Posted by alexandra_k on June 25, 2005, at 17:20:40
Perhaps it is impossible to attempt to come up with necessary and sufficient conditions for civility. Even if this is so, it doesn’t rule out the possibility of an automated civility checker. There has been success with training PDP / connectionist networks to perform functions that resist the sort of neat delineation that is required to write a traditional program. Connectionist networks can be simulated on an ordinary computer. They consist of a series of nodes. The nodes are arranged in layers (to constitute layers of processing). Each node has a threshold, and when that threshold is reached the node sends weighted connections to nodes in a subsequent layer.
http://www.mind.ilstu.edu/images/gnnv3layer.jpg
(Of course you need many more nodes than that, and more than one hidden layer. You would also want two output nodes ‘civil’ and ‘uncivil’ for civility determinations).
Training the network consists in feeding in information (posts) and seeing what determination the machine makes. If the determination is correct then a new example is fed in. When the machine makes an incorrect determination there are statistical algorithms for adjusting the node thresholds / weightings of the connections. The net preserves past determinations and adapts so as to produce the correct result on new determinations. The nets performance improves over time. What is interesting about such nets is that they have had a lot more success with areas in which traditional programs stumble. They are much better at being able to generalize to a novel instance of the same type of phenomena. Due to this they are better at tasks such as face recognition, object recognition, etc and one would expect that they would be better able to handle civility determinations.
Automated Asterisking
Currently there is a list of words that are considered uncivil. The automated system just matches words in the post with the words on the list. The automated system can’t recognize novel variants where there is an idiosyncratic mis-spelling (muthaf[*]cka) and it can’t recognize then a swear word is embedded in a greater word (motherf[*]cker). Nets would have better success with generalizing to new cases such as these. If a net was trained in one task (automated asterisking) then it is also capable of having new tasks superimposed over top of the old tasks (so long as you have enough nodes and weighted connections).
Anyway…
That’s what I have been thinking. I was a bit worried that you would need rules for translating posts into the first layer of nodes. I was worried that you would have to decide upon meaningful units (words – sentences – paragraphs) – but office mate says you don’t have to worry about this – you just let the net figure that out itself. Of course… It would be a pretty major project. You would need a pretty grunty computer to deliver determinations in anything approaching real time. It is possible in principle but at the moment it is probably a little like the thought of building a stainless steel ladder to the moon.But…
What I was thinking was that over time the net would probably get pretty good in its performance. What would be interesting is that when people hit to ‘submit’ their posts then at the confirmation step the civility checker could post a message ‘this post has been determined to be uncivil by the civility checker. Please reconsider your post. If you choose not to and Dr Bob considers this post to be uncivil then you choose to accept the consequences of your post’. If the poster posts their post unaltered then the machine could post a copy of its determination over on admin. If you agree with the determination then the poster has been warned already. If you do not agree then that’s ok, we could just ignore the machine (and alter the weightings so it won’t do that again). That way we would have an automated civility buddy too Of course such a process doesn’t help all that much with respect to coming to understand the civility rules / determinations. Even once you have designed a net it remains something of a mystery as to *why* or *how* it works. The net relies on a trainer to determine when the net is correct and when it needs to have its weightings adjusted. If it gets good at handling the easy cases, then we might be more inclined to trust it on the harder ones, however.
poster:alexandra_k
thread:518792
URL: http://www.dr-bob.org/babble/admin/20050614/msgs/518792.html