MetaGraph: A Large-Scale Meta-Analysis of
GenAI in Financial NLP (2022–2025)MetaGraph：金融自然语言处理（NLP）领域生成式人工智能（GenAI）的大规模元分析（2022–2025）

Paolo Pedinotti ^†^†thanks: Equal contribution. Paolo Pedinotti contributed to this work during his internship at Bloomberg. Affiliation: Leslie Barrett, Enrico Santus¹¹footnotemark: 1 Peter Baumann Affiliation: Bloomberg Nathan Jessurun Affiliation: {pbaumann25, njessurun, lbarrett4, esantus}@bloomberg.net

Abstract摘要

Financial NLP has evolved rapidly since late 2022, outpacing narrative surveys. We introduce MetaGraph, a methodology for extracting typed knowledge graphs from scientific corpora using ontology-guided LLM extraction to enable structured, large-scale trend analysis. Applied to 681 papers on GenAI in Finance (2022–2025), MetaGraph reveals three phases: early LLM-driven expansion of tasks and datasets, growing emphasis on limitations and risk, and a shift toward modular, system-oriented methods (e.g., retrieval-augmented designs). We release the resulting resource and artifacts to support reproducible meta-analysis and future monitoring of the field.自 2022 年底以来，金融自然语言处理（Financial NLP）领域发展迅速，其演进速度已超越了叙述性综述的更新节奏。我们引入了 MetaGraph，这是一种利用本体引导的大语言模型（LLM）提取技术，从科学语料库中提取类型化知识图谱的方法，旨在实现结构化的大规模趋势分析。通过对 2022 年至 2025 年间 681 篇金融领域生成式人工智能（GenAI）论文的应用分析，MetaGraph 揭示了该领域的三个发展阶段：早期以大语言模型驱动的任务与数据集扩展、对局限性与风险的关注日益增加，以及向模块化、系统导向型方法（如检索增强设计）的转变。我们发布了相关的资源与成果，以支持可重复的元分析及对该领域的持续监测。

1 Introduction1 引言

The release of ChatGPT in late 2022 marked a structural shift in NLP, rapidly accelerating the adoption of generative AI (GenAI), particularly large language models (LLMs), in high-stakes domains such as finance. LLMs expanded the scope of Financial NLP beyond traditional supervised pipelines – long dominated by sentiment analysis and structured extraction – toward flexible, generative systems capable of zero-shot reasoning, long-document processing, and multimodal inputs.2022 年底 ChatGPT 的发布标志着自然语言处理（NLP）领域的结构性转变，迅速加速了生成式人工智能（GenAI），特别是大语言模型（LLM）在金融等高风险领域的应用。大语言模型将金融 NLP 的范畴从传统的监督式流水线（长期以来由情感分析和结构化提取主导）扩展到了具备零样本推理、长文档处理和多模态输入能力的灵活生成式系统。

This rapid evolution has outpaced traditional literature review methodologies. Existing surveys of Financial NLP either predate the widespread use of LLMs or rely on narrative summaries that struggle to capture quantitative trends, structural shifts, and emerging research patterns at scale. As a result, the field lacks a systematic, data-driven view of how tasks, datasets, models, and methods have evolved in response to GenAI.这种快速的演进已经超越了传统的文献综述方法。现有的金融 NLP 综述要么早于大语言模型的广泛应用，要么依赖于难以大规模捕捉定量趋势、结构性转变和新兴研究模式的叙述性总结。因此，该领域缺乏一种系统且数据驱动的视角，以审视任务、数据集、模型和方法如何响应生成式人工智能的兴起而演变。

To address this gap, we introduce MetaGraph, a generalizable methodology for extracting structured knowledge graphs from scientific literature using LLMs. MetaGraph combines a manually defined domain ontology with an LLM-based extraction pipeline to transform unstructured papers into a unified, queryable graph capturing research metadata, tasks, datasets, models, techniques, motivations, and limitations. By design, the ontology and prompts are modular, allowing the same framework to be applied beyond Financial NLP. Figure 1 shows the high-level structure of the resulting graph.为了填补这一空白，我们引入了 MetaGraph，这是一种利用大语言模型从科学文献中提取结构化知识图谱的通用方法。MetaGraph 将手动定义的领域本体与基于大语言模型的提取流水线相结合，将非结构化的论文转化为统一且可查询的图谱，涵盖研究元数据、任务、数据集、模型、技术、动机和局限性。该本体和提示词在设计上具有模块化特性，允许将同一框架应用于金融 NLP 之外的领域。图 1 展示了所构建图谱的高层结构。

We apply MetaGraph to a corpus of 681 Financial NLP papers published between 2022 and 2025, producing a structured, queryable representation of the field. This analysis reveals a research landscape undergoing rapid transformation: the rise of financial question answering, the proliferation of datasets enabled by synthetic generation, increasing attention to model limitations and safety, and a gradual shift from model-centric approaches toward integrated systems combining LLMs with retrieval and other auxiliary components.我们将 MetaGraph 应用于 2022 年至 2025 年间发表的 681 篇金融 NLP 论文语料库，生成了该领域结构化、可查询的表征。分析揭示了一个正在经历快速转型的研究图景：金融问答系统的兴起、合成生成技术赋能的数据集激增、对模型局限性与安全性的关注增加，以及从以模型为中心的方法向结合大语言模型与检索及其他辅助组件的集成系统逐渐转变。

Refer to caption — Figure 1: MetaGraph schema: node types and allowed relations (types appear as attributes in the instantiated graph).图 1：MetaGraph 模式：节点类型及允许的关系（类型在实例化图谱中表现为属性）。

Our primary contribution is this large-scale, quantitative mapping of GenAI-driven transformation in Financial NLP. MetaGraph serves as the enabling framework that makes such structured, field-level meta-analysis possible. We do not introduce a novel knowledge graph learning algorithm; rather, our contribution lies in operationalizing ontology-guided LLM extraction at scale to produce a reproducible representation of an evolving research domain.我们的主要贡献在于对生成式人工智能驱动下金融 NLP 转型进行了大规模的定量映射。MetaGraph 作为实现此类结构化、领域级元分析的支撑框架。我们并未引入一种全新的知识图谱学习算法，而是致力于将本体引导的大语言模型提取技术进行规模化应用，从而产出一种可演进研究领域的可重复表征。

Our contributions are threefold: (i) Field-level analysis: We provide a data-driven map of Financial NLP’s evolution, tracing systematic shifts in tasks, data sources, model choices, risks, and system design. (ii) Methodology: We present MetaGraph, a reusable pipeline for ontology-driven extraction and structured meta-analysis of scientific literature. (iii) Open resource: We release the resulting knowledge graph and artifacts¹¹ 1 The released resource is available at https://zenodo.org/records/16968876. (provided on acceptance) to facilitate reproducible meta-analysis and ongoing monitoring of the field.²² 2 58 of the 681 analyzed papers were excluded from redistribution due to CC-BY-NC-ND 4.0 licensing restrictions on arXiv, which prohibit redistribution of derivative content.我们的贡献体现在三个方面： (i) 领域级分析：我们提供了一份数据驱动的金融 NLP 演进图谱，追踪了任务、数据源、模型选择、风险和系统设计方面的系统性转变。 (ii) 方法论：我们提出了 MetaGraph，这是一种用于科学文献本体驱动提取和结构化元分析的可复用流水线。 (iii) 开放资源：我们发布了所生成的知识图谱及相关成果（将在论文录用后提供，资源地址：https://zenodo.org/records/16968876），以促进可重复的元分析及对该领域的持续监测（由于 arXiv 的 CC-BY-NC-ND 4.0 许可限制禁止重新分发衍生内容，所分析的 681 篇论文中有 58 篇被排除在重新分发范围之外）。

2 Related Work2 相关工作

Financial NLP Surveys.金融 NLP 综述。

While the survey by xing2018 centered on traditional tasks such as classification, more recent surveys (li2023; nie2024surveylargelanguagemodels; du2025financial) have focused on the transformative impact LLMs have had on financial applications. Yet, these works rely on a traditional narrative review methodology, qualitatively summarizing the applications in the literature.尽管 xing2018 的综述集中于分类等传统任务，但较新的综述（li2023; nie2024surveylargelanguagemodels; du2025financial）则聚焦于大语言模型对金融应用产生的变革性影响。然而，这些著作依赖于传统的叙述性综述方法，仅对文献中的应用进行了定性总结。

Our approach diverges fundamentally by employing a bibliometric and holistic analysis of the field. We uncover structural shifts and data-driven trends by quantitatively mapping the research landscape into a knowledge graph, offering a comprehensive view of the impact of LLMs on its evolution. Table 1 summarizes how MetaGraph complements prior surveys along these dimensions.我们的方法从根本上与之不同，它采用了一种文献计量学和整体性的领域分析视角。我们通过将研究图景定量映射为知识图谱，揭示了结构性转变和数据驱动的趋势，从而全面展现了大语言模型对其演进的影响。表 1 总结了 MetaGraph 在这些维度上对以往综述的补充作用。

