Dependencies of Discourse Structure on the Modality of Communication: telephone vs. Teletype

Philip R. Cohen, Scott Fertig, Kathy StarrACL 1982 · DOI

Introduction

This foundational 1981 study by Cohen, Fertig, and Starr addresses a question central to modern Natural Language Processing (NLP) and Human-Computer Interaction (HCI): Does the medium of communication—speaking versus typing—fundamentally alter the structure of discourse? As we move toward more sophisticated voice assistants, understanding these differences is critical for building systems that can truly "listen."

The Modality Assumption

In the early 1980s, the prevailing assumption was that a functional voice interface could be created by simply adding a speech-to-text front end to existing keyboard-based systems. The authors challenge this, arguing that spoken discourse is not merely "text with sound." Instead, it possesses a unique structure that reflects the cognitive and physical constraints of the medium. If systems are built based on how people type, they will likely fail to interpret the indirect and fragmented nature of natural speech.

Experimental Design

The researchers employed a "Wizard of Oz" style exploratory study where experts provided instructions to apprentices for assembling a toy water pump. The study compared five communication modes, with this analysis focusing on the contrast between Telephone (interactive speech) and Teletype (interactive typing).

The analysis involved transcribing dialogues and coding them using a speech act taxonomy based on a "plan-based theory of communication." This framework views utterances as strategic actions intended to achieve specific goals, such as a Request for Action (RAACT) or a Request for Identification (RID).

The "Find the X" Strategy

A primary finding is that speech is significantly more "fine-grained" than typing. In Telephone dialogues, experts typically employed a two-step strategy: first, ensuring the apprentice identified the correct part ("See that red ring?"), and second, providing the assembly instruction. In Teletype, these steps were almost always collapsed into a single command ("Fit the red ring over the tube").

Furthermore, identification requests in speech were almost exclusively indirect. Rather than saying "Find the plunger," speakers used forms like "There's a thing called a plunger?" or "It's got a peg in it."

Quantitative Results

The data reveals a stark contrast in how frequently identification is explicitly managed across modalities. Requests for Identification (RID) accounted for 35% of all requests in Telephone mode, compared to a mere 10% in Teletype mode.

Distribution of request types across Telephone and Teletype modalities.

This difference was not merely an aggregate trend but a consistent behavioral pattern among subjects. As shown in the subject-level data, four out of five experts in the Telephone group were "habitual" users of the separate identification strategy, whereas zero experts in the Teletype group adopted this approach.

Frequency of request types by individual subject, highlighting the consistency of the RID strategy in speech.

Conclusion

The study demonstrates that modality dictates strategy. The interactive nature of speech facilitates a high-frequency confirmation loop that the effort of typing discourages. For computational systems, this implies that speech-understanding models must be capable of handling fragments and inferring intent from indirect identification requests. Building effective "mechanical listeners" requires accounting for these fundamental differences in discourse structure.

Summary updated

BibTeX
@inproceedings{cohen1982dependencies,
  author = {Philip R. Cohen and Scott Fertig and Kathy Starr},
  title = {{Dependencies of Discourse Structure on the Modality of Communication: telephone vs. Teletype}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1982},
  doi = {10.3115/981251.981258},
}
Figure