Uniform Information Density - Computational Linguistics and Phonetics

Report
Seminar „Information Theoretic Approaches to the Study of Language “
Redundancy and reduction:
Speakers manage syntactic
information density
T. Florian Jaeger (2010)
Torsten Jachmann
16.05.2014
So far
• Frequent words have shorter linguistic forms
(Zipf)
o Orthographic; PHONOLOGICAL
• Word length (phonemes/syllables) correlated
with predictability
• Information is context dependent
o The more probable, the more redundant
• More predictable instances of the same word are
produced shorter and with less phonological and
phonetic detail
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
• Morphosyntactic:
o E.g. auxiliary contractions
“he is” vs. “he’s”
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
• Syntactic:
o E.g. optional that-mentioning
“This is the friend I told you about”
vs.
“This is the friend that I told you about”
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
• Elide constituents:
o E.g. optional argument and adjunct omission
“I already ate”
vs.
“I already ate dinner”
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
• Production planning:
o E.g. one or more clauses
“Next move the triangle over there”
vs.
“Next take the triangle and move it over there”
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
• Other languages?
o German:
“Er hat es verstanden” vs. “Er hat’s verstanden”
(he understood it)
o Japanese:
“行ってはダメ” vs. “行っちゃダメ”
Itte ha dame
Iccha dame
(you can’t go)
Idea
Speakers manage the amount of information per
amount of linguistic signal (at choice points)
The form with less linguistic signal should be less
preferred whenever the reducible unit encodes a
lot of information (in the context)
Uniform Information Density
(UID)
Optimal:
• On average each word adds the same amount of
information to what we already know
• The rate of information transfer is close to the
channel capacity
❌ Many constraints (grammar; learnability)
Uniform Information Density
(UID)
Efficient:
• Relative uniform information distribution where
possible
• No continuous under- or overutilization of the
channel
?
UID
Definitions:
• Information density:
Information per time
(articulatory detail is left out)
• Choice:
Subconscious
(existence of different ways to encode the intended message)
an Jaeger
UID
Example:
Page 41
UID
Example:
Fig. 1.
UID
Goals:
• UID as a computational account of efficient
sentence production
• Corpus-based studies are feasible and desirable
 Corpus of spontaneous speech
 Naturally distributed data
Data
• 7369 automatically extracted complement
clauses (CC) from “Paraphrase StanfordEdinburgh LINK Switchboard Corpus” (Penn
Treebank)
• - 144 (2%) falsely extracted
• - 71 (1%) rare matrix verbs  extreme
probabilities
Data
Focus
Actually:
I(CC onset|context) =
-log p(CC|context) + -log p(onset|context,CC)
Here:
I(CC|context) = -log p(CC|matrix verb lemma)
Data
Multilevel logit model
• Various factors (might) influence the outcome
• Ability to include several (control) parameters in
one model
• Contribution of each can be estimated
Why?
• Natural (uncontrolled) data
Controls
Dependency
• Distance of matrix verb from CC onset  “THAT”
o My boss thinks [I’m absolutely crazy.]
o I agree with you [that, that a person’s heart can be changed.]
• Length of CC onset (including subject)  “THAT”
• Length of CC remainder
Short sidetrack
Length of CC remainder
• Language production is incremental
(+ heuristic complexity estimates?)
Florian Jaeger
Page 45
NIH-PA Author Manuscript
NIH-PA Author Manus
Fig. 5.
Observed proportions of that by CC length in words (limited to CCs up to 25 words); jittered
points are bottom and top of each cell represent individual cases; error bars indicate 95%
Controls
Availability
• Lower speech rate  “THAT”
• preceding pause  “THAT”
• initial disfluency  “THAT”
Controls
Availability
• Type of CC subject
o It vs. I
o Other PRO vs. above
o Other NP vs. above
• Frequency of CC subject head
• Subject identity
o Identical subject in matrix and CC  ≈ “NONE”
Controls
Availability
• Word form similarity
o Demonstrative pronoun “that”
o Demonstrative determiner “that”
≈ “NONE”
• Frequency of matrix verb
o Higher frequency  “NONE”
Controls
Ambiguity avoidance
• Possible garden path sentence  “THAT” ❌
-PA Author Manuscript
NIH-PA Author Manuscript
o Even unlikely cases were included
“I guess (this doesn’t really have to do with…)”
Table 4
Distribution of disambiguation points for potentially ambiguous CC onsets.