LLM-Assisted Knowledge Graph Construction.大语言模型辅助的知识图谱构建。

carta2023iterativezeroshotllmprompting construct domain-specific knowledge graphs through stepwise prompting strategies, while FunkHJL23 and 10.1007/978-3-031-47240-4_22 use LLMs to learn hierarchical relations among concepts. Our work takes inspiration from GraphRAG edge2025localglobalgraphrag, which uses LLMs to extract a knowledge graph and enrich it with information at different levels of granularity.carta2023iterativezeroshotllmprompting 通过逐步提示策略构建领域特定的知识图谱，而 FunkHJL23 和 10.1007/978-3-031-47240-4_22 则利用大语言模型学习概念间的层次关系。我们的工作借鉴了 GraphRAG (edge2025localglobalgraphrag) 的思路，该方法利用大语言模型提取知识图谱，并以不同粒度级别的信息对其进行丰富。

Survey	GenAI	Quantit.	Taxon.	KG
Xing et al. (2018)	✗	✗	✗	✗
Li et al. (2023)	✓	✗	✓	✗
Nie et al. (2024)	✓	✗	✓	✗
Du et al. (2025)	✓	✗	✓	✗
Ours (2025)	✓	✓	✓	✓

Table 1: Comparison with Financial NLP surveys.表 1：与金融 NLP 综述的对比。

3 Methodology3 方法论

We introduce MetaGraph, a methodology for automatically constructing knowledge graphs from large scientific corpora to support quantitative analysis. By encoding research entities and their relationships in a structured representation, MetaGraph enables analyses that extend beyond frequency-based statistics and uncover relational patterns that are difficult to detect in unstructured text.我们引入了 MetaGraph，这是一种从大型科学语料库中自动构建知识图谱以支持定量分析的方法。通过将研究实体及其关系编码为结构化表征，MetaGraph 能够实现超越基于频率统计的分析，并揭示在非结构化文本中难以察觉的关系模式。

MetaGraph is designed to be generalizable and reusable, although in this work we apply it specifically to Financial NLP papers.³³ 3 Human validation was limited to structured prompt audits on sampled papers and consolidation of taxonomy labels. No large-scale manual correction of extracted entities was performed. The manual validation effort was modest relative to corpus size, and all extraction, normalization, entity resolution, and graph construction steps were fully automated and designed to scale to larger corpora.MetaGraph 被设计为可推广和可复用的，尽管在本研究中我们将其专门应用于金融 NLP 论文。人工验证仅限于对抽样论文进行结构化提示审计以及对分类标签进行整合。未对提取出的实体进行大规模的人工校正。相对于语料库规模，人工验证的工作量适中，且所有提取、归一化、实体解析和图谱构建步骤均已完全自动化，并被设计为可扩展至更大的语料库。

3.1 Method and Implementation3.1 方法与实现

MetaGraph builds a typed knowledge graph by (i) defining an ontology; (ii) OCR-ing papers; (iii) extracting schema-constrained records with LLM abstention; (iv) normalizing and resolving entities; (v) inducing taxonomies and enriching metadata; and (vi) instantiating the graph and derived signals (Figure 1).MetaGraph 通过以下步骤构建类型化知识图谱：(i) 定义本体；(ii) 对论文进行 OCR 处理；(iii) 在大语言模型弃权机制下提取模式约束记录；(iv) 归一化并解析实体；(v) 归纳分类法并丰富元数据；(vi) 实例化图谱及衍生信号（图 1）。

3.1.1 Ontology Definition.3.1.1 本体定义。

This stage defines the graph structure that determines the analytical scope of MetaGraph. The expressiveness of the ontology directly governs which research patterns can be queried and analyzed.此阶段定义了决定 MetaGraph 分析范围的图谱结构。本体的表达能力直接决定了哪些研究模式可以被查询和分析。

Our ontology consists of three components: entity types, attributes, and relationships.我们的本体由三个部分组成：实体类型、属性和关系。

Figure 1 summarizes entity types and allowed relations; co-occurrence is inferred via shared Paper links. Graph nodes correspond to the core elements of research papers in the area of NLP, for example Tasks, Models, Methods and Motivations. Narrowing the focus, we investigate topics specifically in Financial NLP allowing a more limited inventory of datasets and tasks.图 1 总结了实体类型和允许的关系；共现关系通过共享论文链接进行推断。图谱节点对应于 NLP 领域研究论文的核心要素，例如任务、模型、方法和动机。我们将重点缩小到金融 NLP 的特定主题，从而允许使用更有限的数据集和任务清单。

3.1.2 Corpus Acquisition3.1.2 语料获取

We compiled 681 Financial NLP papers (November 2022–April 2025) from ACL Anthology and arXiv:我们从 ACL Anthology 和 arXiv 汇编了 681 篇金融 NLP 论文（2022 年 11 月至 2025 年 4 月）：

•

ACL Anthology: We queried titles and abstracts of conference papers from 2022--2025 using the official Anthology library, applying an initial filter based on finance-related keywords⁴⁴ 4 The set of keywords is financial, fintech, fraud, stock, portfolio, and finance in either the title or abstract. Abstracts were manually screened for relevance to finance as the primary domain.ACL Anthology：我们使用官方 Anthology 库查询了 2022-2025 年会议论文的标题和摘要，并基于金融相关关键词应用了初步过滤（关键词集为 financial, fintech, fraud, stock, portfolio, finance，出现在标题或摘要中。摘要经过人工筛选，确保金融为主要领域）。
•

arXiv: Similarly, we utilized the arXiv API to search 2023–2025 preprints in the Computation and Language (cs.CL) and the Quantitative Finance (q-fin) categories using the same set of finance keywords.arXiv：同样，我们利用 arXiv API，使用相同的金融关键词集搜索了 2023-2025 年计算与语言（cs.CL）和定量金融（q-fin）类别的预印本。

After aggregating and de-duplicating entries from both sources, manual validation resulted in a final corpus of 681 papers. All papers were obtained in PDF format and processed via Mistral OCR⁵⁵ 5 https://mistral.ai/news/mistral-ocr for text extraction. Table 2 reports the number of papers per year.在汇总并去重来自两个来源的条目后，经人工验证，最终语料库包含 681 篇论文。所有论文均以 PDF 格式获取，并通过 Mistral OCR (https://mistral.ai/news/mistral-ocr) 进行文本提取。表 2 列出了每年的论文数量。

Year	Count	Percent (%)
2022	$126$	$18.50$
2023	$169$	$24.82$
2024	$273$	$40.09$
2025	$113$	$16.59$

Table 2: Number of papers per year in our corpus.表 2：语料库中每年的论文数量。

3.1.3 LLM-based Extraction3.1.3 基于大语言模型的提取

We used Gemini 2.5 Flash⁶⁶ 6 https://deepmind.google/models/gemini/flash/ to extract structured information from papers, selecting it for its strong cost-performance trade-off.⁷⁷ 7 https://lmarena.ai/leaderboard The extraction targets standard research entities, including Tasks, Datasets, Models, Motivations, Techniques, Limitations, and Author and Institution metadata.我们使用 Gemini 2.5 Flash (https://deepmind.google/models/gemini/flash) 从论文中提取结构化信息，因其具有强大的性价比（https://lmarena.ai/leaderboard）。提取目标为标准研究实体，包括任务、数据集、模型、动机、技术、局限性以及作者和机构元数据。

We refined prompts via small-sample audits targeting omissions, over-generation, and schema violations; updates emphasized task definitions, abstention, and output constraints (Appendix B). Based on the observed patterns, we updated the prompt instructions (e.g., clarifying task definitions, strengthening abstention rules, and constraining output formats). Refinement stopped once the audit showed no recurring systematic errors. The full set of prompts is provided in Appendix B.我们通过针对遗漏、过度生成和模式违规的小样本审计来优化提示词；更新重点在于任务定义、弃权规则和输出约束（附录 B）。基于观察到的模式，我们更新了提示指令（例如，明确任务定义、加强弃权规则并约束输出格式）。当审计显示不再出现反复的系统性错误时，优化工作停止。完整的提示词集见附录 B。

Multi-instance representation.多实例表征。

Each Paper links to multiple Task/Dataset/Model nodes (many-to-many). Figure 2 illustrates this structure.每篇论文链接到多个任务/数据集/模型节点（多对多关系）。图 2 展示了此结构。

论文 PPFinQAGPT-4FinBERT使用数据集评估模型评估模型测试于

Figure 2: Example of multi-instance representation in the knowledge graph: a paper evaluates multiple models on a dataset (datasets in blue, models in green).图 2：知识图谱中多实例表征的示例：一篇论文在某个数据集上评估了多个模型（数据集为蓝色，模型为绿色）。

