# Exp 1 data brainstorm (2015-11-09) ## Notes from looking through the data (in a detailed fashion) Some changes are more continuous, some are more step-wise, some can be both. E.g. elision of "when you leave the office" happens in several ways: * branch 262: "when you leave the office" -> "when you're finished" -> "when you leave" * branch 264: "when you leave the office" -> "leave the office open when you leave" -> "leave the door open when you leave" -> "leave the door open" * branch 498: direct disappearance Those same transformations take differents paths to get to the same final form. Or different paths to achieve a given transformation. *That's attraction*. Sentences seem to reduce to a minimal content that is relevant (removing unnecessary explanations or complements). Some changes happen very fast, in the first or second transformation in nearly all branches (e.g. in #5, the disappearence of "you can hitch-hike...", #9 disappearence of "by taking all necessary steps"). Some forms are reached so that there are very little transformations for a long time, or at all. *That's a kind of stability* (e.g. #5 branch 252, #9 branch 207, #12 branch 302). For some sentences, a few branches reach a similar form. For other trees the branches really diverge in all directions (e.g. #10). *That looks like different types of trees*. A sizeable part of transformations comes from interaction problems, misunderstandings. A mistake or ambiguity in writing, corrected or interpreted in another way at the next step (e.g. #15, 2545->2761). (Similar things happen orally: when you listen to yourself you see all the ways in which you can be interpreted. Although orally, prosody and interaction gives more cues.) Sounds hard to expose this in the plots. ## Ideas for measures Measure distances between sentences (stopword-filter and lemmatize sentences, then take levenshtein-word distance, or intersection as bags of words). For each tree, downscale space to 2D (tree-dependent at first; could also be over all the trees). Look at spread of final sentences, and path to them. Detect * words that often appear, or often disappear * words that appear then disappear or vice-versa (with final score +-) What are those words? Plot word span/education level/age/gender/job w.r.t. raw transformation rate Check distribution of reading times: * see if data is trustworthy at all * calibrate fixed reading time * also check writing times to calibrate Over the transformations, positionning changes: who's talking, from where. Could codify the dominant person (1st, 2nd, 3rd) and plot that vs. depth for each branch for each tree. ## Ideas on things to fix * remove the "..." in the example * tell subjects to _not_ talk to the experimenter (e.g. "I don't know", "...", "I can't remember"), but to the next person in line. Add an example tree of interpretations with an arrow showing "you are here". * tell subjects to apply themselves to writing properly and cleanly (no chat, no sms, no IM). In fact, mimick the style. * remove the lower-limit on words, or hide it (it makes people fill sentences with garbage, or stop mid-sentence once the button turns active) * make sure what the subjects write makes sense. They must know+feel they're talking to someone, and try to correct any obvious mistakes in the sentence they're copying * ? add a flagging option on sentences to mark them as nonsensical, or incomplete or missing a part (once X people (e.g. 2) flag a sentence, it gets notified to author, and the branch stops) * ? or frame it as an acceptance/refusal of sentence (and refusals would be checked and malused if abused). * prevent people from rushing through * make payment depend on transformation rate and maluses * ? add a confirm/reread step * ? show them their performance on training sentences, just before launch. With diff and #characters changed. * ? add a person/profile page: my transformations, and once exp finished, other's transformations * make reading time fixed (don't let the subject click "next") * hide writing time: when at the end: "time's up" * request a higher education level, at least some good spelling * ? add a scoring system that live-updates. Payment is % to score. * a malus is extremely bad (like -X, and second malus divides score by 2) * abuse of malussing (hand-checked) is also very bad * dropping sentences (reading but not rewriting, without flagging) is bad * going faster or slower doesn't influence score: you have max X time to do it, and you're paid the time you used % score. * In the people pages, make a corresponding scoring sort/high scores