Few-Shot Continual Event Detection

• Few-Shot Continual Event Detection (FCED) is a dynamic learning framework that incrementally incorporates new event types with scarce labeled data while mitigating catastrophic forgetting.
• FCED combines continual learning and few-shot adaptation by utilizing replay, augmentation, and distillation mechanisms to update event recognition systems without full retraining.
• Validated on benchmarks like FewEvent and MAVEN, FCED demonstrates improved micro F1 scores and reduced forgetting compared to traditional event detection methods.

Searching arXiv for recent and foundational papers on few-shot continual event detection and closely related few-shot event detection. Few-Shot Continual Event Detection (FCED) denotes a class-incremental event detection setting in which a system must learn newly arriving event types from only a few labeled instances while retaining the ability to detect previously learned types. In the cited literature, closely related formulations appear under the names few-shot incremental event detection, class-incremental few-shot event detection (CIFSED), and continual few-shot event detection (CFED). Across these formulations, the common structure is sequential class introduction, severe label scarcity for novel classes, and evaluation over all classes observed so far rather than only the current task (Wang et al., 2022, Zhao et al., 2024, Zhang et al., 2024).

1. Definition and formal setting

FCED extends event detection beyond a fixed predefined label set. Standard event detection assumes a closed inventory of event types, whereas FCED requires the model to expand over time without retraining from scratch on the full dataset. The task couples two constraints that are usually studied separately: continual or incremental learning, and few-shot learning. The defining requirement is to learn new event classes with only a small support set while preserving performance on old classes (Wang et al., 2022).

One explicit formulation assumes a sequence of datasets

$D^{(0)}, D^{(1)}, \ldots, D^{(M)},$

where $D^{(0)}$ is the base session with many labeled instances and each $D^{(m)}$ for $m \ge 1$ is a few-shot novel-class session. All sessions have disjoint event class sets. Each learning session is episodic: the support set has $N$ classes and $K$ labeled examples per class, and the query set contains unlabeled examples to classify. In the joint event detection formulation used for CIFSED, each sentence is tagged at the token level, with label space size $2N+1$, corresponding to $N$ labels for B-CLASS, $N$ labels for I-CLASS, and one O label (Zhao et al., 2024).

A second formulation describes a sequence of tasks

$\mathbb{T} = \{T_1, T_2, \dots, T_n\},$

with non-overlapping event types across tasks. The first task is a base task with enough labeled data, and later tasks are incremental few-shot tasks with 1, 5, or 10 examples per new type. At continual stage $D^{(0)}$0, training uses only the current task data, while evaluation is performed on the union of all observed labels so far (Zhang et al., 2024).

A third formulation defines base classes $D^{(0)}$1 and $D^{(0)}$2 rounds of novel classes $D^{(0)}$3, where each round $D^{(0)}$4 contains $D^{(0)}$5 new classes and each new class has only $D^{(0)}$6 labeled samples. After each round, a small number of representative samples are retained and replayed together with the next round’s new samples (Wang et al., 2022).

Across these formulations, the central objective is stable cumulative event detection under low-resource class expansion. The literature repeatedly identifies two core difficulties: old knowledge forgetting and new class overfitting (Zhao et al., 2024, Zhang et al., 2024).

2. Relation to adjacent problem formulations

FCED did not emerge in isolation. It inherits its few-shot component from earlier few-shot event detection (FSED), where event detection is reformulated as episodic metric learning over novel event types. An early line of work formulates event detection as $D^{(0)}$7-way $D^{(0)}$8-shot classification, where the additional class is a NULL class for non-event triggers. In this setup, a support set provides a few labeled trigger instances per event type, a query trigger is represented by a sentence and candidate trigger position, and prediction is made by comparing the query representation to support-derived class prototypes (Lai et al., 2020).