Validation.验证。

Validation on 12 gold papers found only minor omissions (two tasks, one model) and no hallucinations.对 12 篇黄金标准论文的验证仅发现少量遗漏（两个任务，一个模型），未发现幻觉现象。

3.1.4 Entity Resolution3.1.4 实体解析

We resolved surface-form inconsistencies using normalization followed by embedding-based clustering. Mentions were first lowercased and stripped of punctuation and formatting artifacts. For each entity type independently (e.g., Dataset, Model), normalized mentions were embedded using OpenAI’s text-embedding-small. Pairwise cosine similarity was computed, and mentions were merged if cosine similarity was $\geq 0.93$ , using greedy agglomeration.我们使用归一化及基于嵌入的聚类方法解决了表面形式的不一致性。首先将提及内容转换为小写，并去除标点符号和格式伪影。对于每种实体类型（如数据集、模型），使用 OpenAI 的 text-embedding-small 对归一化后的提及进行嵌入。计算两两余弦相似度，若相似度 ≥0.93，则使用贪婪聚合方法合并提及内容。

The threshold $\tau=0.93$ was chosen conservatively to avoid merging semantically related but distinct entities. For example, Finqa $\rightarrow$ FinQA, FIQA-SA variants $\rightarrow$ FIQA-SA, while FinQA and ConvFinQA remained distinct.阈值 τ=0.93 的选择较为保守，以避免合并语义相关但不同的实体。例如，Finqa → FinQA，FIQA-SA 变体 → FIQA-SA，而 FinQA 和 ConvFinQA 则保持独立。

3.1.5 Taxonomy Induction3.1.5 分类法归纳

Selected entity types were organized into taxonomies using a zero-shot LLM-based categorization procedure. We processed batches of maximum 100 entities at a time. For each batch, the LLM was prompted to propose recurring category types and assign entities accordingly. We manually consolidated categories by merging near-duplicates, removing singletons, and standardizing naming. This final step affected only taxonomy naming accuracy; entity extraction remained fully automated.选定的实体类型使用基于零样本大语言模型的分类程序组织成分类法。我们每次处理最多 100 个实体的批次。对于每个批次，提示大语言模型提出重复出现的类别类型并相应地分配实体。我们通过合并近义词、删除单例和标准化命名来人工整合类别。此最终步骤仅影响分类命名准确性；实体提取过程保持完全自动化。

Institution type labeling.机构类型标注。

Institutions were labeled as academic, industry, or mixed using rule-based heuristics applied to affiliation strings and Semantic Scholar metadata⁸⁸ 8 https://api.semanticscholar.org/. Keywords such as University and College triggered the academic label, while terms such as Inc., Ltd., etc., triggered industry. Papers with affiliations spanning categories were labeled mixed. The procedure was fully deterministic and did not rely on LLM inference.机构使用基于规则的启发式方法，根据隶属字符串和 Semantic Scholar 元数据 (https://api.semanticscholar.org/) 标注为学术界、工业界或混合型。诸如 University 和 College 等关键词触发学术界标签，而 Inc.、Ltd. 等术语触发工业界标签。跨类别的机构被标注为混合型。该过程完全确定性，不依赖大语言模型推理。

3.1.6 Relevance Scoring3.1.6 相关性评分

To highlight emerging trends, we define a paper-level relevance score used to prioritize analyses toward more influential work. The score combines three factors: (i) institutional centrality, computed as the PageRank of affiliated institutions in the co-authorship graph; (ii) productivity, measured as the number of papers published by the institution; and (iii) citation normalization, defined as paper citations normalized by the average citation count of the publication year.为了突出新兴趋势，我们定义了一个论文级相关性评分，用于优先分析更具影响力的工作。该评分结合了三个因素：(i) 机构中心性，计算为合作图谱中附属机构的 PageRank；(ii) 生产力，衡量为机构发表的论文数量；(iii) 引文归一化，定义为论文引文数除以出版年份的平均引文数。

4 Findings and Insights4 发现与洞察

In this section, we demonstrate the types of analyses enabled by MetaGraph, focusing on how the release of ChatGPT in late 2022 marked a turning point for Financial NLP. For comparability, we partition the corpus into three chronological subsets of approximately equal size: T1 (January 2022–August 2023), T2 (September 2023–July 2024), and T3 (August 2024–April 2025). All temporal analyses in this section use these fixed partitions. The main findings are:在本节中，我们展示了 MetaGraph 所支持的分析类型，重点关注 2022 年底 ChatGPT 的发布如何成为金融 NLP 的转折点。为了可比性，我们将语料库划分为三个大致等长的时段：T1（2022 年 1 月至 2023 年 8 月）、T2（2023 年 9 月至 2024 年 7 月）和 T3（2024 年 8 月至 2025 年 4 月）。本节中的所有时间分析均使用这些固定分区。主要发现如下：

•

Task emphasis shifts toward QA variants (Figure 3; Table 3).任务重点转向问答（QA）变体（图 3；表 3）。
•

The dataset landscape fragments and data sources diversify (Table 4; Table 5).数据集格局碎片化，数据源多样化（表 4；表 5）。
•

Reported limitations shift from data scarcity toward model- and safety-related concerns (Figure 6; Table 6).报告的局限性从数据稀缺转向模型及安全性相关问题（图 6；表 6）。
•

Methods evolve from prompting-centric approaches toward system-level designs (Figure 7; Table 7).方法从以提示词为中心演进为系统级设计（图 7；表 7）。

More details and insights are provided in Appendices A, B, C.更多细节和洞察见附录 A、B、C。

4.1 Release of ChatGPT4.1 ChatGPT 的发布

Before the release of ChatGPT, the field focused mainly on sentiment analysis, information extraction, and stock prediction. In the period immediately following the release of ChatGPT (November 2022–February 2024), these tasks still constituted 90% of published work, and the most widely used datasets (Table 4(a)) reflected this focus.在 ChatGPT 发布之前，该领域主要关注情感分析、信息提取和股价预测。在 ChatGPT 发布后的初期（2022 年 11 月至 2024 年 2 月），这些任务仍占据发表工作的 90%，且最广泛使用的数据集（表 4(a)）反映了这一重点。

As usage of LLMs matured (February 2024–April 2025, Table 4(b)), the landscape evolved unlocking new applications and attention toward more complex and reasoning-intensive tasks. Financial QA has become the leading focus, rising from 10% to 33% of tasks by 2025, while traditional tasks have steadily declined, as shown in Table 3 for representative tasks (temporal distribution of the full taxonomy in Appendix Table 11).随着大语言模型应用趋于成熟（2024 年 2 月至 2025 年 4 月，表 4(b)），格局发生了演变，开启了新的应用，并对更复杂、推理密集型的任务给予了关注。金融问答已成为主要焦点，其任务占比从 2025 年的 10% 上升至 33%，而传统任务则稳步下降，如表 3 中代表性任务所示（完整分类法的时序分布见附录表 11）。

LLMs have transformed the way researchers approach financial problems. This shift moves from narrow, task-specific pipelines to flexible, generative systems that bridge previously isolated tasks. Between April 2023 and February 2024, the average number of tasks per paper rose from 1.36 to 1.9. Traditional tasks such as sentiment analysis and information extraction are now often used as intermediate steps in broader systems, such as RAG and financial agents.大语言模型改变了研究人员处理金融问题的方式。这种转变从狭窄的任务特定流水线转向了能够桥接先前孤立任务的灵活生成式系统。在 2023 年 4 月至 2024 年 2 月期间，每篇论文的平均任务数从 1.36 升至 1.9。情感分析和信息提取等传统任务现在常被用作更广泛系统（如 RAG 和金融智能体）中的中间步骤。

Data Sources and Datasets.数据源与数据集。

Datasets evolved as well. On the one hand, QA benchmarks now lead the field, overtaking traditional datasets (Table 4). On the other hand, we witnessed an expansion and diversification of the data sources used to generate QA datasets. Recent papers increasingly mention multimodal and structured inputs – such as tables, charts, audio, and analyst commentaries – alongside core sources such as news and company reports (Table 5). This expansion has been supported by synthetic data generation, which reduced the need for expert annotations. The share of synthetic or human-in-the-loop datasets nearly tripled as LLMs became data generators, from 5% in April 2023 to almost 15% by November 2024 (krumdick-etal-2024-bizbench; guo-yang-2024-econnli; liu2025evaluatingaligninghumaneconomic; li2023cfgptchinesefinancialassistant). Data trends for tasks are plotted in Figure 8. Most new datasets target QA tasks, while the development of datasets for other tasks, such as sentiment analysis, has slowed.数据集也在演变。一方面，问答基准现已领先于该领域，超越了传统数据集（表 4）。另一方面，我们见证了用于生成问答数据集的数据源的扩展与多样化。近期论文越来越多地提及多模态和结构化输入——如表格、图表、音频和分析师评论——以及新闻和公司报告等核心来源（表 5）。这种扩展得益于合成数据生成，减少了对专家标注的需求。随着大语言模型成为数据生成器，合成或人机协作数据集的份额几乎翻了三倍，从 2023 年 4 月的 5% 增长到 2024 年 11 月的近 15% (krumdick-etal-2024-bizbench; guo-yang-2024-econnli; liu2025evaluatingaligninghumaneconomic; li2023cfgptchinesefinancialassistant)。任务的数据趋势绘制在图 8 中。大多数新数据集针对问答任务，而其他任务（如情感分析）的数据集开发有所放缓。

