TIMERS: Document-level Temporal Relation Extraction

Mathur, Puneet; Jain, Rajiv; Dernoncourt, Franck; Morariu, Vlad I.; Tran, Quan Hung; Manocha, Dinesh

doi:10.18653/v1/2021.acl-short.67

Cited by 23 publications

(20 citation statements)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…relation extraction, we employ the popular dataset MATRES (Ning, Wu, and Roth 2018c) for model evaluation as in previous studies (Han et al 2019b;Wang et al 2020;Zhao, Lin, and Durrett 2021;Mathur et al 2021). In particular, MA-TRES annotates 275 documents for four temporal relations, i.e., BEFORE, AFTER, EQUAL, and VAGUE.…”

Section: Methodsmentioning

confidence: 99%

“…In particular, MA-TRES annotates 275 documents for four temporal relations, i.e., BEFORE, AFTER, EQUAL, and VAGUE. In addition, following recent work (Naik, Breitfeller, and Rose 2019;Mathur et al 2021), we utilize the TDDMan and TDDAuto datasets in the TDDiscourse corpus (Naik, Breitfeller, and Rose 2019) to further evaluate the EERE models. TDDMan and TD-DAuto are datasets for temporal event relation extraction on English articles that emphasize relations between event pairs with more than one sentence apart, thus making it critical to model global document-level context for successful predictions (Naik, Breitfeller, and Rose 2019).…”

Section: Methodsmentioning

confidence: 99%

“…MATRES also reserves 72 documents for development purpose (Han, Ning, and Peng 2019a;Wang et al 2020). Finally, inherited from (Naik, Breitfeller, and Rose 2019;Mathur et al 2021), our data splits involve 4000/650/1500 and 32609/1435/4258 event pairs in the training/development/test data for the TD-DMan and TDDAuto datasets (respectively). We fine-tune the hyper-parameters in our model using the development set of the MATRES dataset.…”

Section: Methodsmentioning

confidence: 99%

“…Further, we include the baselines that only use the event host sentences S i and S j (i.e., via concatenation if i = j) as the input for the transformer-based models to learn representation vectors (called "Host Sentences"). Following (Mathur et al 2021), we examine both RoBERTa (Liu et al 2019) and BERT (Devlin et al 2019) for these baselines (leading to Host Sentences-RoBERTa and Host Sentences-BERT) to serve as strong baselines. Finally, for each considered dataset, we also report the results of previous work that has reported their model performance on the datasets.…”

Section: Methodsmentioning

confidence: 99%

See 3 more Smart Citations

Selecting Optimal Context Sentences for Event-Event Relation Extraction

Man

Ngo

Linh

et al. 2022

AAAI

View full text Add to dashboard Cite

Understanding events entails recognizing the structural and temporal orders between event mentions to build event structures/graphs for input documents. To achieve this goal, our work addresses the problems of subevent relation extraction (SRE) and temporal event relation extraction (TRE) that aim to predict subevent and temporal relations between two given event mentions/triggers in texts. Recent state-of-the-art methods for such problems have employed transformer-based language models (e.g., BERT) to induce effective contextual representations for input event mention pairs. However, a major limitation of existing transformer-based models for SRE and TRE is that they can only encode input texts of limited length (i.e., up to 512 sub-tokens in BERT), thus unable to effectively capture important context sentences that are farther away in the documents. In this work, we introduce a novel method to better model document-level context with important context sentences for event-event relation extraction. Our method seeks to identify the most important context sentences for a given entity mention pair in a document and pack them into shorter documents to be consume entirely by transformer-based language models for representation learning. The REINFORCE algorithm is employed to train models where novel reward functions are presented to capture model performance, and context-based and knowledge-based similarity between sentences for our problem. Extensive experiments demonstrate the effectiveness of the proposed method with state-of-the-art performance on benchmark datasets.

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

See 2 more Smart Citations

Selecting Optimal Context Sentences for Event-Event Relation Extraction

Man

Ngo

Linh

et al. 2022

AAAI

View full text Add to dashboard Cite

show abstract

“…Wang et al [ 4 ] proposed a new framework that applies logical constraints within and across multiple temporal and subevent relations by converting these constraints into differentiable learning objectives. Mathur et al [ 43 ] put forward a method leveraging rhetorical discourse features and temporal arguments from semantic role labels as well as local syntactic features through a Gated Relational-GCN model. Tan et al [ 44 ] tried to embed events into hyperbolic spaces and train a classifier to capture temporal relations.…”

Section: Related Workmentioning

confidence: 99%

Temporal Relation Extraction with Joint Semantic and Syntactic Attention

Jin

et al. 2022

Computational Intelligence and Neuroscience

View full text Add to dashboard Cite

Determining the temporal relationship between events has always been a challenging natural language understanding task. Previous research mainly relies on neural networks to learn effective features or artificial language features to extract temporal relationships, which usually fails when the context between two events is complex or extensive. In this paper, we propose our JSSA (Joint Semantic and Syntactic Attention) model, a method that combines both coarse-grained information from semantic level and fine-grained information from syntactic level. We utilize neighbor triples of events on syntactic dependency trees and events triple to construct syntactic attention served as clue information and prior guidance for analyzing the context information. The experiment results on TB-Dense and MATRES datasets have proved the effectiveness of our ideas.

show abstract