"Despite" coding#

The problem: countervailing conditions that “failed”#

Narratives often contain claims of the form:

Examples:

• “The heavy rains made the river levels rise, despite the prevention and clearance work.”
• “Despite me reminding him multiple times, he missed the train.”

In ordinary language, the “despite” clause does two things at once:

1. It signals that Z was expected to prevent or reduce E (a countervailing influence).
2. It asserts that E happened anyway (so Z was insufficient, absent, overridden, or ineffective in that case).

If we only code the main causal claim (e.g. Heavy rains -> River levels rose), we lose potentially important information about attempted mitigations, resistance, barriers, or protective factors.

But it is also usually wrong to code the “despite” clause as an ordinary positive causal link, e.g. Prevention work -> River levels rose, because that flips the intended meaning on its head.

The convention: a “despite” link type (or tag)#

We therefore treat “despite” as a link type, not a different semantics for ->.

A minimal encoding is:

• Main link (ordinary): X -> E
• Despite link (typed/tagged): Z -despite-> E

Equivalently (in a plain links table), keep the same Cause and Effect columns but add either:

• a link_type column with values like normal vs despite, or
• a link_tags column containing a tag like #despite.

The key idea is that -despite-> does not mean “Z caused E”. It means:

What can we do with “despite” links?#

Because the distinction is explicit in the links table, we can decide—per analysis—how to treat these links:

• Visualise them distinctly (different colour/line style), so the map shows “tensions” rather than silently dropping them or misreading them.
• Filter: show only despite links to see what people present as failed protections / resistances / mitigations.
• Compare: for similar outcomes, do some sources describe Z as effective (ordinary preventive claim, e.g. Z -> ~E) while others describe “despite” failures (Z -despite-> E)? That contrast can be analytically important.
• Count separately: include them in evidence tallies only under a separate counter (e.g. Despite_Citation_Count) so you don’t accidentally inflate evidence for “Z causes E”.

Worked example#

Text:

Coding:

• Heavy rains -> River levels rose
• Prevention and clearance work -despite-> River levels rose

This preserves the main mechanism while also storing the claim that a mitigation was present but insufficient.

Optional background: force dynamics (why “despite” is cognitively natural)#

Leonard Talmy’s “force dynamics” treats many causal expressions as patterned talk about forces: an Agonist with a tendency (towards motion/rest) and an Antagonist that counters it. “Despite” is a canonical force-dynamics marker: it signals a countervailing force that failed to stop the outcome.

You do not need this theory to use -despite-> coding; it is only a helpful explanation of why “despite” claims feel different from ordinary causal claims, and why it can be worth capturing them explicitly.

Hard cases (brief notes)#

• “Despite” + opposites: should Z -despite-> E be treated as evidence for Z -> ~E? Usually no; treat it as its own link type, or at most as weaker evidence depending on your analysis goal.
• Bipolar concepts: some pairs (employment/unemployment) have high cosine similarity (similar contexts) but are "opposite" in ordinary talk. This is why sentiment often becomes relevant when doing AI-assisted clustering/aggregation.

4.1 Force Dynamics: Agonists and Antagonists in Latent Space#

Leonard Talmy’s theory of Force Dynamics posits that human causal understanding is rooted in the interplay of forces: an Agonist (the entity with a tendency towards motion or rest) and an Antagonist (the opposing force).   

• Linguistic Patterns: "The ball kept rolling despite the grass" (Agonist: Ball; Antagonist: Grass). "He let the book fall" (Removal of Antagonist).

• LLM Evaluation: Recent studies  have tested LLMs on translating and explaining these force-dynamic constructions.   

  • Findings: GPT-4 demonstrates a sophisticated grasp of these concepts. When translating "He let the greatcoat fall" into languages like Finnish or Croatian, the model correctly selects verbs that convey "cessation of impingement" (allowing) rather than "onset of causation" (pushing).   

  • Implication: This suggests that LLMs have acquired a schematic semantic structure of causality. They do not merely predict words; they map the roles of entities in a physical interaction. However, this capability degrades in abstract social contexts. For example, in the sentence "Being at odds with her father made her uncomfortable," models sometimes misidentify the Agonist/Antagonist relationship, struggling to map "emotional force" as accurately as "physical force".   

**