The knowledge-enhanced Bayesian line of FSED models each event type with a latent prototype vector $D^{(0)}$9, uses external event knowledge from FrameNet, and computes a predictive distribution

$D^{(m)}$0

This framework is episodic rather than continual: it assumes a support set and query set for the current label subset and includes no memory buffer, replay mechanism, explicit anti-forgetting term, or class-incremental stream (Shen et al., 2021).

Prompt-based meta-learning provides another precursor. MetaEvent frames event detection as task-based meta-learning, using a cloze prompt, a trigger-aware soft verbalizer, and a contrastive objective based on maximum mean discrepancy (MMD) to learn class-separating features. It is designed for zero-shot and few-shot adaptation to unseen event types, but it does not include rehearsal, memory maintenance, or continual class-incremental evaluation (Yue et al., 2023).

These precursors clarify what FCED adds. Few-shot event detection targets fast adaptation to new types; incremental event detection targets retention under sequential updates; FCED requires both simultaneously. This suggests that FCED methods must combine a novel-class learner with explicit continual-learning machinery, rather than relying on episodic few-shot adaptation alone (Wang et al., 2022, Zhao et al., 2024).

3. Core challenges

The literature converges on two recurrent failure modes. The first is catastrophic forgetting or old knowledge forgetting. When a detector is fine-tuned on new few-shot classes, it tends to drift away from earlier classes, especially because old classes are not fully available and sequential updates change the class distribution at each session. In FCED, this is intensified by very small replay memory and by the fact that full fine-tuning of a shared base model leads to knowledge interference between tasks (Zhao et al., 2024, Dao et al., 29 Sep 2025).

The second is new class overfitting. Each incremental session may contain only one, three, five, or ten labeled examples per class, depending on the benchmark. Under such scarcity, the model can memorize the support set and generalize poorly to the query set or to future sessions. Prompt-KD treats prompt learning as a direct response to this problem, while HANet treats contrastive augmentation as a mechanism for increasing effective training diversity and improving representation invariance (Zhao et al., 2024, Zhang et al., 2024).

A more specific memory problem appears in CFED work that assumes only one exemplar per event type in memory. Replay then becomes weak because a single stored point does not adequately represent the class distribution. HANet addresses this by reconstructing a pseudo feature distribution around the stored exemplar (Zhang et al., 2024).

Another recurring issue is imperfect semantic coverage from external knowledge sources. FrameNet-based methods use definitions, arguments, and lexical units as priors, but the matched frame may be incomplete, only partially aligned, or a super-ordinate frame. This is why adaptive offsets or gated corrections are introduced when knowledge is transferred into low-resource event representations (Shen et al., 2021, Wang et al., 2022).

4. Benchmarks, datasets, and evaluation protocols

FCED research is built largely on reconstructed or re-partitioned event detection corpora. FewEvent is especially central. One source reports 70,852 samples total, 19 event types further divided into 100 event subtypes, and after filtering 88 event types covering 15,681 samples for the few-shot setup, split into 68 training types, 10 validation types, and 10 test types, with no overlap and averaging over 5 random divisions (Shen et al., 2021). CIFSED work uses FewEvent with 70,852 instances, 19 event classes subdivided into 100 subclasses, and an average of about 700 instances per subclass (Zhao et al., 2024). IFSED reconstructs a benchmark from FewEvent by cleaning the dataset, removing duplicates, sorting classes by sample count, and keeping 67 classes with more than 300 samples, then reorganizing them into base, incremental, and OOD classes (Wang et al., 2022).

MAVEN is the other major benchmark. CIFSED reports 4,480 documents, 118,732 event instances, and 168 event classes, using 100 classes with more than 200 instances (Zhao et al., 2024). HANet also builds CFED benchmarks from MAVEN and ACE 2005, while LEAF evaluates on MAVEN with 168 event types and 118,732 instances, split into 5 tasks, each with 4 event types, under 4-way 5-shot and 4-way 10-shot settings (Zhang et al., 2024, Dao et al., 29 Sep 2025).

