Language in humans Today: how do humans process language? - PDF document

Language in humans Today: • how do humans process language? Language in Humans We ’ ve looked above at syntactic processing. There are many other aspects of apparently similar complexity. Human Communication 1 • The main questions are Lecture 16 – the overall organization of the processing, and – the extent to which we can observe modularity. knowledge have to interact. 19/02/09 Susen Rabold 1 19/02/09 Susen Rabold 2 Further points A very simple model • of the processing of speech: We ’ ll see This is highly modular: • how we can use ambiguity to probe • process pressure waves to produce “sounds” certain aspects of the process and • process sounds to produce words • how different sources of knowledge • process words to produce trees have to interact. • process trees to produce DRSs We have already seen that it ’ s too simple in one respect: there are too many interpretations because ambiguities multiply. 19/02/09 Susen Rabold 3 19/02/09 Susen Rabold 4 1

An experiment (a) An experiment (b) • Do we retrieve all possible meanings • Imagine you are conducting an (and syntactic categories) for the words experiment and ask the participants that we hear? CROSS-MODAL priming whether they see a word of English - at allows us to investigate this question. times (1) and (2) present the words hand, tree and other unrelated words Present the following aurally: visually - and note their reaction time • The gypsy read the man ’ s palm (1) (2) (RT). for only a dollar because he was broke. 19/02/09 Susen Rabold 5 19/02/09 Susen Rabold 6 An experiment (c) Conclusions of experiment • We process the sounds to compute all • Your results would be: possible syntactic categories and (1) RT for related words (e.g. hand and tree ) is meanings of those words. These are the same, and quicker than for unrelated then filtered (within a matter of tenths of words. seconds) according to what is (2) RT for hand is quicker than RT for tree and syntactically and contextually unrelated words. The same result is obtained appropriate. Note that this involves the for words of differing syntactic categories interaction of different kinds of (e.g. tyres : wheels vs. wears out ). information. 19/02/09 Susen Rabold 7 19/02/09 Susen Rabold 8 2

And the moral is Shadowing (a) Morals: • We process language fast! Some can listen to a spoken text and repeat it • just because we can ’ t introspect and back with a delay of less than .5s. “feel” the multiple interpretations, that doesn ’ t mean they ’ re not there. • They correct errors (e.g. cigaresh for “cigarette”) as they go. Try it yourself. • at least some part of the human language processor operates in parallel. 19/02/09 Susen Rabold 9 19/02/09 Susen Rabold 10 Shadowing (b) Serial vs. parallel (a) Choices: • We ’ ve seen some evidence of parallelism at the level of words. What about other aspects • opt for a serial syntactic model which always of the process? makes a choice • Garden path sentences provide some • weaken modularity and allow some interesting evidence: interaction between processes. – The man who hunts ducks out on weekends One version of the second is to allow semantic – The cotton shirts are made from grows in and discourse information to rule out Mississippi analyses. This leaves us with a paradox. How – The old train the young can we know what words to retrieve if we – The daughter of the king ’ s son loves himself correct words as we go along? 19/02/09 Susen Rabold 11 19/02/09 Susen Rabold 12 3

A definition of Garden Path Serial vs. parallel (b) • Garden Path sentences = Sentences • Introspection suggests that we “get stuck”. that lead the human sentence • We could model this by our first choice processor (HSP) to construct an initial above: only ever work on one analysis, and have a rule for deciding which syntactic rule syntactic structure, which turns out to to use. be incorrect and thus requires syntactic • There is no or only limited possibility of (and semantic) re-analysis. revision of a choice. 19/02/09 Susen Rabold 13 19/02/09 Susen Rabold 14 Serial vs. parallel (c) Serial vs. parallel (d) • On the other hand, there are some • We can replace the . . . with any sentences in which we seem to be able amount of material. But that doesn ’ t to avoid having to make an early seem to induce the same kind of hiccup choice: as garden path sentences. – Have the police . . . eaten their supper? • One possible conclusion from this – come in and look example is that humans can make around. some use of ambiguity in processing. – taken out and shot. 19/02/09 Susen Rabold 15 19/02/09 Susen Rabold 16 4

Eye tracking your participants Looking at cards • Use an eye tracker to work out where their • Experiment: ask people to move playing attention is. cards around a table. • Results: by the time, (1) is reached, the • For example: 4 ♠ 6 ♣ 6 ♥ 3 ♦ K ♣ subject ’ s gaze alights on the 3 ♦ . • Instruction: “move the six of clubs to • Conclusion: as with shadowing, people can allow their knowledge of the context to allow beneath the three of diamonds (1).” early processing of semantic and discourse information. In this case, the uniqueness of the 3 allows the subject to work out which – (1) is space where card is being moved to card is being referred to. 19/02/09 Susen Rabold 17 19/02/09 Susen Rabold 18 Eye tracking Solitaire Summary • The mechanisms by which humans process speech are still the subject of great controversy. • http://www.youtube.com/watch?v=XwM We can say: oAqgikRM&feature=related – they involve some parallelism, certainly at the level of words, and perhaps at the level of syntactic analyses – they don ’ t conform to the strictest version of modularity – and there can be relatively large-scale interaction between different sources of knowledge. 19/02/09 Susen Rabold 19 19/02/09 Susen Rabold 20 5