SENTiVENT Bench: Economic Events and Implicit Sentiment in Financial News

SENTiVENT is an expert human-annotated financial news dataset for research on what happens in company news and what sentiment those events imply. It supports financial event extraction, implicit financial sentiment analysis, aspect-based sentiment analysis (ABSA), and structured-output evaluation for ML systems and AI agents. Here, implicit sentiment means investor-relevant polarity inferred from factual news through common sense and world knowledge, not only explicit opinion wording. The resource is designed to be representative of the business-news genre: article selection avoids over-representing one company, sector, topic, or news event, and excludes robo-written or templated articles.

How To Use

For most users and AI evaluation agents, start with sentivent_unified_document: text is the input, and annotations plus relations are the gold structured target.

import evaluate
from datasets import load_dataset

test = load_dataset("GillesJacobs/sentivent", "sentivent_unified_document", split="test")

# Runnable empty baseline; replace this list with model predictions.
predictions = [
    {
        "id": row["id"],
        "tokens": row["tokens"],
        "annotations": [],
        "relations": {"event_sentiment": [], "event_coreference": []},
    }
    for row in test
]

metric = evaluate.load("GillesJacobs/sentivent_metric")
result = metric.compute(predictions=predictions, references=list(test))
summary = {
    "sentivent_score": result["sentivent_score"],
    "sentivent_macro_score": result["sentivent_macro_score"],
    "sentivent_relaxed_score": result["sentivent_relaxed_score"],
    "sentivent_relaxed_macro_score": result["sentivent_relaxed_macro_score"],
}
print(summary)

# {
#   "sentivent_score": 0.0,
#   "sentivent_macro_score": 0.0,
#   "sentivent_relaxed_score": 0.0,
#   "sentivent_relaxed_macro_score": 0.0
# }

Score version 0.4.0 defaults to sentivent_score, the strict mean of five extraction component F1 values weighted by fixed train+validation support. By default, the scorer also returns the unweighted macro companion and both relaxed companions. Set include_factuality=True to expand each enabled profile to twelve components by adding seven aligned factuality/context macro-F1 values.

Publication computes and freezes the component supports from the exact staged train+validation dataset; evaluated slices only renormalize those fixed supports over active components. This intentionally lets less-attested subtasks weigh less, while the macro companion preserves the equal-component view. See the metric Space, normative specification, and replication appendix for exact formulas, alternatives, and primary-source literature rationale.

Load common model-family variants directly when you need framework-specific rows rather than the canonical document target:

from datasets import load_dataset

dataset_id = "GillesJacobs/sentivent"

sentence_rows = load_dataset(dataset_id, "sentivent_unified_sentence")  # Sentence-window structured IE
continuous_docs = load_dataset(dataset_id, "sentivent_unified_document_continuous")  # Continuous-span extractors
bio_token_rows = load_dataset(dataset_id, "sentivent_unified_bio_sentence")  # BIO token classifiers
textee_event_rows = load_dataset(dataset_id, "ere_textee_sentence_e2e")  # TextEE event extraction
oneie_joint_rows = load_dataset(dataset_id, "sentivent_unified_oneie_sentence")  # OneIE-style event-sentiment rows
uie_spot_asoc_rows = load_dataset(dataset_id, "sentivent_unified_uie_sentence")  # UIE / Spot-Asoc generation
absa_quad_rows = load_dataset(dataset_id, "iabsa_quad_sentence")  # ABSA quad extraction

What SENTiVENT Provides

SENTiVENT is intended for reproducible evaluation of information extraction and implicit sentiment in financial news.

• Expert annotations for economic events, event arguments, sentiment expressions, sentiment targets, and event-implied sentiment.
• A joint document-level target that links what happened, who or what it involved, and whether the news implies positive or negative sentiment.
• Stable train/dev/test splits matching the SENTiVENT papers for benchmark comparison.
• Ready-to-load document, sentence, token-label, ABSA, TextEE, UIE, OneIE, OmniEvent, DyGIE++, and LLM-evaluation views.

Implicit sentiment captures cases where factual business-news content changes a reader-investor's attitude toward a company, asset, person, product, or event because readers know what is desirable or undesirable in financial markets, even without words like great or bad. Implicit sentiment in financial news is challenging because articles are written objectively and the evidence is lexically diverse and requires common-sense world knowledge to resolve.

Uses

SENTiVENT is intended for researchers and ML practitioners evaluating financial NLP systems, information extraction models, sentiment models, and LLM agents on human-annotated financial news. The corpus covers English company-specific business-news articles about 30 S&P 500 companies from June 2016-May 2017, so results may not transfer directly to other languages, later market periods, informal text, non-US companies, or financial-advice settings.

