Lecture 9: Language and Meaning Flashcards
“Error in perception or interpretation of the heard utterance or the observed scene.”
What challenge of word learning does this definition refer to? Explain in your words. Give an example.
Noise
In most real-life situations, children don’t receive perfect data. They might not hear the entire sentence, or might not notice the fact the chimp is eating something.
Example: They hear not just “apple”, but “The chimp eats an apple”.
“Learners may perceive aspects of a scene that are unrelated to the utterance they hear.”
What challenge of word learning does this definition refer to? Explain in your words. Give an example.
Referential uncertainty
Even if a child knows what the sentence refers to, it still doesn’t necessarily know what it means. The adults might be saying “The monkey is sitting on the rock” or “The black animal is holding a red object” or even “What a beautiful day!” while pointing at the monkey.
Even if the child knows the meaning of the whole sentence, some referential uncertainty still remains
- The words might have multiple meanings.
- There might be multiple words for the same object.
What is the problem with fast mapping in bilingual children?
Bilingual children have the “disadvantage” to be aware of the fact that there are always multiple words that mean the same (“apple”, “manzana”, “apfel”, “mela”, “pomme”, …). When presented with a known and an unknown object together with an unknown word, they don’t directly tend to pick the unknown object because the unknown word could possibly be another word for the known object even though they already one word for it.
“Simple associative mechanisms are used to map a word form with a concept (simply connecting word + concept).” What kind of word learning mechanism does this definition refer to?
Associative learning (Challenging because of noise and/or referential uncertainty)
Trying to narrow down the possible definitions of a word using different attention mechanisms. What kind of word learning mechanism does that definition refer to?
Referential learning
Most words are not used in isolation, but in a multi-word utterance. What challenge of word learning does this definition refer to?
Sentential context.
“Inferring correct word-meaning mappings by observing regularities across usages of a word.”
What kind of word learning mechanism does this definition refer to? Example.
Cross-situational learning.
If the word ‘apple’ is regularly used when an apple is present, apple will be more likely to refer to that object than objects that co-occur less often with the word apple.
How is cross-situational learning computationally modeled? Where does the input come from? What is the output?
Input: pairs of sentences + a representation (image) of their meaning
How would we obtain this input data?
- Utterances: there are big databases (such as CHILDES) with child-parent conversations
- Representations: we try to approximate what the child sees (This is a very difficult task)
We can teach neural networks to combine images with speech. A neural network can find patterns
in the images and speech
- Output: A kind of lexicon with some knowledge about the meaning of every word (from the input)
What are 3 developmental patterns in language learning?
Vocabulary spurt
Fast mapping
(Difficulty with) Second labels
How is the suddenly fast word learning in infants of a certain age called?
Vocabulary spurt
What is the phenomenon that children can map a novel word to a novel object called?
Fast mapping
What is the problem of “second labels” in early word learning?
Early on, children show difficulty in learning homonymous (one word to many concepts) and synonymous (many words to one concept) words
What are the three approaches in simulating perception context?
[go to the lecture slides for visual clarification of the algorithms]
- Artificial word-meaning mappings
- Manually annotate child-adult interaction videos with symbolic representations of the context
- Synthetically construct semantic representations for real-world utterance
Artificial word-meaning mapping is one of three approaches for simulating perception context. Describe it.
Earliest model: Artificial word-meaning mappings (made-up, synthetic input as we simply didn’t have any data we could use as input back then; it was only used to show how those processes work in theory. All possible meanings were shown)
Manually annotate child-adult interaction videos with symbolic representations of the context is one of three approaches for simulating perception context. Describe it.
Manually annotate child-adult interaction videos with symbolic representations of the context. The references consisted only of representations of the visible objects in the scene, so objects the infant could SEE.
Example: utterance = “look at the book”. Meanwhile, the infant was seeing the book, but also a rattle and a ball. The algorithm took therefore as references {rattle, book, ball}
This approach was already better than the first one, but very time-consuming. The experimenter had to sit down, watch the videos one by one and write down every single sentence + the possible references
