Title: A Dynamic Learning Model For Categorizing Words Using Frames
1A Dynamic Learning Model For Categorizing Words
Using Frames
- Hao Wang, Toben Mintz
- Department of Psychology
- University of Southern California
2The Problem of Learning Syntactical Categories
- Grammar includes manipulations of lexical items
based on their syntactical categories. - Learning syntactical categories are fundamental
to the acquisition of language.
3The Problem of Learning Syntactical Categories
- Nativist approach
- Children are innately endowed with the possible
syntactical categories. - How to map a lexical item to its syntactical
category or categories? - Empirical approach
- Children have to figure out the syntactical
categories in their target language, and assign
categories to lexical items. - There is no or little help from syntactical
constraints.
4Approaches Based on Semantic Categories
- Grammatical Categories correspond to
Semantic/Conceptual Categories(Macnamara, 1972
Bowerman, 1973 Bates MacWhinney, 1979 Pinker,
1984)object ? noun action ? verb - But what about
- action, noise, love
- to think, to know(Maratsos Chalkley, 1980)
5Grammatical Categories from Distributional
Analyses
- Structural LinguisticsGrammatical categories
defined by similarities of word patterning
(Bloomfield , 1933 Harris, 1951) - Maratsos Chalkley (1980) Distributional
learning theory - lexical co-occurrence patterns
- (and morphology and semantics)
- the cat is on the mat
- cat, mat
6Grammatical Categories from Distributional
Analyses
- Patterns across whole utterances(Cartwright
Brent, 1997) - My cat meowed.
- Your dog slept.
- Det N X/Y.
- Bigram co-occurrence patterns(Mintz, Newport,
Bever, 1995, 2002 Redington, Chater Finch,
1998) - the cat is on the mat
7Probabilistic Bigram Co-Occurrence Patterns
8Frequent Frames (Mintz, 2003)
- Frames are defined as two jointly occurring
words with one word intervening. - would you put the cans back ?
- you get the nuts .
- you take the chair back .
- you read the story to Mommy .
- Frame you_X_the
9Sensitivity to Frame-like Units
- Frames lead to categorization in adults (Mintz,
2002) - Fifteen-month-olds are sensitive to frame-like
sequences (Gómez Maye, 2005)
10Other Motivation for Frames
- Verb learning in children can be facilitated by
frequent frames (Childers Tomasello, 2001) - Aspects of verb meaning carried by verb frame,
linguistically defined (Gleitman, 1991 Gillette,
Gleitman, Gleitman, Lederer, 1999 etc.)
11Distributional Analyses Using Frequent Frames
(Mintz, 2003)
- Six corpora from CHILDES (MacWhinney, 2000).
- Analyzed utterances to children under 26.
- Accuracy results
- averaged overall corpora.
12Limitation of the Frequent Frame Analyses
- Requires two passes through the corpus
- Step 1, identify the frequent frames by tallying
the frame frequency. - Step 2, categorizing words using those frames.
- Tracks the frequency of all frames
- E.g., approximately 15000 frame types in one of
the corpora in Mintz (2003).
13Goal of current study
- Provides a psychological plausible model of word
categorization - Children possesses limited memory and cognitive
capacity. - Human memory is imperfect.
- Children may not be able to track all the frames
he/she has encountered.
14Features of current model
- It processes input and updates the categorization
frames dynamically. - Frame is associated with and ranked by a
activation value. - It has a limited memory buffer for frames.
- Only stores the most activated 150 frames.
- It implements a forgetting function on the
memory. - After processed a new frame, the activation of
all frames in the memory decreased by 0.0075.
15Child Input Corpora
- Six corpora from CHILDES (MacWhinney, 2000).
- Analyzed utterances to children under 26.
- Peter (Bloom, Hood, Lightbown, 1974 Bloom,
Lightbown, Hood, 1975)Eve (Brown, 1973) Nina
(Suppes, 1974)Naomi (Sachs, 1983)Anne
(Theakston, Lieven, Pine, Rowland, 2001)Aran
(Theakston et al., 2001) - Mean Utterance/Child 17,200
- MIN 6,950 MAX 20,857
16Procedure
- The child-directed utterances from each corpus
was processed individually - Utterances were presented to the model in the
order of appearance in the corpus - Each utterance was segmented into frames
- you read the story to Mommy
- you read the
- read the story
- the story to
- story to Mommy
17Procedure continued
- you read the
- read the story
- the story to
- story to Mommy
18Procedure continued
- The memory buffer only stores most activated 150
frames. - It becomes full very quickly after processing
several utterances.
19Procedure continued
- you put the
- Frame you_X_the
- Look up you_X_the frame in the memory
- Increase the activation of you_X_the frame by 1
- Re-rank the memory by activation
20Procedure continued
- you have a
- Frame you_X_a
- Look up you_X_a frame in the memory
- story_X_Mommy lt 1
- Remove story_X_Mommy
- Add you_X_a to memory, set the activation to 1
- Re-rank the memory by activation
21Procedure continued
- A new frame not in memory
- The activation of all frames in memory are
greater than 1 - There is no change to the memory.
22Evaluating Model Performance
- Hit two words from the same linguistic category
grouped together - False Alarm two words from different linguistic
categories grouped together - Upper bound of 1
23Accuracy Example
- Hits 10
- False Alarms 5
- Accuracy
24Ten Categories for Accuracy
- Noun, pronoun
- Verb, Aux., Copula
- Adjective
- Preposition
- Adverb
- Determiner
- Wh-word
- Negation -- not
- Conjunction
- Interjection
25Averaged accuracy across 6 corpora
26The Development of Accuracy
- Accuracy are very high and stable in the entire
process
27Compare to Frequent Frames
- After processing about half of the corpus, 70 of
frequent frames are in the most activated 45
frames in memory.
28Memory of Final Step of Eve Corpus
29Stability of Frames in Memory
- Big changes of frames in memory in early stage,
but become stable after processing 10 of the
corpus
30Summary
- After processed the entire corpus, the learning
algorithm has identified almost all of the
frequent frames by highest activation. - Consequently, high accuracy of word
categorization is achieved. - After processing fewer than half of the
utterances, the 45 most activated frames included
approximately 70 of frequent frames.
31Summary
- Frames are a robust cue for categorizing words.
- With limited and imperfect memory, the learning
algorithm can identify most frequent frames after
processing a relatively small number of
utterances. Thus yield a high accuracy of word
categorization.