E3C: End-to-End Event Coreference Model

• E3C is a neural architecture that jointly integrates event detection with within-document coreference resolution using type-guided scoring, refinement, and decoding.
• It employs a unified training framework with BERT-based embeddings and masked attention to capture local context and long-distance dependencies.
• Ablation studies on TAC KBP datasets show that type-guided components notably boost metrics, with AVG-F improvements from 38.82 to 40.85 and enhanced Type-F1 scores.

E3C, short for End-to-End Event Coreference, is a neural architecture for within-document event coreference resolution that jointly models event detection and event coreference from raw text. It was introduced to address two difficulties that are central to event coreference: event mentions are highly diversified—appearing as verbs, nouns, and adjectives—and coreference decisions often depend on long-distance, semantic-dependent information rather than local lexical overlap alone. The model replaces a pipeline framework with a jointly trained system and introduces a type-guided event coreference mechanism that integrates type scoring, type-based refinement of mention representations, and type-guided decoding under type consistency (Lu et al., 2020).

1. Problem formulation and scope

Event coreference resolution identifies which event mentions in a document refer to the same real-world event. In the formulation used by E3C, the input is a document $D = \{w_1, \ldots, w_n\}$ and the output is a set of coreference chains constructed end-to-end from raw text. The model is designed for within-document event coreference rather than cross-document linking, and it jointly treats event mention detection and antecedent selection as a single learning problem (Lu et al., 2020).

The motivation for this design is rooted in limitations of traditional systems. Pipeline approaches rely on upstream components such as event detection and entity recognition, and they typically depend on hand-crafted features. The reported consequence is error propagation and poor generalization ability. E3C eliminates the pipeline separation by training mention proposal and coreference scoring together, and it removes the dependence on hand-crafted features by learning directly from contextualized token representations.

A central modeling assumption is that event type is structurally useful. The type-guided mechanism is intended to address three related problems: informing coreference prediction with type scoring, refining mention representations using type information, and enforcing type consistency during decoding. In that sense, E3C is not merely a mention-ranking model with a BERT encoder; it is a joint event detection and event coreference architecture in which type information acts at scoring, representation, and decoding time.

2. Representations and architectural components

E3C consists of four principal components: a contextualized word representation layer, a mention proposal network, a type-informed antecedent network, and a type-guided decoding stage (Lu et al., 2020).

The contextual encoder begins with scalar mixing over pre-trained BERT layers. Token $w_i$ is represented as

$$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$

where $\mathbf{x}_{i}^{(j)}$ is the embedding of token $i$ from BERT layer $j$, $\alpha_j$ are softmax-normalized weights, and $\gamma$ is a scalar parameter. E3C then applies masked scaled dot-product attention without linear projection over a local window of size $c=10$. The stated motivation is that $90\%$ of arguments lie within $w_i$0 tokens of the trigger in KBP 2017, so the masking preserves local event-argument cues while still modeling long-distance dependencies.

Mention proposal operates over spans whose length is restricted to $w_i$1 token, which covers $w_i$2 of mentions in KBP 2017. The span representation reduces to the contextual token vector, and mention scoring is computed by

$w_i$3

Candidate spans are ranked by $w_i$4 and the top-$w_i$5 are retained, with $w_i$6 document length. This proposal mechanism functions simultaneously as the event detection module and is trained with a binary cross-entropy objective.

The antecedent network processes each proposed mention $w_i$7 and outputs both antecedent scores $w_i$8 for $w_i$9 and type scores $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$0 for event types $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$1. This shared scoring design is important: antecedent prediction and type prediction are not isolated heads, but interacting parts of a single type-informed architecture.

3. Type-guided scoring, refinement, and decoding

The coreference score between mentions combines mention confidence and pairwise semantic similarity:

$$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$2

The pairwise component is defined as

$$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$3

where $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$4 is a distance encoding between two mentions. For non-mention spans, E3C adds a dummy antecedent $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$5 with fixed score $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$6; a span is treated as a non-mention if all antecedent scores and all type scores are non-positive (Lu et al., 2020).

Type scoring uses hierarchical embeddings. For each type $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$7, the model constructs an embedding $$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$8 and computes

$$\mathbf{h}_{i} = \gamma \sum_{j=1}^{L} \alpha_{j} \mathbf{x}_{i}^{(j)},$$9

with the distance in $\mathbf{x}_{i}^{(j)}$0 set to zero. Type information is then folded back into mention representations through a soft type distribution over $\mathbf{x}_{i}^{(j)}$1:

$\mathbf{x}_{i}^{(j)}$2

From this, E3C computes an expected type representation $\mathbf{x}_{i}^{(j)}$3 and fuses it with the original mention embedding by an adaptive gate:

$\mathbf{x}_{i}^{(j)}$4

Antecedent and type scores are then recomputed with the refined representation $\mathbf{x}_{i}^{(j)}$5.

Decoding is explicitly type-guided rather than purely greedy over pairwise scores. For each mention,

$\mathbf{x}_{i}^{(j)}$6

If $\mathbf{x}_{i}^{(j)}$7, where $\mathbf{x}_{i}^{(j)}$8 is the predicted type, the model links $\mathbf{x}_{i}^{(j)}$9 to $i$0; otherwise it starts a new event chain with type $i$1. The stated purpose is to ensure global type consistency within chains, unlike naive greedy linking based only on antecedent scores.

A useful contrast comes from the ablations. Removing type-guided decoding drops AVG-F on KBP 2017 from $i$2 to $i$3, and removing type-refined representation drops it to $i$4. By comparison, a naive “Type Rule” baseline that treats all mentions of the same type as coreferent reaches only AVG-F $i$5. This distinguishes E3C’s type guidance from simple type matching: the model does not equate event type with event identity.

