This is my presentation at SQUID 2014 introducing my model for language acquisition that is based mainly on templates. This started with an observation of the patterns used by my son as he was learning to speak. I included a brief survey of other areas in linguistics which also make use of templates; i.e. Information Extraction and Machine Translation.
2. 1- Garbage Disposal
Yaseen Ghaly, 3 Years Old:
• Papy laih zebala zebala? (Dad, why
garbage garbage?) > Why are you
carrying two garbage bags?
• Papy laih bang bang? (Dad, why bang
bang?) > Why are you making this “bang
bang/hammering” sound?
3. 2- Making a Template
• Yaseen seems to be using this template:
• “Papy, laih X?” (Dad, why X)
• Where X can be anything:
– Garbage Garbage
– Bang Bang
– Sleeping
– …
– 天空是绿色的 (foreign word/code switching)
4. 3- To build a language
• How can the linguistic expression from
simple sentences into language such as
ours?
• Answer: Recursion
6. Main Claim
• Language is built using simple (idiomatic)
templates. The complexity comes from
recursion.
7. Outline
• Starting Assumptions
• Learning templates (Language Acquisition)
• Cross Linguistic Template Linearity
• Selecting A template (Semantic-Pragmatic
Prompt)
• Extending A template (Template Malleability)
• Applications of Templates (Information
Extraction and Machine Translation)
8. Starting Assumptions - Syntax
• In the syntax literature, language is a
lexicon of words, and a computational
system to put these words where they
should form a grammatical sentence.
Lexicon
Computation
System
9. Starting Assumptions – Templates
Framework
• The “lexicon”, which is stored in the memory, is extended
with a list of templates, also stored in memory.
• The computational system only manages what to fill the
placeholders within templates.
Word Lexicon Computation
System
Template Lexicon
10. Learning Templates
• The “garbage gabage” example indicates:
– A child can intuitively form a template for
plurals (that is applicable in some human
languages such as Bhasa Malaysia (e.g.
kanak kanak=children)
– A child can put anything in the placeholder X
within the sentence template “Dad, Why X?”
• But these hypotheses would need further
evidence from First Language Acquistion
11. Template Linearity
• English
– I love you.
– I miss you.
– I need you.
• French
– Je t’aime.
– Tu me manques.
– J’ai besoin de toi.
Clearly, the linear order is very different between Constructions in
different languages.
This should entice us to think about how these constructions are
generated.
12. Semantic-Pragmatic Prompt
• An area of overlap between the reason, context,
and information content of some sentence.
• Start with list of arguments (X1: I, X2: You)
• I Want to express [+feeling] [+positive]
[+distance], therefore:
– in English, we invoke the template I miss X2.
– In French, we invoke the template X2 me manques
(with some adjustments depending on pronouns, etc)
• So I can utter the sentence after filling the
template:
– I miss Randa.
– Randa me manque.
13. Template Variability
• Almost everything can be said in an alternative
way:
– Godzilla destroyed the City, which is unfortunate.
– It is unfortunate that Godzilla destroyed the city.
– The destruction of the city by Godzilla is unfortunate.
• So, there are different templates to express the
relation between these four entities (being
unfortunate, the destruction, Godzilla, the City).
This again feeds into the argument of non-
linearity of templates, this time within the same
language.
14. Template Malleability
• Meaning how easy the template can be re-
shaped. This includes the following:
– Tense malleability:
• John was eating fish.
• John has been eating fish.
– Synonym malleability:
• Sarah cannot tolerate this any more.
• Sarah cannot put up with this anymore.
• The idea of malleability enables us to avoid
accounting for hundreds of millions of
combinations of basic templates.
15. Using Templates
• For information Extraction (e.g. Banko and
Etzioni 2008), where templates where
used to extract (is-a) relationships
between entities.
16. Using Templates in Machine
Translation
• Was first suggested by (Nagao, 1984)
under the name of Example-Based
Machine Translation. He also indicted this
approach is relevant to Second Language
Acquisition.
17. Using Templates in Machine
Translation
• Current state of the
art Phrase-Based
Statistical Machine
Translation
techniques uses
contiguous chunks.
(Koehn, 2010)
18. Using Templates in Machine
Translation
• But using contiguous chunks misses many
phrases where there is a difference in
word order between the two languages.
- needs a lot of training data
• To compensate for this, a statistical
reordering model is used
- can make the output unintelligible
19. Using Templates in Machine
Translation
Chunk:
Michael assumes that he will stay in the house ->
Michael geht davon aus, dass er im haus bleibt
Subchunks:
Michael -> Michael
in the house -> im haus
So by removing (stenciling) subcunks from the chunk we get a translation template
X1 assumes that he will stay X2 ->
X1 geht davon aus, dass er X2 bleibt
- preserves word order
- can apply to many sentences not seen before
- requires less training data
- can set restrictions on the type of placeholders (X1: NP , X2: PP)