TL;DR
This paper introduces ABS₁, an artificial belief system designed to engage in dialogues with humans, learn from them, and establish credibility for both information and its sources. The authors explore how such a system can be used to study belief change and the dynamics of credibility in a controlled, artificial environment.
What Problem Does This Paper Address?
The paper tackles the challenge of understanding and simulating human belief systems, particularly the processes of acquiring information, assigning credibility, and how these beliefs change over time. Traditional experiments with humans are complex and difficult to control. By creating an artificial belief system (ABS₁), the authors aim to simplify this problem, allowing for precise study of credibility functions and belief dynamics under controlled conditions. They are interested in how an artificial system, exposed to human dialogues, would form beliefs and how its credibility assignments would evolve.
Methods
The core of the paper is the description of ABS₁, an artificial belief system implemented as a computer program in MLISP. ABS₁ interacts with humans through dialogues, accepting input in specific syntactical forms: statements, rules (implicational statements), and questions.
- Statements: These are the basic units of information, structured as Subject-Field, Verb-Field, Modifier-Field, ending with a period. ABS₁ treats these as facts of observation.
- Rules: These represent expectancies or implications, formatted as
X implies Y, where X and Y are expressions containing variables. These rules are used for inferring new conclusions. - Questions: ABS₁ can answer questions by searching its knowledge base of statements and rules.
- Credibility Assignment: ABS₁ uses a weighted additive function to assign credibility to propositions. This function considers the "foundation" (weighted average of credibilities of beliefs implying the proposition) and "consistency" (converse of foundation). A preliminary credibility (PCRED) is also assigned, initially high for a trusted source.
- Learning and Inference: During "Talktime," ABS₁ absorbs statements and rules. During "Thinktime," it uses inference processes (forward and backward chaining) to generate new statements and re-evaluate credibilities. It also engages in a "Questiontime" phase to elicit new rules from the informant.
Datasets
Not discussed in the paper. The system learns from direct interaction with human informants during dialogues.
Key Insights and Ideas
- Artificial Systems as Experimental Tools: Creating an artificial system like ABS₁ offers a powerful way to study complex cognitive processes like belief formation and change, overcoming the limitations of human experimentation.
- Credibility as a Dynamic Process: Credibility is not static but is dynamically assigned and updated based on incoming information, consistency with existing beliefs, and the perceived credibility of the source.
- Foundation and Consistency as Core Components: The paper highlights "foundation" and "consistency" as key factors in determining the credibility of a proposition.
- The Role of Source Credibility: The initial credibility assigned to a source significantly influences how the system processes and accepts new information, mirroring human child development.
- Polarization of Beliefs: ABS₁ tends to avoid neutral states, striving to polarize its beliefs towards either acceptance or rejection.
- "Self" as an Informant: The system can generate conclusions attributed to "self," treating its own inferences as a form of input, leading to interesting self-referential dynamics.
Key Results
The paper primarily describes the architecture and processes of ABS₁. While it doesn't present quantitative results from extensive experiments, it demonstrates:
- The ability of ABS₁ to accept statements and rules, answer questions, and infer new facts.
- The mechanism for assigning credibility based on foundation and consistency.
- The process of eliciting new rules from informants through question-answering.
- The potential for the system to develop its own "beliefs" and assign credibility to its own conclusions.
Limitations and Future Work
- Limited Natural Language Understanding: The syntax of input is restricted to specific forms to avoid complexities of natural language.
- Conceptual Content of Rules: While the inference process is logically valid, the empirical correctness of conclusions depends on the conceptual content of the facts and rules, which can lead to errors if based on flawed human input.
- Future Work: The authors plan to address questions about how ABS₁ changes as a result of experience with different human sources, and what factors influence this change. They intend to answer these questions in a future report.
Summary updated