semantic glimmers: csdl9
DESCRIPTION
Slides from a talk I gave in 2009 at Conceptual Structure, Discourse, and Language. Research presented was on contributions of semantic and phonological similarity to sentence comprehension. Also see related paper (Otis & Sagi, 2009).TRANSCRIPT
SEMANTIC GLIMMERS: PHONAESTHEMES FACILITATE
ACCESS TO SENTENCE MEANING
Eyal Sagi & Katya Otis, Northwestern University
Phonaesthemes: phonetic clusters that occur in words that have noticeably similar meanings (Firth, 1930)
Example: gl- associated with “words relating to light, vision” (glimmer, glisten, glow, glare, glance)
Violates Saussurean notion of “arbitrariness of the sign” Not sound symbolism
Non-compositional: not properly morphemes
Form and Meaning
Form and Meaning: Evidence Subtle regularities in word form can cue
syntactic category assignment
These regularities facilitate sentence comprehension Ambiguous sentences can be resolved faster
when “nouns sound like nouns; verbs sound like verbs”
Farmer, Christiansen, & Monaghan, 2006Monaghan, Chater, & Christiansen, 2005
Phonaesthemes: Previous Research Hutchins (1998)
Speakers reliably matched phonaestheme-bearing words to glosses of the phonaestheme and vice versa.
Speakers ranked phonaestheme-bearing words’ coherence with a definition higher when the definition was a gloss of that phonaestheme.
Found variability in strength of association between 46 phonaesthemes and their glosses. Probed speakers’ declarative knowledge about
language, not their processing or implicit knowledge Words in isolation Rely on linguists’ intuitions
Phonaesthemes: Previous Research Bergen (2004)
Morphological Priming paradigm Sharing a phonaestheme provided better
priming than either semantic or phonological similarity alone Rely on linguists’ intuitions to decide what
counts as a phonaestheme Words in isolation
Hypotheses
1. Corpus analyses can help reveal regular relationships between form and meaning
2. Speakers use these regular relationships in language processing
Words Sentences
Current Studies
Corpus analysis of phonetic clusters used by Hutchins (1998)
Experiment 1: Sentence completion task Experiment 2: Paraphrase task
Corpus Analysis: Method
Latent Semantic Analysis Words as vectors in a
semantic space Similar meaning
Similar vector direction Standard measure:
Cosine of angle = Correlation between the vectors
Semantic vectors can be combined
gl-(combined vector)
ray
glance
glisten
glare
vision
light’
Corpus Analysis: Method
If a cluster of words is semantically related, the respective word vectors will be more correlated than expected by chance.
Monte-Carlo analysis Compare combined vectors of word pairs within a
phonaestheme cluster with combined vectors of randomly-chosen word pairs.
Quantitative criterion for phonaestheme strength Measure: # of significant t-tests (p < .05)
If 15 of 100 t-tests conducted are significant, then the phonaestheme is statistically supported (overall p < .05)
Corpus Analysis: Input
Public domain literary works from Project Gutenberg 4034 documents Over 290 million words
50 candidate phonaesthemes 46 used by Hutchins (1998) 4 new clusters: br-, -ign, kn-, z-
Corpus Analysis: Results
27 of 50 phonaesthemes met our criterion for significance
Replicated Hutchins’ survey results # of significant t-tests correlates with
Hutchins’ word-gloss relatedness ratings (r = 0.53)
# of stems correlates highly with Hutchins’ # of types (r = 0.93)
Experiment 1
Hutchins’ gloss-matching experiment tested psychological reality of explicit semantic knowledge about phonaesthemes
What about implicit, contextual knowledge about phonaesthemes?
Do phonaesthemes influence our word choice when composing sentences?
Experiment 1: Method
36 phonaestheme-bearing from 6 phonaestheme clusters kn-, gl-, sn-, -oop, -ump, -ign
36 accompanying sentence contexts Highly congruent with one target word,
incongruent with another Transformed targets into nonsense words
Normed for opaqueness: nonsense words whose target was too easily guessed were not used
Experiment 1: Method
The stone's _______ flashed from under the leaves.
1. lague2. glandor3. thoop
Experiment 1: Results
• Phonaestheme vector similarity to sentence vector predicted word choice (r = .4, p < .05)
% of responses
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
Congruent Neutral Incongruent
Experiment 1: Results
Speakers can use phonaesthemes to guess a word’s meaning in sentence contexts
Phonaestheme’s meaning must cohere with the sentence
Experiment 2: Method
Participants asked to read and paraphrase sentences
Same materials as in experiment 1: Congruent: The stone's glandor flashed from under the
leaves. Incongruent: The stone's thoop flashed from under the
leaves. Neutral: The stone's lague flashed from under the leaves.
3 measures: Comprehension latency Typing latency Ratings of paraphrases
Experiment 2: Method
Experiment 2: Results
Experiment 2: Results
Experiment 2: Results
Experiment 2: Results
Conclusions
Form-meaning relationships can benefit from computational methods Detecting semantic clusters Criterion for evaluating form-meaning links
Speakers use phonaesthemes to disambiguate unfamiliar words in sentence contexts
Disambiguation is easier when the context coheres with the phonaestheme’s semantic contribution
Future Directions
Spoken language more appropriate for sound similarity studies? BNC Spoken corpus analysis Experiments using audible stimuli