📄 Current theories of centering for pronoun interpretation: a critical evaluation

Summary

This short critical paper ("squib") in Computational Linguistics asks whether a widely used centering-based pronoun resolution method reflects the reasons centering theory was proposed. Centering (Grosz, Joshi, and Weinstein 1995, GJW) is motivated by garden-path effects, which suggest that listeners commit to a pronoun's referent before the whole sentence is understood. Kehler examines the BFP algorithm (Brennan, Friedman, and Pollard 1987, also used by Walker, Iida, and Cote 1994). BFP ranks candidate assignments using transition types between utterances.

The study uses no corpus or experiments. It argues from constructed examples and informal acceptability judgments. BFP handles GJW's well-known Terry/Tony example. However, the backward-looking center (Cb) of an utterance depends on the entire utterance, so BFP must read a full sentence before choosing referents. As a result, it predicts different preferences for minimally different sentences that speakers judge alike.

A second set of examples shows that coherence relations such as narration versus parallelism also shift pronoun preferences, and BFP cannot capture this. The paper cautions against treating BFP as a model of incremental, human-like pronoun interpretation.

Key points

• Centering concepts reviewed:
  • Cb, the backward-looking center.
  • Cf, the ranked list of forward-looking centers.
  • Cp, the highest-ranked member of Cf.
  • Rule 1, about pronominalizing the Cb.
  • Rule 2, about preferences among transitions.
• What BFP adds: it splits Shift into Smooth-Shift and Rough-Shift. It then orders the transitions Continue > Retain > Smooth-Shift > Rough-Shift and uses a generate–filter–rank procedure.
• Example (6), Terry/Tony variants e1–e3: speakers appear to share the same initial preference across all three variants, which produces a garden path. BFP's analysis, however, differs between them (Table 2).
• Incremental fix also fails: an incremental version of BFP that assigns provisional referents (footnote 7) would wrongly predict a double garden path for variant e2.
• Example (7), Dole/Clinton/Perot: two follow-ups have identical Cb/Cp configurations but different preferred referents. This points to approaches that model coherence relations (e.g., Kameyama 1996; Kehler 1995).