4.2 A Growing Awareness4.2 日益增长的意识

LLMs have lowered key barriers to both adoption and data processing. On the one hand, they remove data format constraints – enabling the processing of unstructured data. On the other hand, they support synthetic data generation, helping mitigate challenges such as cost, scarcity, and domain bias (Table 8). We show how limitations have changed over time in Figure 6. As data constraints eased, research attention increasingly shifted toward model-level challenges – particularly reasoning, interpretability, efficiency, and safety. We observed growing concerns around bias, privacy risks, and potential misuse (Table 6).大语言模型降低了采用和数据处理的关键门槛。一方面，它们消除了数据格式限制——实现了非结构化数据的处理。另一方面，它们支持合成数据生成，有助于缓解成本、稀缺性和领域偏差等挑战（表 8）。我们在图 6 中展示了局限性随时间的变化。随着数据约束的缓解，研究注意力日益转向模型层面的挑战——特别是推理、可解释性、效率和安全性。我们观察到对偏差、隐私风险和潜在滥用的担忧日益增长（表 6）。

This shift toward critical reflection is evident in the evolution of research motivations, which increasingly convey a more cautious stance toward LLMs. By 2024, critical themes such as robustness, efficiency, reasoning, and RAG appeared in nearly 18% of papers – twice the share observed in early 2023 (see Table 9). This marks a shift from earlier studies, which primarily focused on leveraging LLMs through zero-shot learning and fine-tuning.这种向批判性反思的转变在研究动机的演变中显而易见，这些动机越来越多地传达出对大语言模型更谨慎的立场。到 2024 年，稳健性、效率、推理和 RAG 等批判性主题出现在近 18% 的论文中——是 2023 年初观察到的份额的两倍（见表 9）。这标志着与早期研究的转变，早期研究主要关注通过零样本学习和微调来利用大语言模型。

This marks a move from standalone LLMs to system-oriented design. Prompting strategies have evolved as well (Table 10): the progression from in-context learning to augmented methods such as chain-of-thought, retrieval-based prompts, and self-criticism reflects a move away from relying solely on the model’s few-shot capabilities toward more deliberate prompt enrichment aimed at reducing errors.这标志着从独立大语言模型向系统导向设计的转变。提示词策略也发生了演变（表 10）：从上下文学习到思维链、检索增强提示和自我批评等方法的演进，反映了从仅依赖模型少样本能力向旨在减少错误的更审慎的提示词丰富化转变。

Task	T1	T2	T3
Financial Sentiment Analysis	$15.67$	$15.77$	$9.78$
NER	$7.21$	$8.67$	$5.32$
Stock Price Change Prediction	$12.19$	$7.63$	$8.20$
Retrieval Enhanced QA	$5.47$	$7.80$	$12.81$
Numerical QA	$5.47$	$7.45$	$9.50$
Long Document QA	$3.23$	$4.51$	$7.34$
Financial Consulting QA	$2.24$	$4.16$	$6.04$
Claim Verification	$1.24$	$1.21$	$3.31$

Table 3: Distribution (in % of papers) of selected Financial NLP tasks across time periods (T1: Jan 2022–Aug 2023, T2: Sep 2023–Jul 2024, T3: Aug 2024–Apr 2025).表 3：各时段内选定金融 NLP 任务的分布（论文占比 %）（T1：2022 年 1 月至 2023 年 8 月，T2：2023 年 9 月至 2024 年 7 月，T3：2024 年 8 月至 2025 年 4 月）。

Dataset	Task	Freq.
FPB https://doi.org/10.1002/asi.23062	SA	29
FinQA chen-etal-2021-finqa	QA	19
FIQA-SA 10.1145/3184558.3192301	SA	15
ConvFinQA chen-etal-2022-convfinqa	QA	13
RefInd 10.1145/3539618.3591911	RE	7

(a) November 2022 – February 2024(a) 2022 年 11 月 – 2024 年 2 月

Dataset	Task	Freq.
ConvFinQA chen-etal-2022-convfinqa	QA	13
FPB https://doi.org/10.1002/asi.23062	SA	13
FinQA chen-etal-2021-finqa	QA	12
FIQA-SA 10.1145/3184558.3192301	SA	7
FinanceBench islam2023financebenchnewbenchmarkfinancial	QA	7

(b) February 2024 – April 2025(b) 2024 年 2 月 – 2025 年 4 月

Table 4: Top datasets in Financial NLP by usage across two time periods. (QA: Question Answering, SA: Sentiment Analysis, RE: Relation Extraction).表 4：两个时段内按使用量排名的金融 NLP 顶级数据集。（QA：问答，SA：情感分析，RE：关系提取）。

4.3 From Models to Systems4.3 从模型到系统

In the wake of GenAI’s rise, researchers initially focused on adapting general-purpose LLMs to Financial NLP tasks through prompt engineering – especially zero-shot and in-context learning – which quickly gained momentum across applications. This was often complemented by post-training methods such as instruction tuning to further specialize models for the financial domain (Table 10).在生成式人工智能兴起之后，研究人员最初专注于通过提示工程——特别是零样本和上下文学习——将通用大语言模型适配到金融 NLP 任务中，这在各类应用中迅速获得动力。这通常辅以指令微调等训练后方法，以进一步使模型在金融领域专业化（表 10）。

Researchers began to move beyond model-centric approaches as the limitations of reasoning, safety, interpretability, and scalability became more apparent. Over time, these techniques were increasingly complemented by system-level innovations that integrate LLMs into broader frameworks (Figure 7).随着推理、安全性、可解释性和可扩展性的局限性日益显现，研究人员开始超越以模型为中心的方法。随着时间的推移，这些技术越来越多地辅以将大语言模型集成到更广泛框架中的系统级创新（图 7）。

The most prominent of these is RAG (10.1145/3604237.3626866; xue2024weaverbirdempoweringfinancialdecisionmaking; li-etal-2024-alphafin; yepes2024financialreportchunkingeffective; chen-etal-2024-knowledge; zhao-etal-2024-optimizing, among others), which has become a cornerstone of the field. Examining the evolution of RAG (Table 7), we find it mirrors the dataset trend: the spectrum of source types and data formats has widened, knowledge bases have grown, and the size of retrieved context has expanded from single sentences to large document chunks.其中最突出的是 RAG (10.1145/3604237.3626866; xue2024weaverbirdempoweringfinancialdecisionmaking; li-etal-2024-alphafin; yepes2024financialreportchunkingeffective; chen-etal-2024-knowledge; zhao-etal-2024-optimizing 等)，它已成为该领域的基石。审视 RAG 的演进（表 7），我们发现它反映了数据集的趋势：源类型和数据格式的范围已扩大，知识库已增长，检索上下文的大小已从单句扩展到大型文档块。

4.4 Towards Maturity4.4 迈向成熟

As the field matured, researchers began prioritizing shared resources over creating new datasets, increasingly relying on established, literature-backed benchmarks (Figure 4), with growing coverage across tasks. A similar trend emerged on the modeling side, as the community increasingly turned to open-source models valued for their transparency, controllability, and adaptability, alongside a shift from rapid expansion toward critical evaluation (Figure 5).随着领域成熟，研究人员开始优先考虑共享资源而非创建新数据集，越来越多地依赖于既定的、有文献支持的基准（图 4），并在各任务中覆盖面不断扩大。在模型方面出现了类似的趋势，社区越来越多地转向因透明度、可控性和适应性而受到重视的开源模型，同时从快速扩张转向批判性评估（图 5）。

Figure 9 illustrates three key phases: the early dominance of GPT models, the emergence of LLaMA touvron2023llamaopenefficientfoundation, and the current diversification toward a mix of open models – such as Qwen bai2023qwentechnicalreport and DeepSeek deepseekai2025deepseekr1incentivizingreasoningcapability – and proprietary ones. Figure 10 shows how model sizes have also changed over time. The field is revisiting cost-performance trade-offs, driven by the financial and computational cost of large models. This shift is reflected in a recent inflection in model size trends.图 9 展示了三个关键阶段：GPT 模型的早期主导地位、LLaMA (touvron2023llamaopenefficientfoundation) 的出现，以及当前向开源模型（如 Qwen (bai2023qwentechnicalreport) 和 DeepSeek (deepseekai2025deepseekr1incentivizingreasoningcapability)）与专有模型混合的多元化转变。图 10 展示了模型规模随时间的变化。受大模型财务和计算成本的驱动，该领域正在重新审视成本与性能的权衡。这种转变反映在模型规模趋势的近期拐点上。

