Language Learning in Humans and Machines - Making Connections to Make Progress

Explore language learning in humans and machines through this insightful 46-minute talk by Dr. Sharon Goldwater, a Reader at the University of Edinburgh's School of Informatics. Delve into the challenges of natural language processing, the importance of understanding human language acquisition, and the potential for incorporating human-like learning biases into computational systems. Discover how this approach could address the resource-intensive nature of current language processing methods and expand their applicability to a wider range of the world's languages. Examine examples from Dr. Goldwater's research, including word segmentation and meaning as translation, to gain a deeper understanding of the complexities of language learning and the innovative solutions being developed in the field.

Intro
Today's talk
Interpreting a sentence
The real situation
Other examples of ambiguity
Labelled examples
Extracting features
Statistical natural language processing
Labelled data is hard to obtain
Result: unequal access
Human language learning
Must computers learn like humans?
But language isn't "in the world"
Research programme
Learning biases
Stronger bias = less data
Example problem: segmentation
Word segmentation
Statistical learning experiment
Testing for learning
How do they do it?
What about real language?
Another strategy
A model for segmenting words
The right bias can help
The Dirichlet process model
Output of the system
Words aren't marbles
Improved system
Where else can these ideas help?
Continuing work
Meaning as translation
Results so far
Conclusions
Acknowledgements