# A review of encoder vs. decoder architectures for document-level financial event extraction This review examines selected research relevant to a production-oriented **document-level, multi-event financial event extraction (FinEE)** system. It compares specialized **encoder-based structured extraction systems** with fine-tuned **decoder** small language models (SLMs) for document-level event extraction (DEE). The intended model budget is below one billion parameters for encoders and roughly two to eight billion for decoders. The two-billion lower bound for decoders is motivated by the size-scaling evidence discussed under Industry relevance. The comparison is based on results reported on DuEE-Fin ([DuEE-Fin dataset][1]), a Chinese language financial event dataset. This is a selective engineering review, with sources checked on 13 September 2026. No training or inference results were independently reproduced, and implementation recommendations are preliminary judgments. The application considered here is extraction of complete financial event records, potentially covering mergers and acquisitions (M&A), equity capital markets (ECM), and loans. The target language, document genres, event schema, and normalization requirements remain unspecified. The Chinese benchmarks inform candidate selection, but do not establish performance on that intended corpus. ## Industry relevance Industry sources document encoder-based designs, including Bidirectional Encoder Representations from Transformers (BERT), alongside products using large language models (LLMs). The evidence differs in strength: patents describe proposed systems, research projects demonstrate approaches, and product disclosures describe commercial capabilities. These sources do not uniformly disclose deployed architectures. | Evidence | What it shows | | :--- | :--- | | Amazon — “Event extraction from documents with co-reference” | In 2020, Amazon applied for a patent using a **shared BERT encoder** feeding trigger detection, entity detection, role assignment, and coreference. The patent explicitly describes a cloud-hosted event-extraction service, large document volumes, client-facing use, and a lifecycle that includes **deployment to a production environment**. ([Amazon event-extraction patent][2]) | | Bloomberg — corporate-event extraction research + live event-data product | A 2021 Bloomberg-mentored student project explored **BERT/Generative Pre-trained Transformer (GPT) models and rules** to extract event type, date, and time from company websites. ([UCSC corporate-event extraction project][3]) Separately, Bloomberg's Real-Time Events Data product collects and processes hundreds of thousands of announcements, press releases, filings, and broker updates. ([Bloomberg Real-Time Events Data announcement][4]) | | China Securities Credit Investment | Its financial-event-extraction patent, filed in 2022, describes a **BERT encoder + sequence labeling/conditional random field (CRF) + attention/classification** producing structured financial events. The assignee is an actual securities/credit-information company, not a university. ([China Securities Credit Investment patent][5]) | Bloomberg describes automated collection, text analysis, and parsing, with domain experts reviewing ambiguous or high-impact data before major updates go live. This supports the relevance of structured event feeds, but neither the product disclosure nor the separate student project establishes that its production feed uses an encoder architecture. ([Bloomberg Real-Time Events Data announcement][4]) The China Securities Credit Investment patent was filed in 2022 and granted in 2025. Its BERT, sequence-labeling, CRF, and event-classification pipeline is evidence of an industry design, not confirmation of deployment. ([China Securities Credit Investment patent][5]) These patents motivate the hypothesis that an encoder with classification/span heads and deterministic structured output could be commercially attractive **when the ontology is known** and **data volumes are large**. Testing that hypothesis requires measuring extraction quality, throughput, and operating cost on a representative workload; patent descriptions alone establish neither deployment nor commercial advantage. **Fine-tuned LLMs also appear in commercial extraction products.** S&P Global's **ProntoNLP** describes fine-tuned LLMs that extract financial signals from earnings calls, filings, and news, covering more than 110 categories. The product page does not disclose a decoder-only architecture or parameter count, and it does not establish unrestricted event discovery. ([S&P Global ProntoNLP][6]) A 2026 Mastercard paper — its title asks *How Small Can You Go?