Disambiguation point – word:
1
2
3
4
5–9
>9
Number of instances
773
86
61
48
35
9
Cogn Ps
Controls
Matrix
• Position of matrix verb
o Further away from sentence-initial position  “THAT”
• Matrix subject
o You vs. I  “THAT”
o Other PRO vs. above  “THAT”
o Other NP vs. above  “THAT”
Controls
Others
• Random speakers intercept
• Persistence
o Prime w/o that vs. no prime
o Prime w/ that vs. above
• Gender
o Male  “NONE”
Information density
• Clear significance (p < .0001)
• High information density of the CC onset
 use of “that”
• Correlation with other predictors negligible
• Contribution to the model’s likelihood is high
• At least 15% of the model quality due to
information density
• Single most important predictor
Information density
• Verbs’ subcategorization frequency as estimate
for information density
• High CC-biases, low “that”-biases (e.g.: guess)
• Low CC-biases, high “that”-biases (e.g.: worry)
Syntactic reduction is affected by information
density
Results
Information density
• Prediction:
UID can account for any type of reduction
Phonetic and phonological reduction
• So far patterns align with this prediction
• Availability account do not predict this
o But predict lengthening of words
Information density
Optional case markers (or copula)
• Languages with flexible word order
• Japanese
ケーキが大好きだ vs. ケーキ大好き
Keeki ga daisuki da
I love cake
keeki daisuki
Information density
Reduced case markers
• Korean
나는 독일 사람이야
vs. 난 독일 사람이야
Na neun togil saram iya
I am German
Nan togil saram iya
Information density
Optional object clitics
and other argument marking morphology
• Direct object clitics in Bulgarian
o Can’t be predicted by availability account
o Could be predicted by ambiguity avoiding
Information density
Contracted auxiliaries
• English
“he’s” vs. “he is”
o Can’t be predicted by neither availability nor ambiguity
avoidance
Information density
Ellipsis
• Japanese
行きたいけど行けない vs.
ikitai kedo ikenai
行きたいけど
ikitai kedo
I want to go but (I can’t go)
(¬ 行きたいけど(遅くなりそう) I want to go, but I might be late)
ikitai kedo osoku nari sou
Information density
Non-subject-extracted relative clause
• Indefinite noun phrase < definite noun phrases
• Light head nouns < heavy head nouns
(e.g. the way)
(e.g. the priest)
“I like the way (that) it vibrates”
Information density
Whiz-deletion (BE)
• Relativizer + auxiliary can be ommitted
“The smell (that is) released by a pig or a chicken
farm is indescribable”
Information density
Object drop
• Verbs with high selectional preference
“Tom ate.”
vs.
“Tom saw …”
Information density
• Many novel predictions across
o Different levels of linguistic productions
o Languages
o Types of alternations
• Per-word entropy of sentences should stay
constant throughout discourse
o Words with high information density (in the context and
discourse) should come later in the sentence
o A priori per-word entropy should increase
Grammaticalization
• Might interfere with information density?
o Matrix subject “I” or “you”
o Matrix verb “guess”, “think”, “say”, “know”, “mean”
o Matrix verb in present tense
o Matrix clause was not embedded
3033 cases remain
• Still highly significant (p < .0001)
 UID may be a reason for grammaticalization
Noisy channel
• Base of UID
• Audience design
o Speaker considers interlocutors’ knowledge and processor state
to improve chance of successfully achieving their goal
• Modulating information density at choice points
= rational strategy for efficient production
• UID minimizes processing difficulties
Corpus-based research
Claim: “Lack of balance and heterogeneity of data
make findings unreliable”
• Multilevel models
• Avoidance of redundant predictors
• If redundant residualization
• Inter-speaker variance
+ ecological valid
Corpus-based research
• Results easier extend to all of English
• Many previous results replicate
• Provides evidence for so far relatively
understudied effects (e.g. similarity avoidance)
• “effect size” needs to be taken with caution
o Not only strength of effects but also applicability
Ambiguity avoidance (garden path sentences) relatively rare
Questions and discussion

similar documents