📄 Modelling the Interpretation of Discourse Connectives by Bayesian Pragmatics

Summary

The paper asks whether Bayesian pragmatics can explain how readers settle the sense of discourse connectives (DCs) such as 'since' or 'in fact', and of implicit relations where no connective is present. The authors adapt the Rational Speech Acts (RSA) model. A literal listener reads the sense of a DC straight from corpus co-occurrence counts. A pragmatic speaker picks a DC by weighing how informative it is against a production cost, and an implicit ('null') DC costs nothing. A pragmatic listener then inverts that speaker with Bayes' rule, using sense frequency as the prior. All probabilities are estimated from the Penn Discourse Treebank, simplified to 15 senses following the CoNLL shallow discourse parsing convention. The one-level pragmatic listener (L1) predicts the annotated senses of explicit DCs significantly better than the literal listener. Reasoning a second level deeper gives no further gain. Conditioning on the previous relation as context did not help. In a second experiment, the parser's sense probabilities are used as the prior of L1. This raised the accuracy of the winning CoNLL 2015 parser. The work suggests that a cognitively motivated pragmatic layer can be added on top of a statistical classifier.

Key points

• Data: PDTB with 2 form labels (Explicit / Non-explicit) and 15 senses. Sections 2–22 are used for training (33,971 samples) and sections 0, 1, 23 and 24 for testing (6,375 samples).
• On explicit DCs, accuracy is .8767 for the literal listener L0 and .8853 for L1, significant at p < 0.02 (McNemar). L2 matches L1, so a second level of recursion adds nothing.
• Non-explicit relations stay at about .26 accuracy for all listeners, because an absent connective carries little information.
• Three kinds of context were tried: the previous relation's form, its sense, and the form-sense pair. None of them beat the constant (no-context) setting for explicit DCs.
• The parser is Wang and Lan (2015). Using its output as L1's prior raised overall accuracy from .5833 to .5916 (p < 0.05). Concession F1 rose the most, from .3048 to .4382.

Interesting ideas

• I1. The "say nothing" option is modelled as an utterance in its own right: the null DC has zero cost, while every explicit DC has a small constant cost, so the listener can reason about why a writer chose to mark a relation.
• I2. A pragmatic listener can be built entirely from counts: corpus co-occurrences give the literal listener, and a softmax speaker plus one Bayes inversion gives the pragmatic one, with no neural training needed.
• I3. Any classifier's probability distribution can be swapped in as the prior of the pragmatic listener, so the RSA layer acts as a post-hoc reranker that measurably improves the classifier.
• I4. Comparing L0, L1 and L2 directly showed that one level of recursion is enough, which keeps the model cheap.
• I5. The speaker model scores each connective by how well a listener would recover the intended sense from it, minus its cost.

For Mova

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

• I1 (frames): the implicature module in prag could treat an explicit 'but' or 'so' as evidence the writer felt a need to mark the relation, and treat its absence as a weaker default, with each inference explainable as a cost-versus-informativeness trade-off.
• I2 (frames): this fits the explainable design of prag, because every interpretation can be traced back to readable count tables and a single inversion step.
• I3 (learn): the sense or relation distributions from gparse or witness parsers could serve as priors that a speaker-model reranker refines, and the gain could be checked against a literal-listener baseline as a negative control.
• I4 (answer): when readers answer questions about fables and need speaker-intent reasoning, they can stop at depth one and use the L0-versus-L1 comparison as a diagnostic.
• I5 (back): during generation from the meaning graph, the system can choose whether to insert a connective, and which one, by maximising listener recoverability of the relation minus a small cost.