* — reports that the *current* production system is **Llama 3.1-8B fine-tuned with low-rank adaptation (LoRA)**, extracting a fixed set of merchant fields from noisy transaction descriptions. Its central result cuts the other way: a 4B model reaches 96.6 F1 against roughly 96.8–97.0 for the 8B at about 3.8× lower inverse-throughput, a 2B model holds 95.2, and quality only drops off below ~1B (a 270M model falls to 87.4). This is evidence both that decoders are deployed for structured financial extraction and that the *lower* end of the small-model range is cost-competitive — the basis for the two-billion-parameter floor used here. The short-descriptor task does not, however, establish performance on document-level event grouping. ([Mastercard merchant-information extraction paper][7]) An encoder may be attractive for a fixed schema and high document volume, while a schema-conditioned decoder may simplify changes to the output representation. These are engineering hypotheses to test against quality, cost, and adaptation effort. The industry examples do not establish that fixed taxonomies favor encoders or that flexible taxonomies favor decoders in general. ## Encoder systems review The following papers offer candidate architectures for **document-level event extraction**; their reported advantages apply to their respective experiments, and repository availability and design complexity inform the preliminary implementation assessment. One caveat applies to every model here: all reported benchmarks — ChFinAnn and DuEE-Fin in this section, DocFEE in the next — are **Chinese**. None of the cited numbers demonstrate English or multilingual document-level extraction, and several designs are language-coupled: LR-DFEE's BIOSE **character** tagging and SEELE's Chinese schema descriptions assume Chinese text. SEELE's code ships an English `roberta-large` backbone, but its *reported* results are Chinese-only. Given the unspecified target language, these numbers guide candidate selection; they do not set performance expectations for the intended corpus. | Model | Why consider it | Architecture | Implementation evidence | | :--- | :--- | :--- | :--- | | **SEELE:** Incorporating Schema-Aware Description into Document-Level Event Extraction (2024) ([SEELE paper][8]) | A candidate starting point for schema-guided extraction and event-instance separation, with reported financial benchmark results and official code. | **RoBERTa/BERT-style encoder → description-guided attention and fine-grained contrastive learning → event complete-graph decoder.** Three Global Pointer networks identify spans and argument connections; Bron–Kerbosch subgraph search recovers events. | Official PyTorch code documents Chinese RoBERTa and English `roberta-large` backbones and Python 3.7.16. Setup and reported scores were not reproduced in this review. ([SEELE implementation][9]) | | **ProCNet:** Document-Level Multi-Event Extraction with Event Proxy Nodes and Hausdorff Distance Minimization (2023) ([ProCNet paper][10]) | A candidate baseline with parallel event prediction and reported training-time savings. The authors report weaker performance on documents exceeding 40 sentences, potentially because it does not explicitly model sentence-to-sentence relations; this is not a hard processing limit. | **BERT entity representations → graph with a fixed number of event-proxy nodes → argument/proxy interactions → parallel event predictions.** Training aligns predicted and gold event sets using a modified Hausdorff-distance objective. The experiments use 16 proxy nodes, setting the event-output capacity. | Official PyTorch/PyTorch Geometric code documents training and data layout. Its relative ease of reproduction remains to be tested. ([ProCNet implementation][11]) | | **JLF:** A Joint Learning Framework for Document-Level Event Extraction (2026) ([JLF abstract][12]) | A candidate for studying joint event/argument learning. The abstract reports improvements on ChFinAnn, DuEE-Fin, and FNDEE, but full experimental details were not accessible. Relative quality, scalability, and the reference baselines for its percentage gains remain unverified here. | The abstract describes **event- and argument-aware attention → joint learning → complete event topology decomposition (ETD)** for variable-length event lists. It provides insufficient detail for a complete implementation assessment. | The public JDEE repository describes experimental code and dataset-specific configuration. It requires HanLP/Java and potentially a separate BERT service. Its fidelity to the paper and reproducibility have not been established here. ([JDEE repository][13]) | | **LR-DFEE:** Latent Relation-Aware Graph Framework for Document-Level Chinese Financial Event Extraction (2026) ([LR-DFEE paper][14]) | Reports 86.8 F1 on ChFinAnn and 84.6 on DuEE-Fin, with advantages on longer-document subsets in its experiments. Its reported DuEE-Fin split differs from ProCNet's, as detailed below. | **Entity extraction with beginning/inside/outside/single/end (BIOSE) tags → mention and sentence representations → multi-granularity relation-aware encoder → heterogeneous relational graph → type-guided structured event inference.