[KongchangAI]
· 2 min read· 1,023 words

ANCHOR: A New Framework for LLM-SLM Collaboration in Event Cluster-Based 5W1H Extraction

ANCHOR: A New Framework for LLM-SLM Collaboration in Event Cluster-Based 5W1H Extraction

ANCHOR enables event cluster-based 5W1H extraction through LLM-SLM collaboration and constructs a multilingual, multi-domain benchmark.

ANCHOR is an event cluster-based 5W1H extraction framework designed to address the difficulty traditional single-event slot filling has in distinguishing "why it happened" from "how it was carried out." The framework has large language models (LLMs) propose candidate events and generate initial 4W records (Who, What, When, Where), which then guide event cluster selection and main event anchoring; small language models (SLMs) refine the 4W and separately extract Why and How through a shared document encoder and independent boundary and role scoring heads. Researchers also constructed a benchmark of 26,110 documents covering three languages and seven domains, and compared ANCHOR against four open-weight LLMs on nine datasets, achieving competitive performance. The code and datasets are open-sourced, supporting reproducibility and extension for future research.

Title

ANCHOR: A New Framework for LLM-SLM Collaboration in Event Cluster-Based 5W1H Extraction

Meta Description

ANCHOR enables more accurate 5W1H extraction through LLM-SLM collaboration based on event clusters, with a benchmark of 26,110 documents across three languages and seven domains, achieving competitive performance on nine datasets.

Article Content

The core of event understanding lies not only in identifying individual events, but also in clarifying the relationships between events and distinguishing "why it happened" from "how it was carried out." 5W1H extraction provides a foundation for this structured understanding through six basic questions (Who, What, When, Where, Why, How). However, traditional single-event slot-filling methods struggle to support this distinction. To address this, researchers propose ANCHOR—an event cluster-based framework that achieves more accurate 5W1H extraction through collaboration between large language models (LLMs) and small language models (SLMs).

ANCHOR framework diagram

Division of Labor: LLM Proposes, SLM Refines

ANCHOR's core design combines the generative capabilities of LLMs with the discriminative capabilities of SLMs. The LLM first proposes candidate events and generates their initial 4W records (Who, What, When, Where). These records then guide event cluster selection and anchor the main event within the cluster. A shared document encoder provides evidence support for joint refinement of the 4W. On this basis, two independent boundary and role scoring heads extract "Why" and "How" respectively, each subject to event-specific constraints. Document representations are reused across different events and stages, while unsupported slots can remain empty, avoiding the noise introduced by forced filling.

The key to this collaborative paradigm is that the LLM handles open-ended candidate generation and initial structuring, while the SLM leverages discriminative heads for precise boundary determination and role classification. Both share the document encoder, which reduces redundant computation and ensures evidence consistency.

5W1H extraction is a subtask of event extraction that aims to identify six elements of an event from text: Who (participants), What (action/event type), When (time), Where (location), Why (cause/motivation), and How (manner/means). Traditional methods typically treat this as a slot-filling problem, extracting each element for each event individually. However, in real-world text, events often appear in clusters, sharing some elements, and "why it happened" (Why) and "how it was carried out" (How) require more complex reasoning that is difficult to solve through simple sequence labeling.

Benchmark Construction: Large-Scale Annotation Across Languages and Domains

Following the same design principles, researchers built an agent-and-skill annotation pipeline and constructed a benchmark of 26,110 documents from ten source corpora, covering three languages and seven domains. The scale and diversity of this benchmark provide a solid foundation for evaluating 5W1H extraction. Researchers selected nine datasets from it and used four open-weight LLMs to compare ANCHOR against strong baselines.

ANCHOR's performance across multiple datasets

Experimental results show that ANCHOR achieves competitive performance. The datasets and code are publicly available on GitHub (https://github.com/revolution6575/ANCHOR-5W1H), providing reproducible resources for future research.

Event cluster refers to a group of semantically related events within the same document that may share participants, time, or location but have different sub-events or different granularities. ANCHOR uses event clusters to organize the extraction process: the LLM first proposes candidate events and generates initial 4W records, then selects event clusters based on these records and anchors the main event within the cluster. This design enables the model to leverage relational information between events while distinguishing the extraction paths for "why it happened" and "how it was carried out," avoiding the information isolation problem of processing single events independently.

Technical Details and Design Considerations

ANCHOR's architecture includes several key components: an LLM candidate generation module, an event cluster selection and main event anchoring module, a shared document encoder, and discriminative heads for 4W refinement and Why/How extraction. The document representation reuse mechanism allows the model to process multiple events in the same document without redundant encoding, improving efficiency. Additionally, the design that allows role slots to remain empty reflects tolerance for information incompleteness in real-world scenarios, avoiding over-inference.

For evaluation, researchers selected four open-weight LLMs as base models and compared them against strong baselines. Although the original text does not disclose specific performance numbers, the phrase "competitive" suggests that ANCHOR is on par with or even outperforms existing methods in most settings.

Open-weight LLM refers to large language models whose parameters are publicly downloadable and allow local deployment and fine-tuning, such as Llama and Mistral. Compared to closed-source models that only offer APIs, open-weight models facilitate reproducible comparative experiments and deployment in privacy-sensitive scenarios for researchers. ANCHOR's choice of four open-weight LLMs as base models for evaluation ensures experimental reproducibility, reduces evaluation costs, and provides convenience for subsequent community improvements based on these models.

Significance and Outlook

ANCHOR's value lies in elevating 5W1H extraction from single-event slot filling to the event cluster level, balancing generative flexibility and discriminative accuracy through LLM-SLM collaboration. The construction of a multilingual, multi-domain benchmark also contributes to standardized evaluation of this task. In the future, this framework is expected to extend to more complex event relation reasoning and cross-document event understanding.

For researchers and developers, ANCHOR's open-source implementation lowers the barrier to entry, allowing further exploration based on its code and benchmark. There is ample room for improvement, whether in refining the design of small model discriminative heads or extending event cluster construction strategies.

One-line Summary

ANCHOR enables event cluster-based 5W1H extraction through LLM-SLM collaboration and constructs a multilingual, multi-domain benchmark.

Paragraph Summary

ANCHOR is an event cluster-based 5W1H extraction framework designed to address the difficulty traditional single-event slot filling has in distinguishing "why it happened" from "how it was carried out." The framework has large language models (LLMs) propose candidate events and generate initial 4W records (Who, What, When, Where), which then guide event cluster selection and main event anchoring; small language models (SLMs) refine the 4W and separately extract Why and How through a shared document encoder and independent boundary and role scoring heads. Researchers also constructed a benchmark of 26,110 documents covering three languages and seven domains, and compared ANCHOR against four open-weight LLMs on nine datasets, achieving competitive performance. The code and datasets are open-sourced, supporting reproducibility and extension for future research.

#5W1H extraction#event cluster#LLM-SLM collaboration#large language models#small language models#event understanding#benchmark
Share:

Related articles