Data	T1	T2	T3
Sources	(%)	(%)	(%)
News	$27.48$	$29.14$	$25.35$
Social Media / Forums	$21.85$	$14.20$	$14.43$
Company Reports	$28.15$	$27.37$	$28.99$
Company Fundamentals & Indicators	$11.04$	$15.09$	$15.13$
Earnings Calls	$5.18$	$7.10$	$6.58$
Analyst Reports	$4.73$	$3.25$	$3.92$
University Textbooks	$1.13$	$0.89$	$2.38$
Financial Analyst Exams	$0.45$	$2.96$	$3.22$
Signals
Text	$80.42$	$73.72$	$70.44$
Tables	$16.08$	$19.87$	$20.13$
Image	$1.40$	$2.56$	$5.03$
Audio	$0.70$	$1.92$	$1.89$
Other	$2.10$	$1.92$	$2.52$

Table 5: Distribution of data sources and signal types across time periods (T1: Jan 2022–Aug 2023, T2: Sep 2023–Jul 2024, T3: Aug 2024–Apr 2025).表 5：各时段数据源和信号类型的分布（T1：2022 年 1 月至 2023 年 8 月，T2：2023 年 9 月至 2024 年 7 月，T3：2024 年 8 月至 2025 年 4 月）。

Risk / Limitation	T1	T2	T3
Dominant Issues
Misleading Predictions due to Outdated Data	$60$	$100$	$119$
Misleading Predictions due to Inaccuracies	$50$	$66$	$75$
Lack of Transparency/Explainability	$27$	$66$	$94$
Inability to Generalize Across Financial Tasks	$27$	$58$	$64$
Growing Concerns
Biases Towards Stocks/Trends/Products	$12$	$30$	$43$
Breaches of Sensitive Data	$4$	$16$	$25$
Susceptibility to Attacks and Misuse	$10$	$12$	$25$
Misinterpretation of Regulatory Text	$1$	$7$	$21$
Gender or Demographic Bias	$6$	$21$	$17$
Generation of Fraudulent Content	$3$	$7$	$15$
Stable Low-Level Issues
Sensitivity to Data/Market Shifts	$25$	$36$	$33$
Inability to Detect Misinformation	$5$	$7$	$8$
Susceptibility to Corpus Poisoning	$0$	$9$	$3$

Table 6: Classification of LLM risks in finance based on their frequency over time periods (T1: Jan 22–Aug 23, T2: Sep 23–Jul 24, T3: Aug 24–Apr 25).表 6：金融领域大语言模型风险的分类，基于其在各时段的频率（T1：22 年 1 月至 23 年 8 月，T2：23 年 9 月至 24 年 7 月，T3：24 年 8 月至 25 年 4 月）。

RAG Configuration	T1	T2	T3
Data Source Size
Small	$38.24$	$33.33$	$34.15$
Medium	$20.59$	$17.54$	$14.63$
Large	$41.18$	$49.12$	$51.22$
Data Source Type
Text	$66.67$	$64.71$	$60.51$
Table	$18.33$	$21.57$	$20.38$
Database	$5.00$	$7.84$	$10.83$
Graph	$8.33$	$2.94$	$5.10$
Image	$1.67$	$2.94$	$2.55$
Other	$0.00$	$0.00$	$0.64$
Retrieved Text Granularity
Chunk	$84.85$	$87.72$	$95.18$
Sentence	$15.15$	$12.28$	$4.82$

Table 7: Distribution of RAG configurations across time periods.表 7：各时段 RAG 配置的分布。

Data	T1	T2	T3
Data-related Limitations
Costly Human Judging	$3.20$	$3.22$	$2.83$
Insufficient Data Scale/Coverage	$12.45$	$13.00$	$10.59$
Skewed/Imbalanced Classes	$3.08$	$2.29$	$1.47$
Domain/Language Bias	$12.03$	$11.69$	$9.59$
LLM Limitations
Interpretability Gaps	$1.63$	$2.29$	$2.04$
Weak Reasoning	$4.11$	$4.59$	$5.24$
Cost & Environmental Footprint	$2.66$	$2.79$	$3.46$
Hallucination & Bias	$2.18$	$2.95$	$3.72$
Prompt Sensitivity	$1.75$	$2.84$	$2.78$
Latency / Scalability	$2.06$	$2.18$	$2.57$
Synthetic Data / Label Issues	$7.38$	$4.86$	$5.14$
Capacity Constraints	$9.07$	$9.07$	$9.70$
Gaps: Lab vs. Live	$9.43$	$8.68$	$10.27$
Other (Appendix C.2)	$28.81$	$29.55$	$30.06$

Table 8: Distribution of reported limitations across time periods.表 8：各时段报告的局限性分布。

Data	T1	T2	T3
Data
Data Scarcity & Annotation Cost	$32.25$	$28.82$	$23.00$
Other	$37.12$	$35.81$	$35.48$
Exploiting LLMs
Zero/Few-Shot Evaluation	$4.41$	$4.37$	$2.53$
Domain-Specific LLM Training	$10.44$	$8.95$	$11.89$
Solving LLM Limitations
Quantitative Reasoning Gaps	$5.10$	$5.02$	$6.43$
Interpretability & Explainability	$3.25$	$3.71$	$5.07$
Efficiency Constraints	$3.02$	$5.24$	$5.65$
Safety, Robustness, & Fairness	$2.78$	$5.90$	$7.21$
RAG & Retrieval Bottlenecks	$1.62$	$2.18$	$2.73$

Table 9: Distribution of future research directions across time periods.表 9：各时段未来研究方向的分布。

Technique	T1	T2	T3
Core Prompting
Zero-Shot	$69$	$126$	$162$
Few-Shot	$20$	$98$	$74$
Chain-of-Thought	$13$	$61$	$94$
Augmentation Strategies
RAG (Retrieval-Augmented)	$39$	$65$	$92$
Decomposition	$23$	$52$	$58$
Self-Criticism	$4$	$13$	$30$
High-Level Methods
Ensembling	$17$	$14$	$13$
Agents	$9$	$25$	$40$

Table 10: Frequency counts of prompt engineering techniques across time periods.表 10：各时段提示工程技术的频率统计。

One Revolution, Two Speeds. GenAI reshaped industry and academia at different paces (Figure 11). We took all the instances of tasks, models, and datasets in our corpus, and computed the relative proportion of financial QA instances, open-model instances, new datasets (datasets created after 2022), and created datasets (datasets created by the same authors who use them). Industry moved faster – dominating financial QA and driving dataset innovation to stay competitive. Academia responded more cautiously, focusing on established tasks and open-source models, with a stronger emphasis on transparency and reproducibility. This is likely due to academia’s structural constraints, which prioritize transparency, reproducibility, and the use of publicly available data and models – factors that inherently slow down the adoption of cutting-edge approaches. In contrast, industry has largely traded off transparency in favor of rapid experimentation, leveraging proprietary data and closed-source LLMs to push forward advanced use cases such as financial QA.一场革命，两种速度。生成式人工智能以不同的节奏重塑了工业界和学术界（图 11）。我们提取了语料库中所有任务、模型和数据集的实例，并计算了金融问答实例、开源模型实例、新数据集（2022 年后创建的数据集）以及自创数据集（由使用它们的研究人员创建的数据集）的相对比例。工业界行动更快——主导了金融问答并推动了数据集创新以保持竞争力。学术界反应较为谨慎，专注于既定任务和开源模型，更强调透明度和可重复性。这很可能是由于学术界的结构性约束，即优先考虑透明度、可重复性以及公开可用数据和模型的使用——这些因素本质上减缓了尖端方法的采用。相比之下，工业界在很大程度上以牺牲透明度为代价换取快速实验，利用专有数据和闭源大语言模型来推进金融问答等高级用例。

4.5 Looking Ahead4.5 前瞻

Financial NLP is entering a new phase, driven not only by LLMs but by a deeper understanding of their strengths and limitations. As techniques such as RAG and open-source fine-tuning become standard (grey line in Figure 12), multimodal models and small language models (green line) are gaining traction. New trends are also emerging (blue line in Figure 12), most notably multi-agent systems, which range from simple expert–critic setups to more complex architectures. At the same time, the gap between academic research and real-world financial practice remains an open question, as the field shifts its focus from question answering toward reasoning-oriented systems.金融 NLP 正在进入一个新阶段，不仅由大语言模型驱动，还由对其优势和局限性的更深入理解所驱动。随着 RAG 和开源微调等技术成为标准（图 12 中的灰线），多模态模型和小语言模型（绿线）正获得关注。新趋势也在涌现（图 12 中的蓝线），最显著的是多智能体系统，其范围从简单的专家-批评者设置到更复杂的架构。与此同时，随着该领域将重点从问答转向面向推理的系统，学术研究与现实金融实践之间的差距仍然是一个悬而未决的问题。