** | No official public implementation was located in this review. Reproducing its relational encoder, graph, and inference components may require substantial engineering; that effort has not been measured. | For an initial comparison, **SEELE and ProCNet are the candidate starting points**, because their official implementations are available. LR-DFEE offers ideas for modeling dispersed arguments, and JLF merits inspection once its full method and evaluation are accessible. LR-DFEE's paper reports that SEELE exceeds it on multi-event subsets by 0.5 percentage points on ChFinAnn and 0.4 on DuEE-Fin. This supports retaining SEELE as an instance-separation candidate, though the DuEE-Fin evaluation discrepancy (below) remains unresolved. ([LR-DFEE paper][14]) These 2026 numbers also make state-of-the-art claims that cannot all hold. On ChFinAnn, JLF reports a 10.6% *improvement* (no baseline stated in the accessible abstract), LR-DFEE reports 86.8 F1, and MIFEE reports 90.7 — against a benchmark whose prior best sat in the mid-80s. An improvement percentage without a named baseline is uninterpretable, and a ten-point absolute gain over modern baselines is implausible; treat none of these as an established ranking. **EPAL offers another approach to event-instance separation.** Its event-specific probes and argument libraries are relevant when the system mixes repeated events of the same type. An illustrative case shows ProCNet recovering two of four records while EPAL discovers all four, although EPAL still makes argument errors. No public implementation was located in this review. ([EPAL paper][15]) Fine-tuning a Qwen/Llama model with standard supervised fine-tuning (SFT) may reduce implementation effort relative to reproducing a specialized architecture. Whether this outweighs potential differences in extraction quality or operating cost requires measurement, including structured decoding and validation in both systems. ## Decoder-based extraction An alternative to specialized encoder architectures is to formulate document event extraction as **supervised generation**. Event records are serialized as text, and a causal decoder model is fine-tuned to generate that representation from the source document. The output sequence represents the event instances, which can reduce the need for task-specific neural components. A production system still requires data preparation, output validation, and evaluation. **The evidence reviewed here cannot determine whether a fine-tuned 2–8B decoder matches a specialized encoder on complete, multi-event financial records** — a limit of the evidence inspected, not a claim about either architecture in general. Two Chinese journal papers by the same authors provide relevant generative-extraction results: - **MIFEE** reports 90.7 F1 on ChFinAnn and 77.3 on DuEE-Fin, using instruction-based fine-tuning. The accessible abstract does not establish its backbone size or full scoring protocol, so this review cannot confirm that it fits the 2–8B budget. ([MIFEE abstract][16]) - **LREE** reports 89.9 F1 on ChFinAnn using **Qwen-3-4B**, with gains over its untuned base model and DEEM-PT. The public abstract does not substantiate a DuEE-Fin result. Its full experimental protocol remains unverified here. ([LREE abstract][17]) ### Reported results under each paper's evaluation setting The table collects reported DuEE-Fin scores, not a controlled ranking. There is a concrete split discrepancy: ProCNet reports 6,515/500/1,171 training/development/evaluation documents, whereas LR-DFEE reports 5,115/1,024/1,020. SEELE describes following the ProCNet split procedure. The inspected material does not resolve whether the difference reflects filtering, dataset versions, or a reporting error. ([ProCNet, §4.1][10]; [LR-DFEE, §4.1 and Table 1][14]; [SEELE, §4.1][8]) | Model | Reported DuEE-Fin F1 | Evaluation split / protocol | | :--- | :--- | :--- | | ProCNet, specialized encoder ([ProCNet, original Table 1][10]) | 75.6 | 6,515/500/1,171; own split (verified) | | EPAL, specialized encoder ([EPAL, Table 1][15]) | 76.4 | own re-run; lists ProCNet at 75.1 | | MIFEE, fine-tuned generative model ([MIFEE abstract][16]) | 77.3 | protocol not disclosed | | SEELE, specialized encoder ([SEELE, Table 3][8]) | 80.8 | follows ProCNet split (per paper) | | LR-DFEE, specialized encoder ([LR-DFEE results][14]) | 84.6 | different split: 5,115/1,024/1,020 (per paper) | EPAL's comparison reports ProCNet at 75.1, illustrating that later evaluations do not always reproduce the original score. ([EPAL paper][15]) Dataset splits, preprocessing, and scoring scripts must be reconciled before interpreting cross-paper margins. In particular, the 84.6 versus 80.8 comparison does not establish a verified gain on identical evaluation examples. The structured extraction papers score arguments after matching predicted and gold events, allowing partial credit for incomplete records. **Argument-level F1 is not the proportion of complete event records extracted correctly.