Evaluating semantic similarity and sameness in studies of polysemy

Download Report

Transcript Evaluating semantic similarity and sameness in studies of polysemy

Evaluating semantic similarity and
sameness in studies of
polysemy and synonymy
Jarno Raukko (U. Helsinki)
1
For a full version of the PPT, see handout distributed Oct 28, 2010.
SKY WEBPAGE VERSION
2
Examples
1. Are thrifty and stingy
synonyms?
EXPECTED ANSWER:
”Well, not quite.”
2. Are violin and fiddle
synonyms?
EXPECTED ANSWER:
”Well, almost.”
3. Does back have the same meaning
in
My back hurts and
I came back?
EXPECTED ANSWER:
”Not at all. Different.”
4. Does back have the same meaning
in
I came back and
I got it back?
EXPECTED ANSWER:
”Well, almost.”
(SYNONYMY)
(POLYSEMY)
3
Relevance of semantic similarity
(vs. difference)
• In synonymy: you expect similarity for a
pair/(set) of items to be of interest
• In polysemy: primarily, you expect difference
for a pair/(set) of items to be of interest;
secondarily, you group items according to
similarity and difference
4
Yet…
SIMILARITY DIFFERENCE
SYNONYMY
POLYSEMY
(DEFAULT)
Synonymy is about
similar meanings of
different words.
But you are
interested in the
differences between
near-synonyms.
Polysemy is about
related (and
therefore somehow
similar) meanings of
a word.
(DEFAULT)
Polysemy is about
different (related)
meanings of a word.
5
synonymy --- polysemy ?
• Dirk Geeraerts tomorrow in Helsinki:
”The problem of synonymy and the problem of
polysemy are essentially the same”
• Dylan Glynn & Justyna Robinson (eds, in press)
Polysemy and Synonymy. Corpus methods and
applications in Cognitive Linguistics. Amsterdam:
Benjamins.
6
synonymy
WORD 1
WORD 2
If their semantic content is similar or the same,
this is a case of synonymy.
If their semantic content is (very) different, a
researcher of synonymy ignores this case.
7
polysemy
MEANING 1
OF WORD 1
MEANING 2
OF WORD 1
The starting point is that Word 1 has at least 2 (different)
meanings.
If meanings 1 and 2 are very similar, this might be a case of
vagueness.
If meanings 1 and 2 are totally different (and not related
semantically), this might be a case of homonymy.
If meanings 1 and 2 are somewhat different but somehow
relatable (or a bit similar), this is probably a case of
polysemy.
henceforth
W = word
M = meaning
8
scale of similarity: synonymy
The meaning of W1 and W2 is…
THE SAME --------------------------------- DIFFERENT
perfect
nearsynonymy synonymy
full
semisynonymy synonymy
weak
synonymy
quasisynonymy
NOT
WORTH
DISCUSSION
9
scale of similarity: polysemy
The meaning of M1 and M2 (of W1) is…
THE SAME --------------------------------- very DIFFERENT
two
vaguepo
instances ness
of the same
meaning
two instances
of the same
meaning type
lysemy
homonymy
(ambiguity)
instances
of different (yet related)
meaning types
10
Main question
• Is semantic similarity somehow
different when we look at polysemy
than we look at synonymy?
11
Differences so far
• Which is the default, similarity or difference?
• In synonymy, we idealize on the extreme of
the scale, but mainly look at the part of the
scale which is (fairly) close to the extreme.
• In polysemy, we operate pretty much on the
whole scale, with focus on the middle.
12
synonymy --- polysemy ?
• when you study synonymy, the polysemy of the items
gets in the way
– can you ever say “W1 and W2 are synonymous”?
– should you always say “Mx of W1 and My of W2” are
synonymous?
• when you study polysemy, you often use synonyms to
talk about meanings
– “Are get ‘receive’ and get ‘arrive’ meanings of the
same verb?”
13
synonymy --- polysemy ?
• Synonymy occurs when meaning is shared (but form differs)
• Polysemy occurs when form is shared (but meaning differs)
• Synonymy is a relational lexical-semantic property that unites
(parts of the semantic potential of) “accidentally” coinciding
words
– The forms of words involved in the synonymy relationship
are arbitrary (although the relationships might be nonarbitrary, cf. Levin this morning)
– The semantic value (that is shared) is motivating enough
that two or more forms coincide on it
– It is typical that one meaning can be expressed with two
different words.
14
synonymy --- polysemy ?
• Polysemy is a semantic property of one word at a time that
unites meanings. The relationship between them is
motivated, but it is only sometimes predictable.
– It is not accidental or arbitrary that words acquire polysemy. It is in
their nature. :-)
– It is typical for semantic value to be flexible, extended, and
