Introduction
This 1986 study from the early era of computational linguistics investigates the underlying organization of human-to-human technical support conversations. By observing novice users interacting with an adviser to learn a statistical software package, the authors demonstrate that dialogue is not a random stream of consciousness. Instead, it follows a structured hierarchy that directly mirrors the task being performed.
The Problem: Anaphora Resolution
In the mid-1980s, natural language processing (NLP) struggled with anaphora resolution—the challenge of determining what words like "it," "that," or "the vector" refer to in a conversation. Searching through every previous sentence for a matching referent is computationally expensive and error-prone. To address this, researchers proposed focusing: narrowing the search to a small subset of "active" knowledge. The central question of this paper is whether the structure of the task itself can provide the necessary "map" for this focus.
Methodology
The researchers recorded three novice users solving five statistical problems. They employed a three-step analytical approach:
- Task Analysis: Using the GOMS model (Goals, Operators, Methods, and Selection rules), they decomposed statistical problems into a hierarchy of sub-tasks.
- Dialogue Segmentation: Conversations were segmented into "subdialogues" based on the task hierarchy, including "meta-plans" like clarifications or acknowledgments.
- Validation: The derived structure was tested against linguistic markers, specifically noun phrase distribution and antecedent mapping.
Key Insights: Dialogue Mirrors Task
The study's primary finding is that the hierarchy of a conversation (subdialogues within dialogues) parallels the hierarchy of the task (subtasks within tasks). As shown in Figure 1, specific dialogue acts like clarifications and acknowledgments are nested within the broader goal-oriented structure.
Nouns and Pronouns as Structural Signals
The researchers found that full noun phrases (non-pronominal) act as linguistic anchors, appearing most frequently at the start of a subdialogue to signal a topic shift or to reinstate an object. Conversely, pronouns serve as continuity signals; they are rare at the start of a subdialogue but increase in frequency as the conversation maintains a specific topic.
Quantitative Results
The empirical data strongly supported the hierarchical model. In subdialogues of varying lengths, the distribution of noun phrases followed a predictable pattern.
As seen in Figure 3, in two-sentence subdialogues, full nouns appeared 234 times in the first sentence (S1) but dropped to 114 in the second (S2). Pronouns showed the opposite trend, being rare in S1 and increasing in S2. Furthermore, the vast majority of pronoun referents were found within the immediate "parent" or "current" subdialogue node, confirming that the hierarchical structure effectively constrains the search space for anaphora resolution.
Conclusion
The paper concludes that while task structure heavily influences dialogue, it is not the sole determinant. Novices often drift into clarification loops that the task model cannot fully predict. However, the structural evidence remains robust. Future natural language interfaces must incorporate models of user goals and "meta-plans" to navigate the "method in the madness" of human communication effectively.
Summary updated