Lecture: Computational approaches to morphological relatedness: a pilot study on paradigm discovery

Why do natural languages develop morphological structure? Is human cognition biased toward a lexicon that is morphologically structured? And if so, how easily does a learner achieve morphological generalization, that is, how easily may a paradigm be established in a system given some set of cognitive constraints?
Questions such as these can be fruitfully addressed by using computational methods, which provide a highly parametrized test ground for any models of cognition. Importantly, in this case, they allow a great degree of control over crucial confounding factors, such as, for instance, structural bias led by the specific typology of the native language of human subjects in experimental conditions.
In this ongoing project, led by Denis Paperno and Yoad Winter at Utrecht University, we implement a modified skipgram model using FastText word embeddings. Word embeddings allow us to represent both form and meaning using vectors, which are assigned to text words with respect to their distributional properties. In the case of this model, the distributional properties of subword n-grams are take into account to generate the pre-trained word vectors we will use. The goal of this project is to use iterations in order to test the extent to which an ideal learner is able to generalize a morphological paradigm from a limited set of examples in context, thus extracting systematic morphological relations between strings in the input. We plan to use vector cosine similarity, among other test methods, to assess how close together words of a same paradigm are represented. That is, we test for ease of paradigm discovery in an ideal learner, assuming only distributional probabilities of morphemes in the input.
The project is in its early stages, and in this talk we present its methods, advancements, expected results, and current challenges.

Info

Day: 2020-05-23
Start time: 13:30
Duration: 00:30
Room: De Saussure
Track: Computational Linguistics
Language: en

Links:

Feedback

Click here to let us know how you liked this event.

Concurrent Events