Introduction
In the field of computational linguistics, resolving ambiguous pronouns like "he," "she," or "it" remains a significant challenge. This 1987 paper from Hewlett-Packard Laboratories addresses this by extending "Centering Theory"—a framework for modeling how humans track the "center of attention" in discourse. By formalizing these concepts into a concrete algorithm, the authors provide a method for computers to navigate the structured paths of human conversation.
The Centering Model
The authors build upon the work of Grosz, Joshi, and Weinstein (1986), tracking three key components for every utterance ($U_n$):
- Forward-looking centers ($Cf(U_n)$): A list of all entities mentioned, ranked by grammatical role (Subject > Object > Others). The highest-ranked item is the Preferred Center ($Cp$), the most likely candidate for the next sentence's topic.
- Backward-looking center ($Cb(U_n)$): The entity the current sentence is "about," which must have been mentioned in the previous sentence.
Innovation: Extended Transitions
The core contribution of this research is the refinement of transition types between sentences. While previous models used three transitions, this paper proposes four, ranked by their perceived coherence to a listener. This hierarchy allows the algorithm to make deterministic choices in complex cases where multiple pronouns appear in a single sentence.
As shown in the matrix above, the transitions are defined by the relationship between the current $Cb$, the previous $Cb$, and the current $Cp$:
- Continuing: The center stays the same, and it is the most prominent entity ($Cb(U_n) = Cb(U_{n-1})$ and $Cb(U_n) = Cp(U_n)$).
- Retaining: The center stays the same, but the speaker is signaling a move to a new entity ($Cb(U_n) = Cb(U_{n-1})$ but $Cb(U_n) \neq Cp(U_n)$).
- Shifting-1: The center changes, but the new center is the most prominent entity in the current sentence ($Cb(U_n) \neq Cb(U_{n-1})$ and $Cb(U_n) = Cp(U_n)$).
- Shifting: The center changes, and it is not the most prominent entity.
The Resolution Algorithm
The resolution process follows three distinct phases:
- Construct: Generate all possible interpretations (anchors) for the pronouns.
- Filter: Remove interpretations violating syntactic rules (like contra-indexing) or centering constraints.
- Classify and Rank: Categorize the remaining interpretations into the four transition types. The interpretation resulting in the most coherent transition (Continuing > Retaining > Shifting-1 > Shifting) is selected.
Practical Application
The paper demonstrates the algorithm's effectiveness through linguistic examples. By tracking the $Cb$ and $Cf$ lists across a narrative, the system can resolve pronouns even when multiple candidates exist.
In the example above, the algorithm tracks "Carl" as the center through several "Continuing" transitions. When a "Retaining" transition occurs (as in $U_{n+4}$), it signals a potential shift in focus, allowing the algorithm to correctly identify that "She" refers to Lyn and "him" refers to Carl based on the preceding discourse structure.
Conclusion
By prioritizing subjecthood and using coherence as a tie-breaker, the Brennan et al. algorithm successfully resolves edge cases that previously caused centering models to fail. While it focuses on local coherence and lacks broad "world knowledge," it remains a foundational approach to structural pronoun resolution in natural language processing systems.
Summary updated