|
| Model name: | Wikipedia | Model title:
| Simulating Collaborative Writing in EMIL-S: Software Agents Produce a Wikipedia. | Replicated model?
| Yes | | Keywords: | Norm emergence, normative agent, collaborative writing | Model authors:
| Ulf Lotzmann, Robin Emde and Klaus G. Troitzsch
| | Programming language? | Java / Repast / EMIL-S |
This model replicates a model described in Troitzsch (2008). Findings of an empirical analysis of the behavior of contributors to and discussants of Wikipedia articles are used to build a simulation model of norm emergence in collaborative writing. DOCUMENTATION 1. Purpose The purpose of the model is to analyse the processes of norm emergence in collaborative writing. Therefore software agents were endowed with the capability to produce text in an artificial language and to evaluate something like the "style of writing" in this language, thus being able to take offence at certain features of texts and blaming the authors of such text. From this kind of communication, norms emerge in the artificial society of software agents. 2. State variables and scales In this model all agents are instances from the class WikiAgent. Each agent is able to perform the following activities (initialized uniform distributed): - write an article and either submit it or add it to an existing article referring to the same keyword,
- plagiarize, i.e. copying an existing article and publishing it for a new keyword,
- search the current state of the encyclopedia for double entries, for words that do not obey the vowel harmony or for plagiarisms , and reproach the respective author or authors,
- count articles that contain a word about which they wrote an article,
- do nothing.
3. Process overview and scheduling
Each agent decides whether to submit an article / add new content to an existing article, to comment on existing articles or to search for similarities between articles.
In the latter case agents search for similarities between two randomly selected articles. If an agent detects a co-occurrence of more than a certain percentage of words between two articles, it identifies the younger of the two articles as a potential plagiarism and blames its author who loses the profit generated before from intentional plagiarism. If the similarity is just by chance, the same hard punishment occurs.
When an agent finds itself in a situation where it could contribute to the encyclopedia it consults its memory to find how profitable the different actions available in this situation might be.
Based on receiving payoffs, the action selection probabilities will be adapted.
4. Design concepts
4.1 Emergence Emergence of norms in the artificial society of Wiki-Agents about vowel harmony, double entries and/or plagiarism in their Wikipedia. 4.2 Adaptation Agents adapt their behaviour by changing their action preferences, based on received feedback from the environment or received messages from other agents (punishment). Beside learning from own experience about payoffs and from reproaches, agents can in principle also learn from observing other agents’ actions and the resulting payoffs. 4.3 Fitness The fitness of an agent relates to his norm compliant behaviour. 4.4 Prediction No prediction in this model. 4.5 Sensing Agents are able to scan all articles in the Wikipedia and to compare the entries with a given keyword. 4.6 Interaction Individuals interact indirect by the entries in the Wikipedia and direct by sending and receiving norm invocation messages. 4.7 Stochasticity The selection probabilities for actions are randomly initialized. Moreover is stochasticity in the process of writing an article: Writing an article starts with constructing a keyword out of the letters “aei bklsw” (including blank but not the full stop) where the probability of selecting a particular letter as the first letter in the word is equal (with the exception of the blank) whereas the probability of selecting the next letter depends on the previous letter, according to a stochastic matrix which is currently constant (but could as well change over time, according to the practice developing in the community). The blank character is selected with a certain probability, ending the construction of the word. The first word of an article is marked by a following colon as a keyword (and for some trivial technical reason it is preceded by the character “>”). The following words are constructed the same, and after each word a full stop is inserted with a certain (low) probability, such that the chain of words is separated into something like sentences. At the end of a sentence the article ends with another (low) probability.
4.8 Collectives No collectives of groups within the population. 4.9 Observation Results about vowel harmony, plagiarism and double entries in the Wikipedia articles are shown in output graphs during one simulation run (simulation environment Repast). 5. Initialization All agents are initialized with uniform distributed selection probabilities of actions and letters.
6. Input The input parameter of a simulation run are: - Number of agents. Default: 5 agents.
- Length of word list. Defines the maximum length of an article. Default: 2200.
- Threshold for plagiarism. Defines the level of similarity between two articles that identify one as plagiarism. Default: 0.75.
- Chance of non plagiarism. Default: 0.21.
7. Submodels |