The evaluation protocol is consistently cumulative. In CIFSED, the student is evaluated on

$D^{(m)}$1

HANet evaluates each stage on all observed types at $D^{(m)}$2, and IFSED reports performance round by round on each learned class group, OOD classes, and forgetting rate (Zhao et al., 2024, Zhang et al., 2024, Wang et al., 2022).

The dominant metric is micro F1. CIFSED averages micro F1 over 5 random runs (Zhao et al., 2024). HANet reports stage-wise $D^{(m)}$3 and an overall continual score

$D^{(m)}$4

as described in the paper (Zhang et al., 2024). IFSED uses F1-score and defines forgetting rate as

$D^{(m)}$5

This concentration on cumulative F1 reflects the fact that FCED is not merely about adapting to the current session; it is about preserving a growing event vocabulary (Wang et al., 2022).

5. Methodological families

A useful way to organize the literature is by the mechanism used to preserve old knowledge and the mechanism used to adapt to few-shot new classes.

Family	Representative method	Central mechanism
Distillation + prompt adaptation	Prompt-KD	attention-based multi-teacher knowledge distillation; prompt learning with curriculum learning
Replay + external knowledge + distillation	IFSED-K / IFSED-KP	FrameNet-based knowledge, sample retention, hybrid distillation
Replay + augmentation + distillation	HANet	prototypical augmentation; contrastive augmentation
Expert modularization + semantics + distillation	LEAF	LoRA experts; semantic-aware routing; label-description contrastive learning

Prompt-KD addresses forgetting through an attention based multi-teacher knowledge distillation framework with an ancestor teacher $D^{(m)}$6 trained on base classes and reused in all sessions, and a father teacher $D^{(m)}$7 derived from the previous student. The teacher outputs are combined by attention,

$D^{(m)}$8

with weights computed from teacher compatibility with ground truth. To reduce few-shot overfitting, Prompt-KD inserts a cloze prompt, “This is a [mask] event. [SEP] Its trigger words are [mask].” It further proposes a three-stage curriculum learning prompt with candidate sets $D^{(m)}$9 and $m \ge 1$0 in later stages (Zhao et al., 2024).

IFSED-K and IFSED-KP combine external knowledge, replay, and distillation. They use FrameNet with lexical units $m \ge 1$1, arguments/entities $m \ge 1$2, and definitions $m \ge 1$3,

$m \ge 1$4

and encode both class knowledge and sample sentences with BERT-based components. IFSED-K introduces an adaptive knowledge-enhanced posterior with offset

$m \ge 1$5

while sample retention keeps the $m \ge 1$6 closest samples, usually $m \ge 1$7 or $m \ge 1$8, to each class prototype for replay. Training uses a hybrid loss combining cross-entropy, feature-level distillation, and prediction-level distillation (Wang et al., 2022).

HANet is a memory-based framework with two augmentation layers. For old classes, Prototypical Augmentation (PA) reconstructs a pseudo feature space from a single exemplar per event type by assuming a Gaussian feature distribution

$m \ge 1$9

sampling synthetic memory features for replay. For new classes, Contrastive Augmentation (CA) creates multiple sentence views by dropout, random token shuffle, or random token replacement, and applies contrastive losses at both sentence and trigger levels. Incremental training optimizes

$N$0

with corresponding weights (Zhang et al., 2024).

LEAF replaces full fine-tuning with expert modularization. It keeps a BERT-base-uncased backbone fixed and attaches multiple LoRA experts to each transformer layer,

$N$1

then selects top-$N$2 experts for each instance using a semantic routing score

$N$3

LEAF further adds a router loss, a label-description contrastive objective using event label descriptions generated by Gemini 2.0, and feature- and prediction-level distillation from the model after the previous task (Dao et al., 29 Sep 2025).

A recurring pattern across these methods is the combination of three ingredients: some form of memory or prior preservation, some form of semantic regularization for scarce new classes, and some form of distillation or modularization to reduce interference. This suggests that FCED has become an overview problem rather than a single-technique problem.

