📄 Embedded implicatures observed: a comment on Geurts and Pouscoulous (2009)

Summary

This short reply tests a claim by Geurts & Pouscoulous (2009a). They argued that when a scalar word like some is embedded under a universal, for instance in a sentence saying that every square is linked to some of the circles, it does not get strengthened to "some but not all." Their evidence came from a yes/no picture verification task. Clifton and Dube argue that this task is biased. A sentence can count as a "correct" description of a picture even when it is not the best one.

They replaced the yes/no judgement with a forced choice among four options: picture A, picture B, both, or neither. One picture fits the locally strengthened reading and the other fits only the literal reading.

• Non-embedded some (Experiment 1): readers clearly chose the "some but not all" picture.
• Embedded some (Experiment 2): "both" was the most common answer. When readers did pick one picture, they preferred the strengthened one. When neither picture fit the strengthened reading, many rejected both.

The authors conclude that the case against embedded implicatures is not established. Such readings seem possible but not obligatory, which creates problems for both Gricean and mainstream conventionalist accounts.

Key points

• Geurts & Pouscoulous's own data show that the task changes the result. For simple some, 62% accepted the strengthened inference in a written task, but only 34% rejected the picture in verification.

• Experiment 1 involved 36 UMass undergraduates, 4 critical some items, and 84 other items (practice, fillers, other experiments). The strengthened picture B beat A, with a 95% CI of the difference of .54–.82. "Both" did not rise above the .25 chance level.

• In Experiment 2 (all/each … some), when picture B was strengthened, choices were:

  Option	all	each
  B (strengthened)	39%	38%
  Both	57%	63%
  A	3%	0%

• When neither picture was strengthened, "neither" rose to 17% (all) and 21% (each). A was chosen more often than B (28/24% vs 6/4%).

• The authors note that minimal conventionalism, or treating some as simply ambiguous, fits the data. They consider these views too unconstraining to adopt except as a last resort.

Interesting ideas

• I1. Asking which option fits best, rather than whether one option is acceptable, exposes preferences that a yes/no judgement hides, because weak but true readings always pass yes/no.
• I2. Adding a control version where no option satisfies the strong reading, and then counting rejections, gives an independent signal that the strong reading is active.
• I3. The "less-wrong" analysis counts how many entities violate the strengthened reading in each picture, which explains a graded preference without assuming all-or-nothing truth.
• I4. Embedded strengthening seems optional and weaker than unembedded strengthening, so implicatures should carry context-dependent strength rather than always firing.
• I5. Running both a verbal inference task and a picture task on the same items showed that each method has its own bias.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Evaluate implicature handling by having the model rank competing world states for a sentence, rather than checking each one against the text in isolation.
• I2 (learn): Build negative-control items where no candidate world matches the strengthened reading, and check that the system then rejects all candidates instead of picking one.
• I3 (world): When scoring a text's world against a reading, count violating entities to give a graded mismatch score instead of a single true/false.
• I4 (frames): In crate prag, mark some under all/each as an optional, low-confidence "not all" implicature. Unembedded some should get a strong default.
• I5 (answer): Probe the same implicature through both a direct entailment question and a scenario-choice question, and treat disagreement between them as a diagnostic.