The anonymized raw WebAnno/UIMA XMI source export is included under raw_webanno_export/ for users who need the original annotation files. It is a source artifact, not a recommended modeling or evaluation config.

Main Task

• The primary public config is sentivent_unified_document, the compact view for joint economic event extraction, implicit economic sentiment, and event-targeted sentiment in one structured target.
• sentivent_unified_document uses the full article text as input and emits event triggers, participants, fillers, sentiment expressions, non-event targets, and event-sentiment links with token, character, and text spans.
• sentivent_unified_sentence mirrors the same schema one source sentence at a time for smaller models, prompt-size constrained evaluation, and sentence-window pipelines; cross-sentence arguments, targets, and relation endpoints are omitted.
• Use sentivent_unified_document_continuous only for tools that cannot consume discontinuous spans; it fills span gaps and is therefore a derived lossy evaluation view.

Joint Document Row Shape

The default config is an article-level JSON object. text is the model input; annotations and relations are the structured target. This compact example shows the main nested shapes; long token and annotation lists are shortened.

{
  "id": "doc-1:sentivent_unified_document",
  "split": "train",
  "text": "Acme raised its 2017 revenue forecast after strong demand.",
  "tokens": ["Acme", "raised", "its", "2017", "revenue", "forecast", "..."],
  "annotations": [
    {
      "id": "ev-aal00-01",
      "kind": "event",
      "text": "raised",
      "span": {
        "discontinuous": false,
        "tokens": [1],
        "chars": [[5, 11]],
        "text": ["raised"]
      },
      "event_type": "Revenue",
      "event_subtype": "Increase_Revenue",
      "modality": false,
      "negated": false,
      "realis": "actual",
      "polarity": "positive",
      "scoped_polarity": "positive",
      "participants": [
        {
          "id": "part-aal00-01",
          "argument_id": "arg-aal00-01",
          "kind": "participant",
          "role": "Company",
          "text": "Acme",
          "span": {
            "discontinuous": false,
            "tokens": [0],
            "chars": [[0, 4]],
            "text": ["Acme"]
          }
        }
      ],
      "fillers": [
        {
          "id": "fill-aal00-01",
          "kind": "filler",
          "role": "TIME",
          "text": "2017",
          "span": {
            "discontinuous": false,
            "tokens": [3],
            "chars": [[16, 20]],
            "text": ["2017"]
          }
        }
      ]
    },
    {
      "id": "se-aal00-01",
      "kind": "sentiment",
      "text": "strong",
      "span": {
        "discontinuous": false,
        "tokens": [7],
        "chars": [[44, 50]],
        "text": ["strong"]
      },
      "polarity": "positive",
      "scoped_polarity": "positive",
      "negated": false,
      "uncertain": false,
      "targets": [
        {
          "id": "stgt-aal00-01",
          "sentiment_target_id": "stgt-link-aal00-01",
          "kind": "sentiment_target",
          "target_kind": "participant",
          "text": "Acme",
          "span": {
            "discontinuous": false,
            "tokens": [0],
            "chars": [[0, 4]],
            "text": ["Acme"]
          }
        }
      ]
    }
  ],
  "relations": {
    "event_sentiment": [
      {
        "id": "es-aal00-01",
        "from": "se-aal00-01",
        "to": "ev-aal00-01",
        "polarity": "positive"
      }
    ],
    "event_coreference": []
  }
}

Corpus And Annotations

SENTiVENT contains company-specific English financial news about 30 S&P 500 companies from June 2016-May 2017. Articles were selected from diverse online business-news sources for sector balance, temporal spread, topical diversity, source quality, and suitability for detailed linguistic annotation; duplicate, low-relevance, templated, and non-company-specific items were removed.

The annotations cover ACE/ERE-style economic events and implicit economic sentiment in one joint layer: event triggers, event types/subtypes, participants, fillers, event coreference, event attributes, sentiment expressions, entity/aspect/event targets, polarity, and event-implied sentiment.

The holdout test set is a gold-standard expert-adjudicated reference set from a three-annotator agreement study. Reported agreement is almost-perfect for event main type (Fleiss kappa=0.877, Krippendorff alpha=0.874), substantial for event full type/subtype (kappa=0.813, alpha=0.809), and substantial for directly annotated sentiment polarity (kappa=0.778, alpha=0.782).

