A Minimal Computational Improviser Based on Oral Thought

Nick Montfort, Sebastian Bartlett FernandezNAACL 2022

TL;DR

This paper introduces Chain Reaction Time (CRT), a computational model designed to play the improvisational word game "Chain Reaction." The authors ground their work in Walter Ong's theories of oral culture, aiming to create a system that embodies the principles of oral poetics and improvisation, offering a unique approach to computational creativity.

What Problem Does This Paper Address?

The paper addresses the challenge of computationally modeling oral improvisation, specifically within the context of a simple word game. While existing work has explored computational improvisation, particularly in areas like freestyle rap, the authors argue that these systems often lack a deep grounding in the principles of oral culture as theorized by scholars like Walter Ong. They aim to bridge this gap by developing a system that not only plays the game but also embodies the characteristics of oral thought and expression, such as additive structure, aggregation, and situational context. The goal is to explore how computational models can illuminate the nature of orality and inform the design of more sophisticated improvisational systems.

Methods

The core of the paper describes the development of Chain Reaction Time (CRT), a prototype system for playing Chain Reaction. CRT operates by maintaining a knowledge base of valid word pairs. This knowledge base is initially populated by a learning mechanism that extracts word pairs from textual corpora, with a preference for casual spoken-word corpora like the Santa Barbara Corpus of Spoken American English to better reflect oral practices.

CRT's system architecture includes:

  • Learning Mechanism: Extracts word pairs from corpora.
  • Pairs Model: Treats words and symbols as oral tokens, allowing for flexibility in handling homophones and phonetic variations, drawing from the CMU Pronouncing Dictionary.
  • Player Interface: Initially a text-based interface (OralTyping) using the Python curses library to emulate the ephemerality of spoken word. Later, a text-to-speech module (pyttsx3) was added to allow CRT to speak its continuations.

The authors emphasize that CRT is designed to embody principles of oral poetics, such as additive structure (continuing the chain by adding one word) and aggregative tendencies (forming "things" or meaningful pairs).

Datasets

The paper mentions the use of the Santa Barbara Corpus of Spoken American English as a primary source for building the word pair knowledge base. This choice is deliberate, aiming to capture the nuances of casual spoken language and reflect orally-based cognition. The CMU Pronouncing Dictionary is also referenced as a resource for handling phonetic representations of words.

Key Insights and Ideas

  • Orality as a Design Principle: The paper strongly advocates for grounding computational improvisation in the principles of oral culture, as articulated by Walter Ong. This approach offers a distinct perspective compared to systems that might rely solely on large textual corpora or purely statistical models.
  • Minimalism for Insight: By focusing on a "minimal improvisational game" and a "minimal computational improviser," the authors aim to reveal which oral poetic design principles are most broadly applicable and easier to model.
  • The "Thing" as a Unit of Orality: The concept of a "thing" – a meaningful two-word phrase or compound word that forms a coherent pair – is central to the game and the system's understanding of oral expression. This "thing" is determined by consensus, mirroring the fluid nature of oral understanding.
  • Emulating Ephemerality: The challenge of representing the ephemeral nature of spoken word in a textual interface is highlighted, with the OralTyping interface aiming to capture this aspect through timing and visual presentation.
  • Computational Models as Tools for Understanding: The authors view building computational models not just as an engineering task but as a method for exploring and understanding complex phenomena like orality, serving as a "reality check" for theoretical concepts.

Key Results

The paper does not present extensive quantitative results from formal evaluations. Instead, it focuses on the qualitative success of CRT in playing the game and embodying oral poetic principles.

  • Functional Play: CRT successfully participates in Chain Reaction games with human players.
  • Embodiment of Oral Principles: The system demonstrates several characteristics of oral poetics, including additive structure and the formation of "things."
  • Noticeably Nonhuman Play: The authors acknowledge that CRT's pace of recitation is unusually regular, and its continuations are not yet influenced by human context or earlier parts of the chain in the way human players' are. This highlights the limitations but also the unique character of the computational player.
  • Challenges in Modeling: The paper identifies key challenges in modeling oral improvisation, particularly those related to physical, temporal, and social contexts, as well as the fallibility of human memory.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Limited Corpus and Knowledge Base: The current knowledge base is relatively small, leading to situations where CRT cannot find continuations that human players can. Expanding corpora and improving the learning mechanism are key areas for improvement.
  • Handling Solid Compounds: The current learning mechanism struggles to detect valid pairs that are written as solid compounds (e.g., "signpost"). This reflects a fundamental difference between oral and textual perception of words.
  • Fallible Memory: CRT currently lacks a model of fallible memory, meaning it only fails by being unable to find a continuation. Implementing a temporal model where the system can forget parts of the chain is a planned enhancement to better emulate human play.
  • Cultural and Individual Differences: The system does not currently embody cultural or individual differences in player responses. Future work could involve creating custom pairs databases with weighted continuations to model these variations.
  • Contextual Awareness: Modeling the influence of immediate context (physical location, social setting, etc.) on gameplay is a significant challenge and a focus for future research.
  • More Elaborate Improvisation: The authors envision relaxing constraints on pair formation and exploring more complex improvisational forms, such as rhymed couplets for minimal rap cyphers.
  • Improved Interface: Developing a two-way verbal interface with advanced speech synthesis and recognition is a long-term goal.

Summary updated

BibTeX
@inproceedings{montfort2022minimal,
  author = {Nick Montfort and Sebastian Bartlett Fernandez},
  title = {{A Minimal Computational Improviser Based on Oral Thought}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure