SENTiVENT v1.1 Economic Events and Implicit Sentiment

SENTiVENT is a high-quality human-annotated financial news dataset for economic event extraction, implicit economic sentiment analysis, aspect-based sentiment analysis, and joint event-sentiment evaluation. Here, implicit sentiment means investor-relevant polarity inferred from factual news through common sense and world knowledge, not only explicit opinion wording. The corpus contains 170,398 tokens across 288 fully annotated English business-news articles about 30 S&P 500 companies. It includes 6,245 event mentions, 13,780 event arguments, 3,669 sentiment expressions, and 4,429 sentiment-target tuples. Articles were collected from diverse online financial-news sources and selected for sector diversity, temporal spread, topical diversity, and suitability for linguistic annotation. 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.

What SENTiVENT Provides

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

• Human-annotated economic event, argument, sentiment-expression, sentiment-target, and event-sentiment labels.
• A joint document-level target that links event triggers, participants, fillers, sentiment expressions, entity/aspect/event targets, and polarity.
• Consistent train/dev/test splits across configs matching the pre-existing SENTiVENT papers for benchmark comparison.
• Flattened table configs and research-standard document, sentence, token, and token-span task views for corpus inspection, baselines, prompting, and evaluation.

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. It is valuable because much financial news is written objectively, and challenging because the evidence is lexically diverse, event-implied, and target-dependent.

Uses

SENTiVENT is intended for research and evaluation in information extraction, implicit economic sentiment analysis, and structured-output modeling over 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 broader financial-advice settings.

How To Use

from datasets import load_dataset

unified = load_dataset("GillesJacobs/sentivent")
unified_continuous = load_dataset("GillesJacobs/sentivent", "sentivent_unified_document_continuous")
unified_tokens = load_dataset("GillesJacobs/sentivent", "task_sentivent_unified_token")
unified_sentences = load_dataset("GillesJacobs/sentivent", "task_sentivent_unified_sentence")
documents = load_dataset("GillesJacobs/sentivent", "documents")
events = load_dataset("GillesJacobs/sentivent", "events")

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.
• 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:v1",
  "task": "sentivent_unified_event_sentiment",
  "document_id": "doc-1",
  "title": "Acme Raises 2017 Forecast",
  "split": "train",
  "text": "Acme raised its 2017 revenue forecast after strong demand.",
  "tokens": ["Acme", "raised", "its", "2017", "revenue", "forecast", "..."],
  "annotations": [
    {
      "id": "event-1",
      "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": "arg-1",
          "kind": "participant",
          "role": "Company",
          "text": "Acme",
          "span": {
            "discontinuous": false,
            "tokens": [0],
            "chars": [[0, 4]],
            "text": ["Acme"]
          }
        }
      ],
      "fillers": [
        {
          "id": "arg-2",
          "kind": "filler",
          "role": "TIME",
          "text": "2017",
          "span": {
            "discontinuous": false,
            "tokens": [3],
            "chars": [[16, 20]],
            "text": ["2017"]
          }
        }
      ]
    },
    {
      "id": "sentiment-1",
      "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": "target-1",
          "kind": "sentiment_target",
          "target_kind": "participant",
          "text": "Acme",
          "span": {
            "discontinuous": false,
            "tokens": [0],
            "chars": [[0, 4]],
            "text": ["Acme"]
          }
        }
      ]
    }
  ],
  "relations": [
    {
      "id": "event-sentiment-1",
      "kind": "event_sentiment",
      "from": "sentiment-1",
      "to": "event-1",
      "polarity": "positive"
    }
  ]
}

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	Why
sentivent_unified_document	Canonical joint benchmark for ML and LLM evaluation	Use for the main task: read the full article and extract economic events, implicit economic sentiment, event-targeted sentiment links, and span hierarchy together. Format: {text, tokens, annotations, relations}.
sentivent_unified_document_continuous	Systems that cannot represent discontinuous spans	Keeps the joint document target but fills annotation gaps into one continuous extent, making it easier for extractive QA, UI, or span tooling that rejects multi-part spans. Format: {text, annotations with continuous spans, [...]}.
task_sentivent_unified_token	Token-span baselines and sequence-model evaluation	Use when a model consumes sentence-token sequences and predicts trigger, argument, sentiment-expression, target, and link spans with token offsets. Format: {tokens, token_offsets, gold_json.events, gold_json.sentiments, [...]}.
task_sentivent_unified_sentence	Sentence-scoped structured-output prompting	Use for compact LLM evals, debugging prompts, and low-context comparisons where each example carries its instruction, input text, and structured gold target. Format: {instruction, input, gold_json.events, gold_json.sentiments, [...]}.
documents	Corpus browsing, joins, and metadata checks	Use to inspect article text, split membership, annotation counts, and ids before joining into flatter annotation tables. Format: {id, text, event_ids, sentiment_ids, [...]}.
events	Event-only extraction and label analysis	Use for event schema exploration, trigger/type/subtype statistics, argument joins, and event extraction baselines that do not need the sentiment layer. Format: {event_id, event_type, span_set, argument_ids, [...]}.