5 Conclusion5 结论

We presented a structured, quantitative meta-analysis of GenAI-driven transformation in Financial NLP, based on a corpus of 681 papers from 2022–2025. Our analysis reveals systematic shifts in tasks, datasets, risks, and architectural paradigms, marking a transition from model-centric experimentation to system-oriented design.我们基于 2022 年至 2025 年间的 681 篇论文语料库，对生成式人工智能驱动下的金融 NLP 转型进行了结构化、定量的元分析。我们的分析揭示了任务、数据集、风险和架构范式的系统性转变，标志着从以模型为中心的实验向系统导向设计的过渡。

MetaGraph serves as the enabling framework that operationalizes ontology-guided LLM extraction at scale, producing a reproducible and queryable representation of the field. Beyond Financial NLP, this approach illustrates how structured knowledge graph construction can support data-driven monitoring of rapidly evolving research domains.MetaGraph 作为实现本体引导大语言模型规模化提取的支撑框架，产出了该领域可重复且可查询的表征。除了金融 NLP 之外，该方法还展示了结构化知识图谱构建如何支持对快速演进研究领域的实时数据监测。

6 Limitations6 局限性

•

Our approach relies on a manually defined ontology, which introduces an inductive bias in how entities and relations are categorized. While this provides structure and interpretability, it may also limit flexibility and overlook alternative or emergent conceptualizations.我们的方法依赖于手动定义的本体，这在实体和关系分类方式上引入了归纳偏差。虽然这提供了结构和可解释性，但也可能限制灵活性并忽略其他或新兴的概念化方式。
•

Despite continuous human validation and refinement, the entity extraction and taxonomy induction processes remain based on LLMs, which are inherently susceptible to hallucinations, inaccuracies, and bias. These limitations may affect both the precision and completeness of the extracted knowledge.尽管经过持续的人工验证和优化，实体提取和分类法归纳过程仍基于大语言模型，其本质上易受幻觉、不准确和偏差的影响。这些局限性可能会影响提取知识的精确性和完整性。
•

The initial selection of papers was based on heuristic keyword search. This approach may have missed some papers that should be considered part of Financial NLP.论文的初步筛选基于启发式关键词搜索。这种方法可能遗漏了一些应被视为金融 NLP 一部分的论文。

Acknowledgments致谢

We thank David Rosenberg for his detailed feedback, which substantially improved this paper.感谢 David Rosenberg 提供的详细反馈，这些反馈极大地改进了本文。

References参考文献

Appendix A The Geography of Financial NLP附录 A 金融 NLP 的地理分布

The map depicted in Figure 13 illustrates the geographical distribution of institutions represented in our corpus. It highlights that Financial NLP research predominantly clusters around three major global hubs. In the United States, research activity is highly concentrated along the Atlantic Coast, with a distinct epicenter in New York City. In East Asia, significant research centers have emerged in major economic and technological hubs, notably within China, Korea, Japan, Hong Kong, and Singapore. Europe, on the other hand, presents a different pattern: research activities are more dispersed, reflecting a fragmented landscape with multiple smaller centers rather than a single dominant hub. This contrasting geography suggests regional differences in collaboration patterns, institutional density, and possibly cultural or economic factors influencing research organization in Financial NLP.图 13 中描绘的地图展示了我们语料库中所代表机构的地理分布。它突出表明金融 NLP 研究主要集中在全球三个主要中心。在美国，研究活动高度集中在大西洋沿岸，纽约市是明显的中心。在东亚，重要的研究中心已在主要经济和技术枢纽中涌现，特别是在中国、韩国、日本、香港和新加坡。另一方面，欧洲呈现出不同的模式：研究活动更为分散，反映了一个由多个较小中心组成的碎片化格局，而非单一的主导中心。这种地理差异暗示了在合作模式、机构密度以及可能影响金融 NLP 研究组织的文化或经济因素方面存在区域差异。

A.1 Financial NLP Graph Sub-treesA.1 金融 NLP 图谱子树

We show a snippet of the ontology emphasizing the connections between the papers on financial topics, the models used and datasets used. For readability, we do not include author names and institutions, paper motivations or limitations. Figure 14 shows the subtree.我们展示了本体的一个片段，强调了金融主题论文、所用模型和所用数据集之间的联系。为了可读性，我们没有包含作者姓名和机构、论文动机或局限性。图 14 展示了该子树。

Appendix B Prompts for Graph Extraction and Enrichment附录 B 图谱提取与丰富的提示词

This section details the prompts designed for extracting and enriching the graph structure from the corpus of Financial NLP research papers. The prompts were crafted to ensure clarity, conciseness, and consistency in extracting different entity types and relationships, facilitating accurate and systematic analysis.本节详细介绍了旨在从金融 NLP 研究论文语料库中提取和丰富图谱结构的提示词。这些提示词经过精心设计，旨在确保提取不同实体类型和关系时的清晰度、简洁性和一致性，从而促进准确和系统的分析。

B.1 Graph ExtractionB.1 图谱提取

For entities such as limitations, motivations, and techniques, a unique standardized prompt format was adopted to maintain consistency across extraction tasks. Given an entity type represented by X (e.g., limitations), the prompt follows the format shown in Listing 1.对于局限性、动机和技术等实体，采用了独特的标准化提示格式，以在提取任务中保持一致性。给定由 X 代表的实体类型（例如，局限性），提示遵循清单 1 中所示的格式。

⬇

1 You你 will将 be given获得 the全文 full论文 text该 of a paper论文 that应用了 applies技术 NLP techniques于 to financial金融 data数据 analysis分析. Your你的 goal目标 is to identify是识别 the文中提到的 {ENTITY TYPE} mentioned该 in the论文 paper你.

2 You应该 should总结 summarize每一个 each简短 {ENTITY TYPE} in a brief且 and简洁的 concise描述 description不超过 (no more50 than个 50)单词 words算法 1：用于提取局限性、动机和技术的提示词格式.

Algorithm 1: Format of the prompt used for extraction of limitations, motivations, and techniques清单 2 中详述的另一个单独提示词被开发出来，专门用于提取任务、数据集和模型。此特定的提示词对于准确捕捉这些实体之间的关系是必要的。其他类型的关系，如在同一篇论文中共同出现的实体，是从提取的信息中隐式推断出来的，因此不需要额外的提示。

A separate prompt, detailed in Listing 2, was developed to specifically extract tasks, datasets, and models. This distinct prompt was necessary to accurately capture relationships among these entities. Other types of relationships, such as entities co-occurring within the same paper, were implicitly inferred from extracted information and thus did not require additional prompting.你

⬇

1 You将 will获得 be given鉴于 the全文 full全文 text文章 of an article文章 that该 applies应用 NLP to financial金融 data数据. Your你的 goal目标 is to answer回答 with使用 the以下 following信息 information信息:

3 - General常规 Reasoning推理: reason推理 step逐步-by-step逐步 to answer回答 the以下 following以下 questions问题: which这些 are是 the该 tasks任务 the论文 paper论文 evaluates评估 the模型 models模型 on? Which哪些 are是 the用于 datasets数据集 used用于 for每个 each每个 task任务?

4 Use使用 the上一步 result推理 of the输出 previous的 reasoning结果 to output输出:

5 - Tasks任务: a list列出 of tasks论文评估模型所用的任务 the论文 paper论文 evaluates评估 the模型 models模型 on. For对于 each每个 task任务, you你 have必须 to specify明确:

6 - Datasets数据集: a list列出 of datasets用于测试任务模型的数据集 used用于 to test测试 the模型 models模型 on the该 task任务.

7 For对于 each每个 dataset数据集, you你 have必须 to provide提供:

8 - Dataset数据集 Reasoning推理: reasoning逐步推理并提取以下信息 step逐步-by-step逐步 to extract提取 the以下 following以下 information信息

9 - Dataset数据集 Name名称: the数据集的名称 name名称 of the数据集 dataset数据集. If the如果 dataset数据集 is referred被 generically统称, specify请注明 "generic通用"

10 - Dataset数据集 Created创建: "yes如果是作者创建的数据集，请回答“是”，否则回答“否”" if the数据集 dataset数据集 is created创建 by the作者 authors作者, "no" otherwise否则

11 - Reference参考 name名称: name提出该数据集的论文名称，或作者引用的参考论文名称 of the论文 paper论文 that提出 proposes提出 the数据集 dataset数据集, if the参考 reference参考 paper论文 is cited引用 by the作者 authors作者

12 - Reference参考 Year年份: Year年份 of the提出该数据集的论文年份，或作者引用数据集来源的参考论文年份 paper论文 that提出 proposes提出 the数据集 dataset数据集, if the参考 reference参考 paper论文 is cited引用 by the作者 authors作者 where其中 the数据集 dataset数据集 was被 extracted提取.

