Milin Et Al.

• Towards cognitively plausible data science in language research (2016), Milin, Divjak, Dimitrijevic and Baayen
• Identify difficult and easy forms (from lemma to plural form)
• Check if human participants also react differently to independently identified difficult and easy forms Compare NDL learning model to TiMBL and human results
• MDVM computes the distance between two values of a feature to reflect their patterns of co-occurrence with categories
• Using MDVM adds an Unsupervised Learning component to MBL Hoste (2005) because essentially it clusters feature values and uses that information
• Using larger values of k with MDVM is helpful
• Easy words that are frequent tokens (forms) are reacted to faster
• Maybe this interaction doesn’t occur with difficult words because there is less variation in the frequency of the difficult words?
• This seems similar to results with regular past tense forms in English:
• Strikingly, TiMBL’s inflectional class probabilities turn out to be predictive in production and comprehension, i. e., for lexical decision latencies.
• Two Grapheme to Lexeme Measures
• Diversity Sum of the absolute values of the activations of all possible outcomes, given a set of input cues.
• Input cues that activate many different outcomes give rise to a highly diverse activation vector, which in turn indicates a high degree of Uncertainty about the intended outcome.
• G2L-Prior Sum of the absolute values of the weights on the connections from all cues to a given outcome.
• independent of the actual cues encountered in the input
• reflects the prior availability of an outcome, its entrenchment in the learning network
• TiMBL assigns higher probabilities to forms belonging to lemmas with letter trigraphs that yield more diverse activations
• Those trigraphs belong to a rich exemplar space in the memory
• it would be expected that higher probabilities would result in shorter response latencies
• However, NDL’s G2L-Diversity was in fact positively correlated with RTs, indicating inhibition, i. e. slower recognition.
• TiMBL probabilities are intended to capture the likelihood of a form’s occurrence in production.
• in comprehension (lexicality judgments) high trigraphs diversity may hurt results
• Spontaneous recovery from extinction
• After a CS is learned to associated with a given Conditioned Response (CR), this association is unlearned
• Theoretically, it can not arise again without retraining
• But in real life, sometimes seemingly completely forgotten associations are reactivated
• shows extinction is not unlearning
• responses that disappear are not necessarily forgotten
• Suggests loss of activation is not simply the mirror of acquiring associations
• Given two conditions stimuli, (CS) where one is more salient, the more salient CS will develop a strong association with the CR (Conditioned Response)
• Some linguistic things can be learned with NDL and this might show use something about the problem
• What made NDL so nice for animal learning might not scale up to linguistic phenomena
• Inductive approaches to cognition