Coding Structures and Actions with the COSTA Scheme in Medical Conversations

Nan Wang, Yan Song, Fei XiaBioNLP 2018 · DOI · OpenReview

TL;DR

This paper introduces COSTA, a novel annotation scheme for capturing the multi-layer structure and actions within conversations, particularly in medical settings. They've created a rich corpus of pediatric consultations annotated with COSTA, enabling them to analyze treatment decision-making and the impact of communication on antibiotic prescribing.

What Problem Does This Paper Address?

Understanding human conversation is a long-standing challenge across various fields, from linguistics to AI. While existing approaches have analyzed conversational structures and actions, there's a lack of a comprehensive, hierarchical scheme that captures the full complexity of conversations. Furthermore, medical communication, a critical area for patient outcomes, has been hampered by a scarcity of annotated language resources. This paper aims to bridge this gap by providing a robust annotation framework (COSTA) and a valuable corpus of medical conversations to facilitate deeper analysis of these interactions.

Methods

The COSTA scheme is deeply rooted in Conversation Analysis (CA), a sociological approach that views conversation as systematically organized. Unlike text-based discourse analysis, CA emphasizes the dynamic, interactive nature of talk. COSTA formalizes CA's insights into a multi-layer annotation system with four key components:

  • Turn: The basic unit of conversation, marked by a change in speaker.
  • Adjacency Pair: Pairs of turns that form a coherent unit (e.g., question-answer, offer-acceptance). COSTA distinguishes between base adjacency pairs and their expansions (insert, pre-, post-).
  • Sequence of Actions: A collection of turns organized around a base adjacency pair and its expansions, representing a specific social action.
  • Phase: The highest level of organization, segmenting conversations into distinct stages (e.g., history taking, diagnosis, treatment).

A key innovation is the structure-informed taxonomy of actions. Instead of relying solely on the surface form of an utterance (like Speech Act Theory), COSTA uses the turn's sequential position within the conversation's structure to determine its action. This allows for more accurate identification of indirect speech acts.

Datasets

The paper introduces a newly created corpus of 318 video-recorded and manually transcribed pediatric consultations collected from five hospitals in China in 2013. This corpus is unique in its multi-layer structural annotations using the COSTA scheme. The data includes detailed transcriptions capturing both spoken words and speech production features, as well as non-verbal activities.

Key Insights and Ideas

  • Hierarchical Structure is Key: The paper strongly advocates for understanding conversations not as flat sequences of turns, but as nested structures (turns within sequences, sequences within phases). This hierarchical view is crucial for accurately interpreting actions.
  • Structure Informs Action: The core idea is that a turn's meaning and the action it performs are heavily dependent on its position within the conversational structure. This is a significant departure from approaches that focus solely on the utterance itself.
  • Medical Conversations are Complex: The analysis of pediatric consultations reveals that treatment recommendation phases are particularly expanded, suggesting that communication challenges and negotiations are most likely to occur during these critical stages.
  • Caregiver Influence on Antibiotic Prescribing: A surprising finding is the significant influence of caregiver advocacy on antibiotic prescribing outcomes, suggesting physicians play a less dominant role than previously assumed in this specific context.

Key Results

  • The COSTA scheme was applied to a corpus of 318 pediatric consultations.
  • Analysis of the treatment decision-making process showed that physicians' treatment recommendations are resisted by caregivers 41% of the time, leading to expanded action sequences.
  • The treatment phase (P5) exhibits the longest action sequences on average, indicating more complex communication dynamics.
  • Caregiver advocacy for antibiotics was significantly associated with increased antibiotic prescribing (OR = 9.23).
  • Physicians' resistance to caregiver advocacy reduced the likelihood of antibiotic prescriptions by 77% (OR = 0.23).
  • The inter-annotator agreement for structural annotations was high at 94.43%, and for task-specific annotations, it was 97%.

Limitations and Future Work

  • Task-Specific Domain: While the COSTA scheme is designed to be extensible, it was developed and tested on pediatric consultations. Its applicability to other general domains and languages needs further validation.
  • Turn Construction Units (TCUs): The current annotation focuses on turns as the smallest unit. Future work aims to extend COSTA to mark Turn Construction Units (TCUs) for even finer-grained analysis.
  • Dataset Release: The authors plan to release the annotated dataset once it is completed, which will be a valuable resource for the research community.
  • Cross-Cultural Application: The paper suggests that COSTA can be applied to languages other than Chinese, given that social norms underlying conversations don't vary drastically across cultures, but this requires empirical testing.

Summary updated

Abstract

This paper describes the COSTA scheme for coding structures and actions in conversation. Informed by Conversation Analysis, the scheme introduces an innovative method for marking multi-layer structural organization of conversation and a structure-informed taxonomy of actions. In addition, we create a corpus of naturally occurring medical conversations, containing 318 video-recorded and manually transcribed pediatric consultations. Based on the annotated corpus, we investigate 1) treatment decision-making process in medical conversations, and 2) effects of physician-caregiver communication behaviors on antibiotic over-prescribing. Although the COSTA annotation scheme is developed based on data from the task-specific domain of pediatric consultations, it can be easily extended to apply to more general domains and other languages.
BibTeX
@inproceedings{wang2018coding,
  author = {Nan Wang and Yan Song and Fei Xia},
  title = {{Coding Structures and Actions with the COSTA Scheme in Medical Conversations}},
  booktitle = {BioNLP},
  year = {2018},
  doi = {10.18653/v1/w18-2309},
}
Figure