13 - Signals信号: List列出 of signals数据集包含的信号 included包含 in the数据集 dataset数据集. Can可以 be one一个 or more更多 from来自 [Text文本, Tables表格, Image图像, Audio音频, Code代码, Time时间 Series序列, Video视频, Charts图表, Equations公式]

14 - Sources来源: List列出 of sources数据集的来源 from来自 which其中 the数据集 dataset数据集 is taken获取. Can可以 be one一个 or more更多 from来自 [News新闻, Social社交 Media媒体/Forums论坛, Company公司 Reports报告, Company公司 Fundamentals基本面 \& Indicators指标, Earnings财报 Calls电话会议, Analyst分析师 Reports报告, University大学 Textbooks教科书, Financial金融 Analyst分析师 Exams考试]

15 - Annotation标注: The数据集的标注方式 way方式 the数据集 dataset数据集 was被 annotated标注. Can可以 be one一个 or more更多 from来自 [Manual人工, Synthetic合成, Spontaneous自发, Human人工-in-the循环-loop循环]

16 - Models模型: a list列出 of models该任务中评估的模型 that被 are评估 evaluated评估 on the任务 task任务:

17 - Models模型 reasoning推理: reasoning逐步推理并提取以下信息 step逐步-by-step逐步 to extract提取 the以下 following以下 information信息

18 - Model模型 Position定位: "main主要" if it is the模型 model模型 on which其中 the主要 main主要 contribution贡献 of the作者 authors作者 is based基于, "comparison比较" if the模型 model模型 was被 used用于 as a comparison比较.

19 - Model模型 Name名称: the模型名称 name名称 of the模型 model模型

20 - Reference参考 Name名称: Name提出该模型的论文名称，或作者引用的参考论文名称 of the论文 paper论文 that提出 proposes提出 the模型 model模型, if the参考 reference参考 paper论文 is cited引用 by the作者 authors作者 where其中 the模型 model模型 was被 extracted提取.

21 - Reference参考 Year年份: Year年份 of the提出该模型的论文年份，或作者引用模型来源的参考论文年份 paper论文 that提出 proposes提出 the模型 model模型, if the参考 reference参考 paper论文 is cited引用 by the作者 authors作者 where其中 the模型 model模型 was被 extracted提取.

22 - Size规模: Can可以 be one一个 of small小型（小于 10 亿参数） (up to 8B parameters参数), medium中型（10 亿至 100 亿参数） (between之间 8B and和 80B parameters参数), and和 large大型（超过 100 亿参数） (more更多 than超过 80B parameters参数).

23 - Parameter参数 Size规模: Number参数数量（十亿） of parameters参数 (in billions十亿) if it is specified如果作者指定了，请填写；否则回答“无” by the作者 authors作者, None无 otherwise否则.

Algorithm 2: Format of the prompt used for extraction of tasks, datasets, models, and their relationships算法 2：用于提取任务、数据集、模型及其关系的提示词格式

B.2 Taxonomy ExtractionB.2 分类法提取

To systematically derive taxonomies for specific entity types, we employed an iterative process, where multiple representative samples of entities were provided to the LLM using the prompt shown in Listing 3. This iterative approach allowed for the identification of recurrent patterns and commonly occurring entity types, ensuring robust and meaningful taxonomic categories.为了系统地推导特定实体类型的分类法，我们采用了迭代过程，通过列表 3 中所示的提示词向大语言模型提供多个具有代表性的实体样本。这种迭代方法有助于识别重复出现的模式和常见的实体类型，从而确保分类类别稳健且具有意义。

⬇

1 You你 will将 be provided提供 with带有 a list列表 of descriptions描述. Each每个 description描述 corresponds对应 to a {ENTITY TYPE} extracted提取 from来自 a Financial金融 NLP paper论文.

2 Your你的 goal目标 is to identify识别 recurring重复 patterns模式 in the列表 list列表 of {ENTITY TYPE}. In particular特别地, you你 should应该 identify识别 types类型 of {ENTITY TYPE} that出现 appear出现 multiple多次 times次数 in the列表 list列表. Do not不要 try尝试 to find找到 types类型 of {ENTITY TYPE} that涵盖 cover涵盖 every每一个 single单个 example示例-only仅 identify识别 those那些 that重复 recur重复 with具有 a certain一定 frequency频率.

Algorithm 3: Format of the prompt used for taxonomy extraction算法 3：用于分类法提取的提示词格式

B.3 Entity ClassificationB.3 实体分类

To classify entities consistently across different categories, we utilized a universal prompt structure, as illustrated in Listing 4. This approach ensures uniformity and accuracy in the classification of entities based on predefined categories.为了在不同类别之间一致地对实体进行分类，我们使用了通用的提示词结构，如列表 4 所示。这种方法确保了基于预定义类别对实体进行分类的统一性和准确性。

⬇

1 You你 will将 be given给定 a description描述 of an entity实体 extracted提取 from来自 a Financial金融 NLP papers论文. The实体 entity实体 is of type类型 {ENTITY TYPE}. Your你的 goal目标 is to classify分类 the实体 entity实体 into进入 one一个 or more更多 of the以下 following以下 categories类别:

3 {LIST OF POSSIBLE CATEGORIES}

Algorithm 4: Format of the prompt used for entity classification算法 4：用于实体分类的提示词格式

Appendix C Additional Material for Analysis附录 C 分析补充材料

⬇

1 You你 will将 receive接收 a JSON array数组 containing包含 every每一个 *motivation动机 paragraph段落* for用于 ONE

2 research研究 paper论文 in NLP-for用于-Finance金融.

4 **Task任务 (for用于 EACH element元素, in the相同 same相同 order顺序)**）**

5 1. Think思考 step逐步-by-step逐步: does该 the段落 paragraph段落 primarily主要

6 - show展示 *enthusiasm热情 / experimentation实验* with带有 LLMs大语言模型, **or**

7 - propose提出 *hybrid混合 / alternative替代* solutions解决方案 (i.e. highlight强调 LLM limitations局限性)?

8 - If no stance立场 towards朝向 LLMs大语言模型 is expressed表达, set设置 Enthusiasm热情 = "other其他"

9 2. Summarise总结 your你的 reasoning推理 in 2-32-3 sentences句子.

Algorithm 5: Prompt used for detection of the stance towards LLMs from motivations算法 5：用于从动机中检测对大语言模型立场的提示词

This section provides supplementary analyses and detailed data supporting the insights discussed in the main body of the paper. It encompasses comprehensive overviews of task trends, evolving methodologies, and changing attitudes within Financial NLP research, presented through tables and figures to offer clear visual representations of temporal developments.本节提供了补充分析和详细数据，以支持论文正文中讨论的见解。它通过表格和图表呈现了任务趋势、不断演变的方法论以及金融自然语言处理（NLP）研究中态度的变化，从而直观地展示了时间维度的发展。

C.1 Temporal Evolution of Financial NLP TasksC.1 金融 NLP 任务的时间演变

Table 11 presents a detailed percentage distribution of Financial NLP tasks across three different timeframes: January 2022–August 2023 (T1), September 2023–July 2024 (T2), and August 2024–April 2025 (T3). This analysis is based on categories derived by prompting Gemini-2.5-Flash to output a taxonomy from samples of task descriptions extracted from the paper, and we used the same model to assign each task to one of the categories.表11展示了三个不同时间段内金融自然语言处理（NLP）任务的详细百分比分布：2022年1月至2023年8月（T1）、2023年9月至2024年7月（T2）以及2024年8月至2025年4月（T3）。此分析基于通过提示 Gemini-2.5-Flash 从论文中提取的任务描述样本生成的分类法，我们使用同一模型将每个任务分配到相应的类别中。

The data illustrates limited innovation in information extraction: the tasks remain the same since the first period, and the percentage of papers practicing them is decreasing (with the exception of claim verification, which is a recent novelty). We can observe that the sharpest decline concerns financial sentiment analysis, which was the most practiced task in the first period and has experienced a significant drop in the last, reaching a percentage that is, by itself, lower than retrieval-enhanced QA. Finally, we can see that the true core of innovation is concentrated around financial QA. Tasks that were already practiced in the first period have attracted increasing attention over time, and new, more complex forms of QA (such as financial consulting and multimodal QA) have become widespread in the latest period.数据显示信息抽取领域的创新有限：相关任务自第一个时期以来保持不变，且采用这些任务的论文比例正在下降（声明验证除外，这属于近期出现的新颖任务）。我们可以观察到，降幅最明显的是金融情感分析，它在第一个时期是应用最广泛的任务，但在最后一个时期出现了显著下降，其占比甚至低于检索增强型问答。最后，我们可以看到创新的核心真正集中在金融问答领域。在第一个时期就已经存在的任务随着时间的推移受到了越来越多的关注，而更复杂的新型问答形式（如金融咨询和多模态问答）在最近的时期变得十分普遍。

