When Only Time Will Tell: Interpreting How Transformers Process Local Ambiguities Through the Lens of Restart-Incrementality

TL;DR

This paper investigates how restart-incremental Transformers handle local ambiguities by formalizing RI as a transition-system around a base model, and by representing internal state evolution with triangular 3D structures that track $s^t$ over prefixes. It develops glass-box interpretability methods to measure how future context updates past representations and aligns these state revisions with output edits, focusing on two RI scenarios: meaning construction in bidirectional LMs and incremental dependency parsing. The empirical results show that RI models exhibit garden-path-like revisitations, with initial commitments revised when disambiguating region tokens are integrated, and that revision dynamics are more pronounced in middle-to-upper layers; dependency parsing reveals that shifts in attention distributions correlate with edits, especially for arcs. Overall, the work demonstrates that RI operates as a non-monotonic, revisable processing regime in BiDi transformers, offering insights into garden-path phenomena and informing the design of incremental NLP systems that require non-monotonic reasoning and revision capabilities.

Abstract

Incremental models that process sentences one token at a time will sometimes encounter points where more than one interpretation is possible. Causal models are forced to output one interpretation and continue, whereas models that can revise may edit their previous output as the ambiguity is resolved. In this work, we look at how restart-incremental Transformers build and update internal states, in an effort to shed light on what processes cause revisions not viable in autoregressive models. We propose an interpretable way to analyse the incremental states, showing that their sequential structure encodes information on the garden path effect and its resolution. Our method brings insights on various bidirectional encoders for contextualised meaning representation and dependency parsing, contributing to show their advantage over causal models when it comes to revisions.

Figures (25)

• Figure 1: A prefix with multiple valid continuations. A causal decoder is forced to output only one POS-tag for the token can at this point and cannot change it anymore, even if its internal state encodes the local ambiguity. In contrast, a restart-incremental model can perform revisions and would thus be able to recover if the selected label turned out to be incorrect (as in left).
• Figure 2: Triangular structures representing states built step by step in restart-incremental sequential processing.
• Figure 3: We realign tokens, after computing the states, by removing the states of additional token(s) from the triangular structure. That way we can directly compare the states of a locally ambiguous sentence with its unambiguous counterpart.
• Figure 4: The three types of stimuli (s) and their corresponding reference baselines (b) used in our analyses. The words in lilac/bold are locally ambiguous until the underlined/yellow token is observed.
• Figure 5: BERT's mean effect of the second noun on the tokens in the prefix. Absolute difference over baseline.