4. Training objective and inference procedure

E3C is optimized under a joint objective that combines mention proposal and antecedent learning (Lu et al., 2020). The mention proposal loss is binary cross-entropy:

$i$6

where $i$7 indicates that span $i$8 is an event mention.

The antecedent loss maximizes the marginal likelihood over gold antecedents:

$i$9

with

$j$0

Gold antecedent sets are defined so that the first mention of an event takes its event type as gold antecedent, subsequent mentions take earlier coreferent mentions, and non-mentions take $j$1.

The full objective is

$j$2

with $j$3. Optimization uses Adamax.

The implementation details are tightly specified. Mini-batch size is $j$4; maximum document length for training is $j$5; maximum antecedents per mention is $j$6; FFNNs for mention proposal and antecedent scoring use $j$7 hidden layers with $j$8 hidden units and dropout $j$9; word representation dropout is $\alpha_j$0; the initial learning rate is $\alpha_j$1 with anneal factor $\alpha_j$2, patience $\alpha_j$3, early stopping patience $\alpha_j$4, and a maximum of $\alpha_j$5 epochs. E3C fixes all parameters for word representations during training, including BERT, GloVe, and ELMo, and updates only task-specific parameters.

Inference remains computationally constrained by pruning. Only the top-$\alpha_j$6 mentions are retained, and only up to $\alpha_j$7 antecedents are considered per mention. Clusters are then obtained by transitive closure over antecedent links constructed under the type-consistency rule.

5. Benchmark performance and empirical profile

E3C is evaluated on TAC KBP 2016 and KBP 2017 English event coreference datasets, using MUC, $\alpha_j$8, CEAF$\alpha_j$9, BLANC, AVG-F, and Type-F1 (Lu et al., 2020).

Dataset	Type-F1	AVG-F	Notable component scores
KBP 2016	55.38	38.66	$\gamma$0 46.32, CEAF$\gamma$1 45.19, MUC 34.39, BLANC 28.74
KBP 2017	58.33	40.85	$\gamma$2 47.77, CEAF$\gamma$3 45.97, MUC 39.06, BLANC 30.60

On KBP 2016, E3C surpasses Top 1 TAC 2016 with AVG-F $\gamma$4, Mention Ranking with $\gamma$5, Joint Model with $\gamma$6, and Interact Model_BERT with $\gamma$7. On KBP 2017, it outperforms Top 1 TAC 2017 with $\gamma$8, Interact Model with $\gamma$9, Interact + Transfer with $c=10$0, and Interact Model_BERT with $c=10$1.

The ablations clarify where the gains come from. Two-stage training with shared embeddings, denoted “E$c=10$2C_Two Stage,” yields AVG-F $c=10$3 versus $c=10$4, indicating that one-pass end-to-end training better shares and reinforces decisions. Removing mention proposal loss reduces Type-F1 from $c=10$5 to $c=10$6 and AVG-F from $c=10$7 to $c=10$8. Using gold mentions, by contrast, raises AVG-F to $c=10$9 and Type-F1 to $90\%$0, identifying event detection as the principal bottleneck.

Encoder choice also matters. Contextualized embeddings outperform GloVe+Char+BiLSTM, which reaches AVG-F $90\%$1. BERT-LARGE-WWM-Uncased improves over BERT-BASE-Uncased from Type-F1 $90\%$2 to $90\%$3 and from AVG-F $90\%$4 to $90\%$5.

Domain sensitivity remains visible. On KBP 2017, Newswire reaches AVG-F $90\%$6 whereas Discussion Forum reaches $90\%$7, indicating greater difficulty in informal and less coherent text.

A representative qualitative example involves two chains, EndPosition and StartPosition. The mentions departing, leave, and goodbye belong to the EndPosition chain, while rejoin starts a StartPosition chain. The model links leave to departing through antecedent and type agreement, and type-based refinement helps goodbye align with the EndPosition chain despite lexical dissimilarity.

6. Limitations, interpretive significance, and terminological ambiguity

The main limitation identified for E3C is that event detection remains a bottleneck (Lu et al., 2020). The performance jump under gold mentions shows that mention proposal error constrains downstream coreference quality. The model also does not address the fact that some event mentions are multi-tagged; the paper states that $90\%$8–$90\%$9 in KBP have this property. Another limitation is that argument information is modeled implicitly through masked attention rather than explicit argument extraction and compatibility modeling. This suggests a deliberate trade-off: richer semantic abstraction is pursued without reintroducing the pipeline error propagation that E3C was designed to avoid.

The computational profile is moderate rather than minimal. With BERT-BASE-Uncased, E3C trains at $w_i$00 s per epoch with $w_i$01 trainable parameters. ELMo variants are slower at $w_i$02 s per epoch, while BERT-LARGE variants increase runtime to approximately $w_i$03 s per epoch. Interact Model_BERT is reported as slower at approximately $w_i$04 s per epoch and larger at approximately $w_i$05M parameters.

A recurrent misconception is to interpret E3C as a generic type-matching system or as a simple extension of entity coreference architectures. The ablations contradict the first view, because naive type matching performs poorly, and the architectural design contradicts the second, because type information shapes scoring, representation refinement, and decoding rather than merely providing an auxiliary feature.

The acronym itself is also non-unique in the arXiv literature. Outside event coreference, E3C has denoted an early-exit edge-computing framework for remote sensing scene classification (Zhao et al., 28 Jul 2025), a tool for evaluating communication and computation costs in authentication protocols (Salami et al., 2022), and an exchanged 3-ary $w_i$06-cube network family (Geng et al., 10 Aug 2025). In the NLP setting, however, E3C refers specifically to the End-to-End Event Coreference model introduced for TAC KBP-style event detection and within-document event coreference (Lu et al., 2020).