Annotation Guidelines

How causal relations are annotated in the Causal Semantics framework

Overview

This page documents the annotation schema and guidelines used to produce the training data for C-BERT and the underlying (C, E, I) tuple representation. If you are using C-BERT or working with annotated data from this project, this page serves as the reference for how annotations are structured, what decisions were made, and how edge cases are handled.

The annotation was developed over three iterations (2024–2025) using the INCEpTION platform, applied to 4,753 sentences from German environmental discourse (1990–2022). The resulting corpus contains 2,391 manually annotated causal relations.

Annotation Schema

The schema consists of two layers: span annotations that mark causal components in text, and relation annotations that link them into structured causal relations.

Spans

Each token or token sequence in a sentence receives at most one span label. There are two span categories:

Role marks the primary causal components:

• Indicator — the lexical marker that projects the causal relation (e.g. verursachen, ist Ursache, weil, stoppen)
• Entity — a cause or effect entity involved in the relation (e.g. Klimawandel, Pestizide, Artensterben)

Coefficient captures semantic modifiers:

Coefficient	Function	Examples
Negation	Negates or inverts the causal relation	nicht, kein, Verlust, Schwund
Temporality	Temporal framing	seit, bereits, künftig, Jahre
Spatiality	Spatial framing	global, lokal, weltweit, Deutschland
Uncertainty	Modality and hedging	möglicherweise, könnte, vermutlich
Quality	Qualitative modification	industrielle, parasitischer, schwefelhaltiger
Quantity	Quantitative modification	großes, fünftausend, zwölf
Representation	Evidentiality marker	laut, sagt, sei
Representation Entity	Source/provenance	Studie, Bericht, Forscher

Relations

Annotated spans are linked by directed relations:

• Cause: Indicator → Entity (the entity fills the Cause role)
• Effect: Indicator → Entity (the entity fills the Effect role)
• Constraint: Entity or Indicator → Coefficient (the coefficient modifies its governor)

Worked Example

Consider the sentence:

Klimawandel und intensive Landwirtschaft verstärken möglicherweise das weltweite Artensterben.

(Climate change and intensive agriculture possibly intensify global species extinction.)

Span	Type	Relation
Klimawandel	Entity	verstärken → Cause
Landwirtschaft	Entity	verstärken → Cause
verstärken	Indicator	—
möglicherweise	Uncertainty	verstärken → Constraint
weltweite	Spatiality	Artensterben → Constraint
Artensterben	Entity	verstärken → Effect

Key points illustrated here: coordinated entities (Klimawandel und Landwirtschaft) are annotated separately with parallel Cause relations; the attributive adjective intensive is extracted as a Quality coefficient (token minimization principle); the sentence adverb möglicherweise is linked to the indicator since it modifies the causal proposition.

Annotation Principles

Four principles guide annotation decisions in ambiguous cases.

Minimal Principle

Only explicitly marked causal relations are annotated — no inference. When a lexically specific marker (e.g. verursacht) and a functional marker (e.g. durch) co-occur, only the lexically richer element is annotated as indicator. Functional prepositions and connectors are annotated as indicators only when no richer causal lexeme is present.

For light verb constructions, only the semantically loaded element is annotated (e.g. hat etwas mit X zu tun → Indicator: tun).

Token Minimization

Entities are reduced to their head token. Attributive modifiers are extracted as separate coefficients:

• industrielle Landwirtschaft → Entity: Landwirtschaft, Coefficient: industrielle (Quality)
• Einsatz von Pestiziden → Entity: Pestiziden
• Exception: named entities (Europäische Union) and fixed multi-word expressions (saurer Regen) are annotated as single units.

This prevents proliferation of marginal entity variants and facilitates downstream aggregation.

Syntactic Proximity

When multiple potential entities compete for a role, the syntactically closest entity to the indicator is preferred, unless semantic considerations override this.

Coefficient Conservatism

Coefficients (other than Representation) are only annotated when they stand in a direct syntactic dependency relation with an indicator or entity.

Indicators

Indicators are the lexical or syntactic markers that project causal relations and establish (C, E, I) tuples. They fulfill two functions: projecting causal roles onto syntactic positions in their co-text, and encoding inherent information about polarity and salience.

The annotation corpus contains 642 distinct indicator forms, grouped into 192 indicator families by morphological and semantic criteria. A family subsumes all realizations of a shared lexical core (e.g. the Ursache family includes verursachen, ist Ursache, Ursache sein, Teilursache sein).

Top 10 Indicator Families

Family	Forms	Instances		Family	Forms	Instances
Ursache	21	167		Beitrag	10	70
Verantwortung	11	122		Folge	6	69
Stoppen	3	111		Kampf	8	69
Gegen	3	95		Führen	9	59
Durch	2	70		Grund	7	57

Polarity and Salience

Each indicator family carries an inherent polarity (promoting + or inhibiting −) and a default salience class:

Family	Polarity	Default Salience	Discourse Function
Ursache	+	variable	Prototype; full salience spectrum
Verantwortung	+	variable	Causal-moral attribution
Folge	+	monocausal	Effect-centered perspective
Beitrag	+	polycausal	Distributional attribution
Durch	+	monocausal	Grammaticalized preposition
Stoppen	−	monocausal	Intervention framing
Gegen	−	monocausal	Variable condensation (Kampf gegen, Protest gegen)

Syntactic Projection Patterns

The syntactic realization of each indicator family determines how Cause and Effect roles are projected:

Family	Example	Cause Projection	Effect Projection
Ursache	X ist Ursache für Y	Subject	PP (für)
Folge	Y ist Folge von X	PP (von)	Subject
Verursachen	X verursacht Y	Subject	Accusative object
Beitragen	X trägt zu Y bei	Subject	PP (zu)
Stoppen	X stoppt Y	Subject	Accusative object
Durch	Y durch X	PP complement	Head noun
Gegen	X gegen Y	Subject/Agent	PP complement

Context Markers

Context markers modify and contextualize the causal relations projected by indicators. Unlike indicators, they are not lexically specialized for causality but operate at the predication or proposition level. Three structural marker types directly affect the INFLUENCE computation:

Division

Division markers signal polycausal structures with implicit co-causes. Typical realizations include unter anderem (“among other things”), auch (“also”), ebenfalls (“likewise”), and the composite nicht nur (“not only”).

Division markers reduce the salience to |I| = 0.5 regardless of the indicator’s default.

Priority

Priority markers establish asymmetric weighting within polycausal sets: vor allem (“above all”), hauptsächlich (“mainly”), maßgeblich (“significantly”). They set |I| = 0.75 for the prioritized cause.

Both marker types affect only the salience (|I|), leaving polarity (\pm) unchanged.

Negation

Negation is the structurally most influential marker. Two types are distinguished:

Object-based negation operates at the entity level through markers like Verlust (“loss”), Schwund (“decline”), Rückgang (“decrease”). These invert the polarity when the count of negations is odd:

Verlust von Lebensräumen verursacht Bienensterben.

Indicator verursachen: default I > 0; object negation on Cause (Verlust) → polarity inverted: I < 0.

Propositional negation operates at the relation level (nicht, kein) and neutralizes the entire causal relationship (I = 0):

Pestizide verursachen nicht Insektensterben.

Indicator verursachen: default I > 0; propositional negation → I = 0.

From Annotations to INFLUENCE

The annotations documented above — indicators, entities, and context markers — are the inputs to a deterministic computation that produces the final INFLUENCE value I \in [-1, +1]. In brief: entity identification follows the indicator’s syntactic projection pattern, polarity is determined by indicator class and negation markers, and salience is computed through a cascading hierarchy of morphological, determiner, and syntactic markers.

For the full formal specification — including the cascade rules, coordination normalization, and worked examples — see Tuple Construction.

Data Format

Annotated data is exported as JSON. Each entry represents a sentence with its metadata and extracted relations.

Schema

{
  "subfolder": "Artensterben_oa",
  "global_sentence_id": 1148482,
  "text_id": "FAZ_200204_384209",
  "text_date": "2002-04",
  "sentence_id": "12",
  "sentence": "...",
  "relations": [
    {
      "indicator": "Folge",
      "entities": [
        {
          "entity": "Kleinplanet",
          "relation": "Cause",
          "dependent_coefficients": [
            {"coefficient_text": "Jahren", "coefficient": "Temporality"},
            {"coefficient_text": "Mexikos", "coefficient": "Spatiality"}
          ]
        },
        {
          "entity": "Artensterben",
          "relation": "Effect"
        }
      ],
      "coefficient": "Division",
      "dependent_coefficients": [
        {"coefficient_text": "könnte", "coefficient": "Uncertainty"}
      ],
      "representation": "berieten",
      "representation_entities": [
        {
          "entity": "Teilnehmer",
          "relation": "Constraint",
          "dependent_coefficients": [
            {"coefficient_text": "fünftausend", "coefficient": "Quantity"}
          ]
        }
      ]
    }
  ]
}

Field Reference

Sentence-level fields:

Field	Description
subfolder	WABI subcorpus and syntactic position (e.g. Artensterben_oa = accusative object)
global_sentence_id	Unique sentence identifier across the full corpus
text_id	Source document identifier (format: SOURCE_YYYYMM_ID)
text_date	Publication date (YYYY-MM)
sentence	Full sentence text
relations	Array of causal relations found in this sentence (empty if none)

Relation-level fields:

Field	Description
indicator	The causal indicator lexeme
entities	Array of entities with their causal role (Cause or Effect)
coefficient	Structural marker on the relation level (Negation, Division)
dependent_coefficients	Contextual coefficients attached to the indicator
representation	Evidentiality/speech-act verb (if present)
representation_entities	Source entities for reported speech

Entity-level fields:

Field	Description
entity	Head token of the entity (token-minimized)
relation	Causal role: Cause, Effect, or Constraint
dependent_coefficients	Array of coefficients modifying this entity

Sentences Without Relations

Sentences where no explicit causal relation was identified have an empty relations array. These are not noise — they were reviewed during annotation and determined to contain no explicit causal markers per the minimal principle.

Corpus Statistics

Total sentences	4,753
Sentences with ≥1 WABI relation	1,797 (37.8%)
Total WABI-relevant relations	1,867
Distinct indicator forms	642
Indicator families	192
Mean relations per sentence	0.39

Per WABI term:

Term	Sentences	Relations	Rel./Sent.
Waldsterben	1,818	633	0.35
Artensterben	1,854	744	0.40
Bienensterben	536	257	0.48
Insektensterben	545	233	0.43

Further Reading

• For the theoretical motivation behind polarity and salience, see Framework
• For how annotations are transformed into (C, E, I) tuples, see Tuple Construction
• For the C-BERT model trained on this data, see C-BERT
• Full annotation data (Bundestag subset): HuggingFace Dataset