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.
• task_sentivent_unified_document, task_sentivent_unified_sentence, and task_sentivent_unified_token provide secondary task-row views with structured gold JSON.
• task_ere_document, task_ere_sentence, and task_ere_token expose the event extraction labels.
• task_iabsa_document, task_iabsa_sentence, and task_iabsa_token expose implicit aspect-based economic sentiment labels.
• Canonical configs such as documents, sentences, tokens, events, event_arguments, sentiment, sentiment_targets, and event_sentiment expose the annotation layers as tables.

Config	Train	Validation	Test	Format
sentivent_unified_document	228	30	30	Format: nested article JSON {text, tokens, annotations, relations}
sentivent_unified_document_continuous	228	30	30	Format: nested article JSON with continuous spans {text, annotations, [...]}
documents	228	30	30	Format: article table {id, text, event_ids, sentiment_ids, [...]}
sentences	5475	681	727	Format: sentence table {sentence_id, span, text, token_ids, [...]}
tokens	135317	17955	17126	Format: token table {token_id, token_index, span, text, [...]}
participants	11188	1459	2098	Format: mention table {participant_id, kind, role, span_set, [...]}
events	4636	625	984	Format: event table {event_id, event_type, span_set, argument_ids, [...]}
event_arguments	10135	1346	2299	Format: edge table {event_id, role, target_id, target_kind, [...]}
sentiment	2954	387	328	Format: sentiment table {sentiment_id, polarity, span_set, target_ids, [...]}
sentiment_targets	3531	483	415	Format: target edge table {sentiment_id, target_id, target_kind, [...]}
event_sentiment	71	11	20	Format: event-sentiment edge table {sentiment_id, event_id, polarity, [...]}
relations	1710	200	397	Format: relation table {relation_id, relation_type, source_id, target_id, [...]}
task_ere_document	228	30	30	Format: prompt JSON {instruction, input, gold_json.events}
task_ere_sentence	5475	681	727	Format: prompt JSON {instruction, input, gold_json.events}
task_ere_token	5475	681	727	Format: token prompt {tokens, token_offsets, gold_json.events}
task_iabsa_document	228	30	30	Format: prompt JSON {instruction, input, gold_json.sentiments}
task_iabsa_sentence	5475	681	727	Format: prompt JSON {instruction, input, gold_json.sentiments}
task_iabsa_token	5475	681	727	Format: token prompt {tokens, token_offsets, gold_json.sentiments}
task_sentivent_unified_document	228	30	30	Format: prompt JSON {instruction, input, gold_json.events, gold_json.sentiments, [...]}
task_sentivent_unified_sentence	5475	681	727	Format: prompt JSON {instruction, input, gold_json.events, gold_json.sentiments, [...]}
task_sentivent_unified_token	5475	681	727	Format: token prompt {tokens, token_offsets, gold_json.events, gold_json.sentiments, [...]}

Splits

The public configs preserve the original SENTiVENT v1 split policy v1: 228 train documents, 30 dev documents, and 30 test documents. Hugging Face exposes original dev rows as validation; every row also keeps original_split.

References And Citation

Please cite the relevant SENTiVENT paper for your task:

For event extraction tasks:

@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}
}

For sentiment analysis tasks:

@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}
}

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}
}

License

SENTiVENT is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). Non-commercial use is permitted under the license terms. When using the dataset, cite the relevant papers listed above.

Release Provenance

• Release tag: sentivent-guidelines-v1.1-20260702-g1b099a5d8fe9.
• Annotation guidelines version: v1.1.
• Build timestamp: 2026-07-02T23:27:30.050473+00:00.
• Builder git commit: 1b099a5d8fe9.
• Machine-readable release tag metadata is in metadata/release_tags.json.