** Exact record recovery and same-type instance separation therefore need separate evaluation for the intended application. ([ProCNet evaluation metrics][10]; [SEELE evaluation metrics][8]) ### DocFEE: a financial decoder implementation example **DocFEE** provides a dataset and training code for document-level financial event extraction. It contains 19,044 Chinese listed-company announcements, nine event types, and 38 argument types. Documents average 2,277 Chinese characters and 1.86 events. Its human–LLM collaborative annotation process achieved a reported annotation accuracy of 84.04% under the authors' verification procedure. This qualifies the benchmark's label quality; it is not a direct estimate of a model's accuracy ceiling. ([DocFEE paper][18]) The paper's fine-tuned decoder baseline, LLM_SFT, achieved an event-argument extraction $F_1$ score of 72.5% with 18k training examples (Table 7). **This result is not directly comparable to the DuEE-Fin scores above because it was evaluated on the DocFEE dataset.** It does not establish an advantage or disadvantage against the encoders on a shared benchmark, nor does it measure complete-record correctness. ([DocFEE Table 7][19]) The released implementation uses **Qwen1.5-7B-Chat**, Hugging Face `AutoModelForCausalLM`, parameter-efficient fine-tuning (PEFT) with LoRA, and DeepSpeed. ([DocFEE implementation][20]) The launch script enables LoRA, leaves quantized low-rank adaptation (QLoRA) disabled but configurable, and sets an 8,192-token maximum sequence length. Its default data path names the 4k subset, so the command alone does not reproduce the best 18k result. ([DocFEE launch script][21]) The training code enables truncation. Instructions, document text, and target records share the sequence budget, so complete-announcement coverage cannot be assumed. Longer examples require an explicit strategy for preserving evidence and labels. ([DocFEE training code][22]) A generative workflow combines extraction instructions, document text, and serialized target records into training examples. Standard causal language modeling and LoRA can train the decoder to produce the desired representation. Output conformity and correct event grouping still require evaluation and validation. A new event type can often be represented through a data/schema change without redesigning the neural architecture, but this is not evidence that the model can extract it accurately without further training. One concrete experiment: replace Qwen1.5-7B with a small Qwen3-class model and reuse the existing fine-tuning pipeline, testing whether it improves extraction quality with less engineering than reproducing a specialized encoder. A newer base model is a starting point, not a guarantee of quality. If the taxonomy evolves, adding `DividendRecapitalization`, for example, may require label, output-head, or structured-inference changes in an encoder system. A generative representation may make that interface change easier. However, schema conditioning also exists in encoders such as SEELE, and both approaches may require annotated examples, retraining, and evaluation. Easier schema representation should not be equated with demonstrated transfer to unseen event types. ([SEELE paper][8]) DocFEE illustrates an available generative extraction workflow. Establishing its quality relative to a specialized model requires evaluation on the same held-out documents, with the same event schema and scoring procedure. ### Important trade-offs A decoder's autoregressive generation and larger parameter count may increase inference cost. Encoder pipelines can require document segmentation, multiple encodings, graph operations, and structured inference. Several cited timing results concern training rather than complete production inference. ([ProCNet runtime comparison][10]; [EPAL runtime comparison][15]) Controlled benchmarks on adjacent tasks show substantial inference-efficiency advantages for small specialized encoders over some larger generative decoders. The magnitude depends on the workload and inference implementation; these results do not establish equivalent complete-record quality or a specific deployment saving for financial event extraction. ([Just Pass Twice — NER inference timings, §4.4 and Table 3][23]; [GLiNER Guard — moderation latency, Tables 1–2][24]; [Power Hungry Processing — inference energy, §4.2][25]) Compare the full pipelines on the same workload, including encoding or prefill, event decoding, output parsing, and validation. A cost advantage for the intended financial workload remains a hypothesis until measured. Generative extraction requires checks for malformed output and arguments that do not occur verbatim in the source, alongside extraction errors such as duplicated or omitted events and incorrect fields. Structured encoders can