratively merge the active event mention into a prior event (largest probability higher than some threshold) or start the event mention as a new event
Clustering Method 2: Spectral Graph Clustering explosion blast Trigger Arguments Trigger Arguments Role = Place Role = Time a cafe Tuesday Role = Place Role = Time site morning explosion Trigger Arguments explosion Trigger Arguments Role = Time a month after Role = Place Role = Time restroom morning rush hour exploded Trigger Arguments Role = Place restaurant explosion Trigger Arguments Role = Place building bombing Trigger Arguments Role = Attacker groups (Chen and Ji, 2009)
Spectral Graph Clustering 0.8 A 0.7 0.9 0.8 0.9 0.6 0.3 0.8 0.2 0.7 0.2 0.3 B 0.1 cut(A,B) = 0.1+0.2+0.2+0.3=0.8
Spectral Graph Clustering (Cont) Start with full connected graph, each edge is weighted by the coreference value Optimize the normalized-cut criterion (Shi and Malik, 2000) ( , ) min ( , ) NCut A B vol A ( , ) ( ) vol B cut A B cut A B = + ( ) vol(A): The total weight of the edges from group A Maximize weight of within-group coreference links Minimize weight of between-group coreference links
Performance MUC metric does not prefer clustering results with many singleton event mentions (Chen and Ji, 2009)
Remaining Challenges The performance bottleneck of event coreference resolution comes from the poor performance of event mention labeling
Beyond ACE Event Coreference Annotate events beyond ACE coreference definition ACE does not identify Events as coreferents when one mention refers only to a part of the other In ACE, the plural event mention is not coreferent with mentions of the component individual events. ACE does not annotate: Three people have been convicted Smith and Jones were found guilty of selling guns The gunman shot Smith and his son. ..The attack against Smith.
CMU Event Coref Corpus Annotate related events at the document level, including subevents. Examples: drug war (contains subevents: attacks, crackdowns, bullying ) attacks (contains subevents: deaths, kidnappings, assassination, bombed )
Graph Decoding (Liu et al., 2018) 13
Graph Decoding (Liu et al., 2018) 14
Adding Event Sequencing Features Surface-Based Script Compatibility: these features capture whether two mentions are script compatible based on the surface information, including: Mention headword pair. Event type pair. Whether two event mentions appear in the same cluster in Chambers s event schema database (Chambers and Jurafsky, 2010). Whether the two event mentions share arguments, and the semantic frame name of the shared argument (produced by the Semafor parser (Das and Smith, 2011)). Discourse-Based Script Compatibility: these features capture whether two event mentions are related given the discourse context. Dependency path between the two mentions. Function words (words other than Noun, Verb, Adjective and Adverb) in between the two Mentions. The types of other event mentions between the two mentions. The sentence distance of two event mentions. Whether there are temporal expressions (AGM-TMP slot from a semantic parser (Tratz and Hovy, 2011)) in the sentences of the two mentions. Event Ordering: this feature set tries to capture the ordering of events. We use the discourse ordering of two mentions (forward: the antecedent is the parent; backward: the antecedent is the child), and temporal ordering produced by Caevo (Chambers et al., 2014). 15
Results (Liu et al., 2018) 16
Weak Supervision (Peng et al., 2016) 17
Event Representation (Peng et al., 2016) 18
Domain Transfer Results (Peng et al., 2016) 19
Incorporating Argument Matching (Huang et al., 2019) 20
Incorporating Argument Matching (Huang et al., 2019) 21
Incorporating Argument Matching (Huang et al., 2019) 22
Incorporating Argument Matching (Huang et al., 2019) 23
Incorporating Argument Matching (Huang et al., 2019) 24
Incorporating Argument Matching (Huang et al., 2019) 25
Event Mention Pairwise Similarity Threshold >=0.8 They share many coreferential entity arguments Its now clear that many of the anti-junta activists who occupied the Trade Unions House were neither burned to death nor died of smoke inhalation , but were savagely shot at point-blank range by agents and thugs who had infiltrated the building to kill as many of the occupants as possible , burn the corpses , and then slip away without notice . The article , of course , fails to explain how many of the people inside the building were either shot or strangled to death . But how many will have died by then ? 26
Event Mention Pairwise Similarity Threshold >=0.8 They share many coreferential entity arguments Its now clear that many of the anti-junta activists who occupied the Trade Unions House were neither burned to death nor died of smoke inhalation , but were savagely shot at point-blank range by agents and thugs who had infiltrated the building to kill as many of the occupants as possible , burn the corpses , and then slip away without notice . ( Odessa Building Fire Kills Dozens , AP ) In a piece titled Deadly Ukraine Fire Likely Sparked by Rebels , Government Says , the WSJ pushes the improbable theory that the anti- coup activists inside the building actually burned the building down themselves , a pathetic attempt to blame the victims of a ruthless government crackdown . 27
Event Mention Pairwise Similarity [0.3, 0.8) HC000030E: super-event and subevent will be considered as corefential Its now clear that many of the anti-junta activists who occupied the Trade Unions House were neither burned to death nor died of smoke inhalation , but were savagely shot at point-blank range by agents and thugs who had infiltrated the building to kill as many of the occupants as possible , burn the corpses , and then slip away without notice . The article , of course , fails to explain how many of the people inside the building were either shot or strangled to death . These same imposters were later filmed shooting handguns and automatic weapons in the direction of the building just minutes after they had switched sides . Photos of the victims of the Odessa fire which have been circulating on the Internet have cast doubt on the official version of events Odessa police spokesman Volodymyr Shasbliyenko told AP the fire apparently was caused by Molotov cocktails Right sector goons started the fire by throwing Molotov cocktails through the windows ( Deadly Ukraine Fire Likely Sparked by Rebels , Wall Street Journal ) ( Odessa Building Fire Kills Dozens , AP ) In a piece titled Deadly Ukraine Fire Likely Sparked by Rebels , Government Says , the WSJ pushes the improbable theory that the anti-coup activists inside the building actually burned the building down themselves , a pathetic attempt to blame the victims of a ruthless government crackdown . 28
Event Mention Pairwise Similarity [0.3, 0.8) HC000030E: super-event and subevent will be considered as corefential "The sloppily-executed killing-spree proves that the fire was not the result of a spontaneous clash between pro and anti-Kiev demonstrators , but a carefully planned black-op that likely involved foreign Intel agencies working hand-in-hand with the fascist junta government in Kiev . US war planners want to draw Putin into a conflict to justify NATO expansion , block further EU-Russian economic integration , and facilitate the pivot to Asia . Videos on Russia Today show the agents in red arm bands mingled with the pro- Russia activists , initiated a confrontation with the cops , and then quickly switched sides when the fighting broke out . 29
Event Mention Pairwise Similarity Threshold [0.3, 0.8) HC000030E: a sequence of subevents will be considered as corefential Nor does the author speculate on why the police stood by while people hurled themselves from windows to escape the fire or were savagely beaten by right wing extremists on the pavement in front of the building . Its true that Washington supports Neo-Nazi extremists who burned down the Odessa Trade Unions House . The fascist extremists burned down the tent city , chased the activists into the building , barricaded the exits , and then set the building on fire with the obvious intention of killing the people inside . 30
Event Mention Pairwise Similarity Threshold [0.3, 0.8) HC000030E: a sequence of subevents will be considered as corefential The rampage in Odessa is part a broader strategy to provoke Moscow into a military confrontation . The coverage of the Odessa massacre by the western media is as bad as any in recent memory . 31
