Politeness and frustration language in child-machine interactions

Sudha Arunachalam, Dylan Gould, Elaine Andersen, Dani Byrd, Shrikanth S. NarayananInterspeech 2001 · DOI

TL;DR

This paper investigates how children aged 6-14 express politeness and frustration when interacting with a computer game. They found that younger children are less likely to use overt politeness markers and more prone to verbalizing frustration, with gender playing a role in frustration expression. This research is crucial for designing more effective and engaging conversational interfaces for children.

What Problem Does This Paper Address?

As conversational interfaces become more sophisticated, a key challenge is tailoring them to diverse user groups. Children, in particular, represent a significant user segment for applications like educational games and interactive media. However, children's language use, social understanding, and interaction strategies differ from adults. This paper addresses the gap in understanding these specific linguistic behaviors in child-computer interactions, focusing on how children express politeness and frustration. Without this understanding, systems risk being ineffective, frustrating, or simply not engaging for young users. The authors aim to inform the design of more natural and efficient spoken language systems for children by analyzing their unique discourse patterns.

Methods

The study employed a "Wizard of Oz" (WOZ) experimental design. In this setup, a human operator (the "wizard") secretly controlled the computer's responses, simulating a fully functional speech recognition and natural language understanding system. This allowed researchers to collect realistic child-computer dialogue data without the limitations of current ASR/NLP technology for children.

The core methodology involved analyzing transcripts of these WOZ-simulated interactions. Specifically, the researchers focused on:

  • Frustration Markers: Identifying and cataloging words children used to express frustration, difficulty, or annoyance. They created a lexicon of 21 such words.
  • Rejection Language ("no"): Analyzing the use of "no" as a systematic indicator of a child rejecting a system response or action, particularly when they changed their mind.
  • Politeness Markers: Searching for overt politeness terms like "please," "thank you/thanks," and "excuse me."
  • Information Request Forms: Examining the linguistic structure of requests, categorizing them into five forms ranging from least to most polite based on modal verbs (e.g., "can," "will," "could," "would," "may").

Quantitative analysis involved calculating the frequency of these markers and comparing them across different age groups (6-8, 10-11, 13-14 years), genders, and game outcomes (win/loss).

Datasets

The study utilized a dataset from the "ChIMP - Children's Interactive Multimedia Project" database, collected in 1997. This database comprised spoken interactions from 160 children (boys and girls) aged six to fourteen years old, playing a computer game ("Where in the USA is Carmen Sandiego?"). The data was collected via a Wizard of Oz setup. No new benchmarks were introduced; the paper focuses on analyzing this existing dataset.

Key Insights and Ideas

  • Age Matters for Politeness: Younger children (6-8) are significantly less likely to use overt politeness markers like "please" or "thank you" compared to older children. This suggests their cognitive resources might be prioritized for task negotiation, or they may not yet fully grasp the social nuances of politeness with a non-human interlocutor.
  • Frustration is Age and Gender Dependent: Younger children express frustration verbally more often than older children. Interestingly, male children were found to verbalize frustration more frequently than female children.
  • "Shut Up" is King (of Frustration): The word "shut up" was the most common frustration marker identified, far exceeding others like "oh man" or "heck." This highlights a direct and less filtered expression of annoyance.
  • "No" as a System Rejection Signal: The word "no" systematically indicated a child rejecting a system's action or response, often when they changed their mind. Losing games and playing the second game in a series correlated with increased use of "no."
  • Modal Politeness Increases with Age: While overt politeness markers show age-related differences, the form of information requests also indicates politeness. Older children increasingly use more polite modal verbs (e.g., "could," "would," "may") for requests, suggesting a developing understanding of register and social hierarchy in their interactions.
  • Older Children's Variable Politeness: Some older children showed less overt politeness, potentially viewing the computer as less "animate" or more of a peer, thus not requiring the same level of deference.

Key Results

  • Frustration: Younger children (8-10) used frustration markers more than older children (11-14). Male children used frustration markers four times more often than female children (0.16% vs. 0.04%). Frustration occurred more than twice as often in lost games (0.13%) compared to won games (0.06%).
  • Rejection ("no"): Males rejected system actions more frequently than females. Younger children (8-9) made more frequent rejections than older children (12-14). Losing games and playing the second game in a series led to more "no" utterances.
  • Politeness Markers: No significant differences were found in the frequency of "please," "thank you/thanks," or "excuse me" based on age group or sex in a general ANOVA. However, a more fine-grained analysis suggested that "please" and "thanks/thank you" usage varied with age, being more common in younger and middle groups, with older children showing more variability.
  • Information Requests: Older children used more polite modal forms (e.g., "could," "would," "may") for requests compared to younger children. No significant differences were found for less polite forms ("can," "will") or between genders in modal politeness.

Limitations and Future Work

Limitations:

  • The study acknowledges that the WOZ setup assumed "perfect" speech recognition and understanding. This means the collected data doesn't reflect the impact of actual system errors on child behavior.
  • The analysis of politeness markers was normalized by total user turns, which the authors suggest might be less informative than normalizing by turns belonging to specific dialog states.
  • The age grouping for politeness analysis (younger, middle, older) was found to be potentially too broad, with significant variability observed within these groups, particularly among the youngest children.

Future Work:

  • The authors emphasize the crucial need for future research to analyze child-machine conversations under errorful scenarios to better understand how children react to actual system failures.
  • Further investigation into child preferences regarding the linguistic register used by systems in generating spoken responses is suggested.
  • The study mentions ongoing work on "insistent requests" involving similarity scores of consecutive utterances, indicating a continued exploration of complex interaction patterns.

Summary updated

BibTeX
@inproceedings{arunachalam2001politeness,
  author = {Sudha Arunachalam and Dylan Gould and Elaine Andersen and Dani Byrd and Shrikanth S. Narayanan},
  title = {{Politeness and frustration language in child-machine interactions}},
  booktitle = {Interspeech},
  year = {2001},
  doi = {10.21437/eurospeech.2001-626},
}
Figure