constrain output format and source spans, but they can also omit, duplicate, or misassign event arguments; the relative error rates require evaluation. In production, useful safeguards include JSON/schema-constrained generation, deterministic parsing, span validation against the source document, and confidence or consistency checks. Another concern is **taxonomy specialization**: success on a fixed ontology does not establish performance on new event types. LoRA freezes base weights while learning adapters, but this does not guarantee preservation of the model's behavior. Studies have found less forgetting with LoRA in some settings, while others show that optimizer choice can favor full fine-tuning. Neither finding establishes ontology transfer for these financial extractors. ([LoRA Learns Less and Forgets Less][26]; [Optimizer-Model Consistency][27]) ### Evaluation for the intended application **ProCNet or SEELE versus a schema-conditioned 2–8B Qwen3 decoder** is a practical experiment to run. Prompted commercial or large models (like GLM-5.3) or should remain evaluation baselines; their relative quality and cost depend on the model, prompt, examples, and task. ## References 1. [DuEE-Fin — Baidu dataset][1] 2. [Amazon — Event extraction from documents with co-reference (US20220100963A1)][2] 3. [UCSC — Information Extraction of Corporate Events from the Web][3] 4. [Bloomberg — Bloomberg Elevates Front Office Efficiency With Real-Time Events Data][4] 5. [China Securities Credit Investment — A method and device for extracting financial events (CN115062615B)][5] 6. [S&P Global — ProntoNLP][6] 7. [Mastercard — How Small Can You Go? LoRA Fine-Tuning 270M–8B Models for Merchant Information Extraction in Financial Transactions][7] 8. [SEELE — Incorporating Schema-Aware Description into Document-Level Event Extraction (IJCAI 2024)][8] 9. [SEELE — Official implementation][9] 10. [ProCNet — Document-Level Multi-Event Extraction with Event Proxy Nodes and Hausdorff Distance Minimization (ACL 2023)][10] 11. [ProCNet — Official implementation][11] 12. [JLF — A Joint Learning Framework for Document-Level Event Extraction][12] 13. [JLF — JDEE experimental repository][13] 14. [LR-DFEE — Latent relation-aware graph framework for document-level Chinese financial event extraction][14] 15. [EPAL — Joint Learning Event-Specific Probe and Argument Library with Differential Optimization for Document-Level Multi-Event Extraction (Findings of NAACL 2025)][15] 16. [MIFEE — Multi-dimensional Instruction Set Tuning for Financial Event Extraction with Large Language Models][16] 17. [LREE — Label reconstruction-based financial event extraction model with large language models][17] 18. [DocFEE — A dataset for document level Chinese financial event extraction][18] 19. [DocFEE — Table 7: Event-argument extraction results][19] 20. [DocFEE — Implementation][20] 21. [DocFEE — Fine-tuning launch script][21] 22. [DocFEE — Decoder training code][22] 23. [Just Pass Twice: Efficient Token Classification with LLMs for Zero-Shot NER (ACL 2026)][23] 24. [GLiNER Guard: Unified Encoder Family for Production LLM Safety and Privacy (technical report, 2026)][24] 25. [Power Hungry Processing: Watts Driving the Cost of AI Deployment? (FAccT 2024)][25] 26. [LoRA Learns Less and Forgets Less][26] 27. [Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less][27] [1]: https://aistudio.baidu.com/datasetdetail/186939 [2]: https://patents.google.com/patent/US20220100963A1 [3]: https://nlp.ucsc.edu/capstone-projects/information-extraction-of-corporate-events-from-the-web/ [4]: https://www.prnewswire.com/news-releases/bloomberg-elevates-front-office-efficiency-with-real-time-events-data-302419269.html [5]: https://patents.google.com/patent/CN115062615B/en [6]: https://www.spglobal.com/market-intelligence/en/solutions/products/pronto-nlp [7]: https://arxiv.org/abs/2606.08051 [8]: https://www.ijcai.org/proceedings/2024/0729 [9]: https://github.com/TheoryRhapsody/SEELE [10]: https://aclanthology.org/2023.acl-long.563/ [11]: https://github.com/xnyuwg/procnet [12]: https://pubmed.ncbi.nlm.nih.gov/41926403/ [13]: https://github.com/jokerGrimm/JDEE [14]: https://www.sciencedirect.com/science/article/abs/pii/S0306457326003778 [15]: https://aclanthology.org/2025.findings-naacl.42/ [16]: https://opaj.napstic.cn/periodicalArticle/0120260902360427 [17]: https://www.joca.cn/EN/10.11772/j.issn.1001-9081.2025121480 [18]: https://www.nature.com/articles/s41597-025-05083-9 [19]: https://www.nature.com/articles/s41597-025-05083-9/tables/7 [20]: https://github.com/tongzhou21/DocFEE [21]: https://github.com/tongzhou21/DocFEE/blob/main/baseline/finetune_gpt.sh [22]: https://github.com/tongzhou21/DocFEE/blob/main/baseline/baseline_gpt_sft.py [23]: https://aclanthology.org/2026.acl-long.526.pdf [24]: https://arxiv.org/html/2605.05277v1 [25]: https://arxiv.org/html/2311.16863v3 [26]: https://arxiv.org/abs/2405.09673 [27]: https://arxiv.org/abs/2605.06654