Item	Total	Train/Dev	Test
documents	288	228 train / 30 dev	30
sentences	6,883	5,475 train / 681 dev	727
tokens	170,398	135,317 train / 17,955 dev	17,126
event mentions	6,245	4,636 train / 625 dev	984
event arguments	13,780	10,135 train / 1,346 dev	2,299
participant arguments	10,581	7,751 train / 1,048 dev	1,782
filler arguments	3,199	2,384 train / 298 dev	517
event coreference links	1,364	1,021 train / 125 dev	218
event type labels	18	-	-
event subtype labels	42	-	-
sentiment expressions	3,669	2,954 train / 387 dev	328
sentiment-target tuples	4,429	3,531 train / 483 dev	415
event-sentiment links	102	71 train / 11 dev	20

Recommended Configs

Config	Best for	Task and references
sentivent_unified_document	Canonical joint benchmark for ML and LLM evaluation	Full-article structured IE: extract economic event triggers, types/subtypes, participants, fillers, implicit sentiment expressions/targets, and event-sentiment links. See the event extraction paper, implicit sentiment paper, and SENTiVENT thesis.
sentivent_unified_document_continuous	Systems that cannot represent discontinuous spans	Same joint document task for span-limited extractive tools; discontinuous spans are filled to one covering extent, so this is a lossy convenience view. See the event extraction paper and implicit sentiment paper.
sentivent_unified_sentence	Smaller-context joint sentence evaluation	One row per source sentence with the same event/sentiment schema as the document target. Cross-sentence arguments, targets, and relation endpoints are cut so each row is self-contained. See the event extraction paper, implicit sentiment paper, and SENTiVENT thesis.
sentivent_unified_sentence_continuous	Sentence-window tools that cannot represent discontinuous spans	Same sentence-level joint task, after row-local filtering, with retained discontinuous spans filled to one covering extent. This is a lossy convenience view for span-limited tools. See the event extraction paper and implicit sentiment paper.

Dataset Configs

• sentivent_unified_document is the default and recommended config for the full joint extraction task over article text.
• sentivent_unified_document_continuous is a derived evaluation variant that fills discontinuous spans to continuous ranges.
• sentivent_unified_sentence is the smaller-context sentence-window version of the same joint task; cross-sentence endpoints are cut so every row is self-contained.
• sentivent_unified_sentence_continuous applies the same sentence-window filtering and fills retained discontinuous spans to continuous ranges.
• sentivent_unified_document_with_canonical_referents is an opt-in variant that keeps document-level canonical-reference evidence inline.
• sentivent_unified_sentence_with_canonical_referents is the row-local canonical-reference audit variant for sentence windows.
• Interoperability configs expose complete task views for common event extraction, implicit ABSA, UIE, OneIE, OmniEvent, DyGIE++, TextEE, and BIO pipelines.

Config	Description	train / validation / test
sentivent_unified_document	Primary joint document-level benchmark: extract economic events, participants, fillers, implicit sentiment expressions, entity/aspect/event targets, polarity, and event-sentiment links. See the event extraction paper, implicit sentiment paper, and SENTiVENT thesis.	228 / 30 / 30
sentivent_unified_document_continuous	Lossy continuous-span variant of the joint document extraction task for tools that cannot represent discontinuous spans. See the event extraction paper and implicit sentiment paper.	228 / 30 / 30
sentivent_unified_document_with_canonical_referents	Opt-in joint document config with the same resolved spans as the default view plus inline resolved_from evidence for canonical referents, including standalone canonical-reference annotation items when needed. See the event extraction paper, implicit sentiment paper, and SENTiVENT thesis.	228 / 30 / 30
sentivent_unified_sentence	Sentence-window version of the joint event and implicit-sentiment task: one row per source sentence, with cross-sentence arguments, targets, and relation endpoints omitted so each row is self-contained. See the event extraction paper, implicit sentiment paper, and SENTiVENT thesis.	5475 / 681 / 727
sentivent_unified_sentence_continuous	Lossy continuous-span variant of the sentence-window joint task. It keeps the same sentence filtering, then fills retained discontinuous spans without crossing sentence boundaries. See the event extraction paper and implicit sentiment paper.	5475 / 681 / 727
sentivent_unified_sentence_with_canonical_referents	Sentence-window audit config with inline resolved_from evidence only when both the weak mention and canonical referent are inside the same sentence row. Cross-sentence referents are not imported. See the SENTiVENT thesis.	5475 / 681 / 727
sentivent_unified_bio_sentence	Joint event and sentiment sequence labeling: one BIO tag per token, with ner_tags for Hugging Face token-classification code and sidecars for overlaps. See HF token classification, the event extraction paper, and the implicit sentiment paper.	5475 / 681 / 727
ere_textee_sentence_e2e	TextEE-style sentence/window rows for event detection plus argument extraction; sentence configs follow the processed JSONL keys used by TextEE. See TextEE and the event extraction paper.	5475 / 681 / 727
ere_textee_sentence_eae	TextEE-style sentence/window rows for event argument extraction; sentence configs follow the processed JSONL keys used by TextEE. See TextEE and the event extraction paper.	5475 / 681 / 727
ere_textee_sentence_ed	TextEE-style sentence/window rows for event trigger detection/classification; sentence configs follow the processed JSONL keys used by TextEE. See TextEE and the event extraction paper.	5475 / 681 / 727
iabsa_asqp_sentence	ASQP-style aspect sentiment quad prediction as delimiter-line rows with character spans and 0/1/2 polarity ids. See ASQP and the implicit sentiment paper.	5475 / 681 / 727
iabsa_aste_sentence	ASTE-style aspect sentiment triplet extraction as delimiter-line rows and companion .txt split files. See ASTE and the implicit sentiment paper.	5475 / 681 / 727
iabsa_quad_sentence	Implicit economic ABSA quad extraction: target, aspect category, opinion expression, and polarity. See ASQP and the implicit sentiment paper.	5475 / 681 / 727
ere_dygiepp_document	Upstream DyGIE++ document-level event extraction rows with document-token indices, NER spans, and event argument records. See DyGIE++ and the event extraction paper.	228 / 30 / 30
sentivent_unified_uie_sentence	UIE text-to-spot-asoc sentence rows for joint event and sentiment extraction as UIE spot-association targets. See UIE, the event extraction paper, and the implicit sentiment paper.	5475 / 681 / 727
ere_omnievent_sentence	OmniEvent unified event/entity rows; intended for legacy parity or framework adapters. See OmniEvent and the event extraction paper.	5474 / 681 / 727
sentivent_unified_oneie_sentence	OneIE/TextEE-style event rows with SENTiVENT sentiment-target relation extensions; intended for legacy parity or framework adapters. See OneIE, TextEE, and the event extraction paper.	5474 / 681 / 727
ere_oneie_sentence	OneIE/TextEE-style event rows; intended for legacy parity or framework adapters. See OneIE, TextEE, and the event extraction paper.	5474 / 681 / 727