Task Category & Name	T1(%)	T2(%)	T3(%)
Information Extraction
Intent Detection	$1.99$	$1.73$	$1.73$
Stance Detection	$1.00$	$1.56$	$0.58$
NER	$7.21$	$8.67$	$5.32$
Relation Extraction	$7.21$	$4.16$	$4.03$
Semantic Annotation	$12.69$	$7.11$	$4.17$
Claim Verification	$1.24$	$1.21$	$3.31$
Event Based Text Annotation	$6.47$	$9.53$	$6.47$
Event Extraction	$3.23$	$2.08$	$0.86$
Argument Mining	$0.75$	$1.56$	$0.58$
Financial Sentiment Analysis
Financial Sentiment Analysis	$15.67$	$15.77$	$9.78$
Financial Emotion Analysis	$0.50$	$0.87$	$0.43$
Stock Market Prediction
Stock Price Change Prediction	$12.19$	$7.63$	$8.20$
Stock Volume Prediction	$0.25$	$0.35$	$0.58$
Stock Return Prediction	$4.73$	$3.47$	$2.59$
QA & Specialized Tasks
Retrieval Enhanced QA	$5.47$	$7.80$	$12.81$
Numerical QA	$5.47$	$7.45$	$9.50$
Conversational QA	$2.24$	$3.12$	$3.60$
Long Document QA	$3.23$	$4.51$	$7.34$
Causal QA	$1.99$	$1.04$	$2.30$
Tabular QA	$2.24$	$3.99$	$4.46$
Financial Consulting QA	$2.24$	$4.16$	$6.04$
Temporal Reasoning QA	$0.50$	$0.52$	$0.86$
Financial Terminology Explanation	$0.75$	$1.21$	$2.73$
Multimodal QA With Images	$0.75$	$0.52$	$1.73$

Table 11: Percentage distribution of Financial NLP tasks across different timeframes (T1: Jan 2022–Aug 2023, T2: Sep 2023–Jul 2024, T3: Aug 2024–Apr 2025).表11：不同时间段内金融NLP任务的百分比分布（T1：2022年1月至2023年8月；T2：2023年9月至2024年7月；T3：2024年8月至2025年4月）。

C.2 Trends in LLM Risks and LimitationsC.2 大语言模型（LLM）的风险与局限性趋势

Table 6 categorizes the evolving risks associated with large language models (LLMs) in financial applications, as identified in the reviewed corpus. Recall that these categories are obtained by prompting Gemini-2.5-Flash to output a taxonomy from samples of safety limitations extracted from the paper, and we used the same model to assign the limitations to one of the categories.表6对所审查语料库中确定的金融应用中大语言模型（LLM）的相关风险演变进行了分类。请注意，这些类别是通过提示 Gemini-2.5-Flash 从论文中提取的安全局限性样本生成的分类法，我们使用同一模型将这些局限性分配到相应的类别中。

Notably, safety concerns have sharply increased from the first to the second period, and are still sharply rising today. However, we can distinguish groups with different trends: concerns that were already prevalent, such as inaccuracy and the inability to adapt to new data, have become even more pronounced. At the same time, limitations rarely highlighted by paper authors before have emerged. These include the impact that well-known limitations of LLMs, such as bias, can have in the financial sector, as well as privacy and misuse issues.值得注意的是，安全顾虑从第一个时期到第二个时期急剧增加，且至今仍在迅速上升。然而，我们可以区分出具有不同趋势的群体：一些早已普遍存在的顾虑，如不准确性和无法适应新数据，变得更加突出。与此同时，一些此前很少被论文作者强调的局限性也开始显现。这些包括大语言模型的已知局限性（如偏见）在金融领域可能产生的影响，以及隐私和滥用问题。

C.3 Evolution of Retrieval-Augmented Generation (RAG) ApproachesC.3 检索增强生成（RAG）方法的演变

Table 7 summarizes the distribution of different retrieval-augmented generation (RAG) configurations over the studied timeframes. It should be noted that we are using the distinguishing factors for RAG approaches as proposed in the survey by Gao2023RetrievalAugmentedGF.表7总结了在所研究的时间段内不同检索增强生成（RAG）配置的分布情况。需要说明的是，我们使用了Gao2023RetrievalAugmentedGF综述中提出的RAG方法区分因素。

We can observe a clear trend that follows the expansion of data sources within the sector: the documents from which retrieval occurs are becoming increasingly longer, and the landscape of modalities is growing more diverse. In particular, there has been an increased use of structured data sources such as tables and databases. As RAG tasks become more complex and refined, the model is asked to retrieve more and more information, which is then incorporated into the prompt. This testifies to the progressive transformation of information extraction tasks into retrieval tasks.我们可以观察到一个明显的趋势，即随着该领域数据源的扩展：检索所依据的文档正变得越来越长，模态格局也日益多样化。特别是结构化数据源（如表格和数据库）的使用有所增加。随着RAG任务变得更加复杂和精细，模型被要求检索越来越多的信息，并将其整合到提示词中。这证明了信息抽取任务正在逐步向检索任务转型。

C.4 Shifts in Data Reuse PracticesC.4 数据重用实践的转变

Figure 4 illustrates the changing proportions of datasets explicitly referencing prior literature over time, relative to all datasets used in published studies. The trend suggests a shift in data reuse practices: following an initial phase characterized by widespread creation of new datasets, researchers increasingly began to adopt and build upon existing resources.图4展示了随着时间推移，在已发表研究中明确引用先前文献的数据集比例变化。这一趋势表明了数据重用实践的转变：在以广泛创建新数据集为特征的初期阶段之后，研究人员越来越多地开始采用并基于现有资源进行研究。

C.5 Prompt Engineering TechniquesC.5 提示工程技术

Table 10 highlights the evolving landscape of prompt engineering techniques over three distinct timeframes. The categories are taken from schulhoff2025promptreportsystematicsurvey. We can see an explosion in the use of prompt engineering techniques in the period following the advent of LLMs. While zero-shot remains the most widespread and dominant technique, few-shot has seen a recent decline in usage, being increasingly replaced by newer techniques to improve output accuracy. Chain-of-thought prompting enhances the model’s reasoning, improving performance on complex tasks and supporting interpretability. RAG adds contextual information to reduce inaccuracies and provide updated knowledge, while techniques such as self-criticism use another or the same LLM to check the output. These techniques share a common goal: they add complexity and sophistication to the prompting process to mitigate model shortcomings.表10强调了在三个不同时间段内提示工程技术不断演变的格局。这些类别取自 schulhoff2025promptreportsystematicsurvey。我们可以看到，在大语言模型出现后的时期，提示工程技术的使用出现了爆发式增长。虽然零样本（zero-shot）仍然是最广泛和占主导地位的技术，但少样本（few-shot）的使用近期有所下降，正逐渐被旨在提高输出准确性的新技术所取代。思维链（Chain-of-thought）提示增强了模型的推理能力，提升了复杂任务的表现并支持可解释性。RAG通过增加上下文信息来减少不准确性并提供更新的知识，而自我批评等技术则利用另一个或同一个大语言模型来检查输出。这些技术都有一个共同目标：增加提示过程的复杂性和精密程度，以缓解模型的短板。

C.6 Stance Detection towards LLMs in Papers MotivationsC.6 论文动机中对大语言模型的立场检测

In order to obtain a quantitative insight into how the stance towards LLMs has changed over time, we applied Gemini 2.5 Flash zero-shot on the papers’ motivation, using the prompt in Listing 5.为了定量了解对大语言模型的立场如何随时间变化，我们使用列表5中的提示词，对论文动机应用了 Gemini 2.5 Flash 零样本（zero-shot）分析。

Figure 15 provides a quantitative depiction of evolving attitudes toward LLMs over time, as reflected in the motivations stated in their work. We observe that papers framing LLMs as a solution to the limitations of earlier models have steadily declined, giving way to a growing number of studies motivated by the goal of mitigating the limitations of the LLMs themselves.图15定量描绘了对大语言模型的态度随时间演变的情况，正如其工作动机中所反映的那样。我们观察到，将大语言模型视为早期模型局限性解决方案的论文数量稳步下降，取而代之的是越来越多以缓解大语言模型自身局限性为动机的研究。

MetaGraph: A Large-Scale Meta-Analysis of GenAI in Financial NLP (2022–2025)MetaGraph：金融自然语言处理（NLP）领域生成式人工智能（GenAI）的大规模元分析（2022–2025）