Language Modelling Professor's stimulating trip to Saarland, Germany

Language Modelling Professor's stimulating trip to Saarland, Germany

The Centre's Professor of Language Modelling Janet B. Pierrehumbert travelled to Saarland in south-western Germany recently for two speaking engagements relating to her research on the relationship between the dynamics of language and the structure of linguistic systems.

On 8th December she was the guest speaker at Saarland University’s IDeal Distinguished Speakers in Language Science Colloquium. IDeaL (Information Density and Linguistic Encoding) is a collaborative research centre that investigates the hypothesis that language use may be driven by the optimal use of the communication channel.

Professor Pierrehumbert gave a talk on the topic of ‘Remembering and generalizing from examples of words’: People learn words from experienced examples of them. The words in turn provide the statistical foundation for learning word-formation patterns. What information about words is encoded and remembered? How are generalisations formed from stored examples? When people encounter variable input, do they simply remember and reproduce this variation, or do they systematise it in their own outputs?

In this talk, Professor Pierrehumbert presented results from corpus analyses and online game-like experiments that address these questions [see Wordovators]. These indicate that mental representations of words include much detail. There is a great amount of individual variation in what information is encoded and the way abstract generalizations are formed. People have a propensity to systematise the input, but differ in what contextual associations they notice, how open they are to unexpected input, and how aggressively they systematise.

While in Saarland, Professor Pierrehumbert was also invited to speak at a language workshop organised by is a network of close interdisciplinary collaboration between linguists and computer scientists, funded by DFG, the largest independent research funding organisation in Germany.

The workshop (pictured) was aimed at models of language use, the speech production–perception loop and phonetic forms. The main theme of the event, held at Schloss Dagstuhl - Leibniz Center for Informatics in Germany, was exemplar-theoretic models and alternative approaches in usage-based models of language and speech production/perception.

Professor Pierrehumbert's talk discussed hybrid exemplar models. Exemplar theory was a response to findings that the mental representations of words and word phrases include a lot of phonetic and contextual detail. The initial computational models made the assumptions that 1) Words are associated directly with parametric phonetic representations and 2) Every experience with a word caused the memory representation to be updated. In her talk, Pierrehumbert presented findings that require more complex assumptions.

She concludes that a hybrid exemplar model is needed, combining much of the detail present in the first-generation models with the phonological parsing present in classical models of sound structure.