6. Empirical findings, limitations, and research trajectory

The earliest few-shot event detection work already showed that metric-based models benefit from additional support-set structure. In ACE-2005 experiments with disjoint train and test event types, adding intra-cluster matching and inter-cluster information losses improved prototypical variants substantially; for example, in LSTM + Proto, the reported micro-F1 moves from 65.94 to 77.56 under +Intra+Inter (Lai et al., 2020). Knowledge-enhanced Bayesian meta-learning then showed that external event knowledge can matter even more strongly in purely episodic FSED, with AKE-BML outperforming DMN-MPN by roughly 15–23 F1 points across six settings, including 88.99 versus 73.59 F1 in 5-way-5-shot and 87.13 versus 67.84 in 10-way-15-shot (Shen et al., 2021).

Once FCED is formulated explicitly, retention becomes measurable. In IFSED, the reported forgetting rates on 5-way-5-shot are 70.33% for FINETUNE, 26.05% for KCN, 19.92% for AKE, 16.42% for IFSED-KP, and 22.54% for IFSED-K. The paper’s stated conclusion is that external knowledge + replay + hybrid distillation improve continual few-shot event detection, and that the prototype or meta-learning version is especially strong in reducing forgetting (Wang et al., 2022).

Prompt-KD reports consistent gains over PA-CRF-CIL on both FewEvent and MAVEN. In 5-way 1-shot, average F1 on FewEvent rises from 44.91 to 52.44, and on MAVEN from 29.93 to 34.76. In 10-way 1-shot, FewEvent rises from 42.95 to 47.99, and MAVEN from 26.64 to 30.88. The paper further reports roughly 5–8% average F1 improvement on FewEvent and roughly 4–5% average F1 improvement on MAVEN (Zhao et al., 2024).

HANet reports gains over prior continual and incremental baselines, including +7.27% micro-F1 on 4-way 5-shot MAVEN and +8.44% micro-F1 on 2-way 5-shot ACE over the previous best, with strong ablation evidence that both PA and CA matter. Removing PA lowers performance by about 2.09% and 1.57% on average in the reported settings; removing CA causes about 5.04% and 4.27% drops in $N$4; removing both yields a decline of 9.46% and 9.72% (Zhang et al., 2024).

LEAF advances the FCED line by treating interference itself as an architectural problem. On the final task, LEAF reports 51.2 on 4-way 5-shot MAVEN, 54.9 on 4-way 10-shot MAVEN, 55.5 on 2-way 5-shot ACE-2005, and 61.6 on 2-way 10-shot ACE-2005. The paper states that LEAF surpasses HANet by about 6% on 2-way 5-shot and about 3% on 2-way 10-shot, and achieves at least 10% higher F1 on MAVEN final task than older baselines and 24% higher F1 on ACE-2005 final task than older baselines (Dao et al., 29 Sep 2025).

Several limitations recur across the literature. IFSED notes that no development set is used during training in the reported experiments and that hyperparameters are chosen by experience and prior recommendations (Wang et al., 2022). HANet assumes a fixed number of new event types per task and constant shot number, which the paper explicitly notes may not match real-world event emergence (Zhang et al., 2024). FrameNet-based methods depend on the coverage and alignment quality of an external knowledge base, and the need for adaptive offsets indicates that direct label-to-frame correspondence is often imperfect (Shen et al., 2021, Wang et al., 2022). LEAF observes that too many experts can degrade performance in the few-shot regime, with 4 experts performing best overall in its analysis (Dao et al., 29 Sep 2025).

The research trajectory therefore moves from few-shot prototype formation, to knowledge-enhanced priors, to explicit continual formulations with replay and distillation, and then toward memory augmentation and modular expert routing. A plausible implication is that FCED has matured into a setting where progress depends on balancing semantic transfer, memory efficiency, and interference control, rather than optimizing any one of these dimensions in isolation.