Splits

The public configs preserve the original SENTiVENT v1 paper split membership v1: 228 train documents, 30 dev documents, and 30 test documents. Hugging Face publishes the source dev split as validation; document membership is unchanged.

Citation and License

If you use SENTiVENT, please cite the dataset itself:

@dataset{jacobs2020sentivent,
  title={SENTiVENT Bench: Economic Events and Implicit Sentiment in Financial News},
  author={Jacobs, Gilles},
  publisher={OSF},
  year={2020},
  version={1.1},
  doi={10.17605/OSF.IO/8JEC2},
  url={https://doi.org/10.17605/OSF.IO/8JEC2}
}

For event extraction or annotation methodology, also cite:

@article{jacobs2022sentiventenablingeconomicevents,
  title={SENTiVENT: enabling supervised information extraction of company-specific events in economic and financial news},
  author={Jacobs, Gilles and Hoste, Veronique},
  journal={Language Resources and Evaluation},
  volume={56},
  number={1},
  pages={225--257},
  year={2022},
  doi={10.1007/s10579-021-09562-4},
  url={https://doi.org/10.1007/s10579-021-09562-4}
}

When using the sentiment annotations or sentiment configurations, also cite:

@article{jacobs2021finegrainedimplicitsentiment,
  title={Fine-Grained Implicit Sentiment in Financial News: Uncovering Hidden Bulls and Bears},
  author={Jacobs, Gilles and Hoste, Veronique},
  journal={Electronics},
  volume={10},
  number={20},
  pages={2554},
  year={2021},
  doi={10.3390/electronics10202554},
  url={https://doi.org/10.3390/electronics10202554}
}

For annotation guidelines and more details:

@phdthesis{jacobs2021dissertation,
  title={Extracting Fine-Grained Events and Sentiment from Economic News},
  author={Jacobs, Gilles},
  school={Ghent University},
  year={2021},
  url={https://biblio.ugent.be/publication/8728891}
}

The SENTiVENT dataset, annotations, schemas, release metadata, dataset card, and documentation are licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0); see LICENSE. Executable integration adapters and configuration under integrations/ are licensed under the Apache License 2.0. See the license scope notice for the complete file-level boundary and third-party exclusions. The citation requests above are not additional license conditions.

Release Provenance

• Release tag: sentivent-guidelines-v1.1-20260714-g7b759a38329c.
• Annotation guidelines version: v1.1.
• Build timestamp: 2026-07-14T23:04:09.400815+00:00.
• Builder git commit: 7b759a38329c.
• Machine-readable release tag metadata is in metadata/release_tags.json.