📄 Computational Investigations of Pragmatic Effects in Natural Language

Summary

This student-workshop thesis proposal argues that NLP has focused on literal meaning while neglecting pragmatics, the context-dependent intended meaning. It reports two finished studies that use function words as testbeds for pragmatic reasoning. The first is definiteness prediction: choosing the, a(n) or no article for a noun phrase. An attention-equipped LSTM beats a feature-engineered classifier. Longer context helps it handle article choices that seem to depend on coreference, including synonymous mentions. The second introduces a new task: deciding whether a context licenses an adverbial presupposition trigger such as again or also. The author builds binary datasets from Penn Treebank and English Gigaword and proposes a parameter-free weighted pooling attention model that outperforms the baselines on most triggers. The remaining plan has four parts:

• a crowdsourced corpus of presuppositional effects,
• summarizers that insert or drop adverbs to stay coherent,
• reading-comprehension questions whose answers are implied rather than stated,
• multi-task reuse of presupposition-aware representations.

The work matters because it treats implicit, presupposed content as a core modeling target rather than an afterthought.

Key points

• Definiteness model: an LSTM with attention, POS tags, pre-trained vectors and extended context reaches 96.63% accuracy on WSJ/PTB. This is a 51% relative error reduction over the logistic regression of De Felice (2008), which used 10 types of hand-crafted features.
• Ablations showed that pre-trained initialization, extra context, POS tags and attention all help, but by different amounts. Attention weights were used to interpret predictions.
• The presupposition trigger datasets come from Penn Treebank and English Gigaword and cover five adverbs: too, again, also, still, yet. Class balance and the adverb's syntactic governor were controlled so that models cannot succeed through shallow correlations.
• The weighted pooling attention is a self-attention variant that learns how to reweight attention scores rather than averaging hidden states uniformly. It adds no parameters over the LSTM baseline and reaches 82.42% accuracy for "also" on Gigaword. Baselines were logistic regression, RNN and CNN models.
• The author notes that SQuAD is unsuitable for pragmatic QA because its answers are always explicit text spans.

Interesting ideas

• I1. When building positive vs. negative contexts for a trigger, the author matches confounds such as label balance and the governing word, so a classifier cannot win through syntax alone.
• I2. Adverbs like again, also, still and yet signal that a similar or earlier event exists, or that a state continues or ends. Reading them as event-history markers turns a function word into a world-state constraint.
• I3. Article choice reflects discourse status: indefinites introduce a new referent, and definites point back to a known one, sometimes through a synonym.
• I4. The proposed QA probes ask yes/no questions whose answer follows only from a presupposition. Hard negatives put the trigger on a different person, so a pure trigger detector fails.
• I5. A summarizer should check, sentence by sentence, whether adding or dropping a trigger adverb keeps the presupposed context true and the output coherent.

For Mova

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

• I1 (learn): Our negative controls could pair each trigger-bearing test sentence with a governor-matched sentence lacking the trigger, checking that rules respond to discourse rather than to head verbs.
• I2 (world): Each "again" could add a presupposed earlier event with the same frame before the current one in the event order, and "still/yet" could assert that a state persists.
• I3 (frames): The coreference step could treat "a N" as opening a new entity and "the N" as a strong cue to link to an earlier, possibly synonymous mention.
• I4 (answer): We could write fable and tale tests like "Had X done this before?" with distractors where another character carries "again", checking that answers come from the world model.
• I5 (back): The graph-to-text generator could emit "again/also/still" only when the meaning graph holds the matching prior or parallel event, and omit it otherwise.