“multiplied”.
– Polysemy is about categorization, both between words (W1 covers a
semantic territory) and within a word (M1 and M2 are categories too).
• One form : One meaning
• a principle that cognition may strive for / take as a default
• synonymy breaks it
• polysemy breaks it
15
synonymy --- polysemy ?
• The role of co(n)text
• You can evaluate synonymy in identical co(n)texts:
I like to play the fiddle in bars.
I like to play the violin in bars.
• Usually you evaluate polysemy in non-identical co(n)texts
I got to Zabriskie Point.
I got to a point in my life where…
• But you can use identical co(n)texts as well.
I got to be the last one.
I got to be the last one.
16
evaluating
• To study shades of semantic similarity, we need to
evaluate it.
• A corpus cannot tell us if two instances are
semantically similar
– It requires human judgement
• The main use of evaluating in this paper:
– How informants / test subjects / speakers
evaluate the semantic similarity (or difference)
of linguistic items in a more or less experimental setting
(e.g., similarity rating test)
≈ Data elicitation ≈ Population test
17
evaluating
• quantitative:
– Estimate the degree of synonymy
(or semantic distance between two meanings in
polysemy)
• qualitative:
– Justify / explain / explicate
the nature of / the reason for
semantic similarity
18
evaluating takes place in real life as well
•
•
synonymy (examples)
– in linguistic production, you e.g. estimate which of the near-synonyms might
suit your needs best
– in comprehension, you e.g. estimate whether near-synonyms that you have
encountered refer to the same semantic value
– in communication, when you negotiate meaning, you e.g. operate with
synonymous alternatives
polysemy (examples)
– in production, you e.g. apply words to new contexts
– in comprehension, you e.g. approximate meanings according to related
meanings of the same word
– jokes often exploit polysemy
– polysemy may cause misunderstandings
– in communication, when you negotiate meaning, you e.g. cross-check with
polysemy of other words
19
(Back to experiments/elicitation.)
Expected difference between
synonymy and polysemy, 1
• If an informant is asked to rate the semantic
similarity/difference of two words,
– the very fact that they are different words might cause
her/him to presuppose that there is at least some
semantic difference.
– Therefore, rating two words ”semantically identical”
requires a marked choice.
– However, if the informant realizes that the researcher
is after synonymy, then evaluating W1 and W2 as
semantically similar is more likely.
20
Expected difference between
synonymy and polysemy, 2
• If an informant is asked to rate the semantic
similarity/difference of two meanings of one
word,
– the very fact that they are uses/instances of the same
word might cause her/him to presuppose that there is
at least some semantic similarity.
– Therefore, rating two words ”semantically totally/very
different” requires a marked choice.
– However, if the informant realizes that the researcher
is after polysemy, then evaluating M1 and M2 as
semantically different is more likely.
21
Factors that influence
• In both cases (synonymy and polysemy)
– it matters a great deal
• Which test (type) we use
• What the instructions (exact phrasings) are
• Whether there is an example rating given by the
researcher
• What the selection of stimuli is
• What the linguistic context of each stimulus is
• Which types of cases have been placed in the
beginning of the test (or, the order in general)
22
Factors that influence
• Should we expect (total) consensus?
• No. There will be subjective differences.
• Why?
• The nature of semantics:
• Based on intersubjective convention
• Based on negotiation and flexibility
• Must allow for variability and
variation
23
Examples from (more or less)
experimental studies on synonymy
and polysemy
24
Whitten & al. 1979 (synonymy)
• “Indicate the degree to which two words have the
same meaning by writing a digit from 1 to 7.”
• 7 =excellent synonymy
• 1 = poor synonymy
• All 464 stimulus noun pairs were listed as synonyms in
standard references.
• The rated degree of synonymy ranged from 6.79 to
2.24. The median was 5.08.
• If placed within context of nonsynonym pairs, the
ratings for the low end might have been higher.
25
Whitten & al. 1979 (synonymy)
cont’d
• Stimulus pairs at the high end:
• purchase – buy
6.79
• lawyer – attorney
6.78
• autumn – fall
6.72
• penny – cent
6.71
• taxi – cab
6.71
• Stimulus pairs close to the median
• college – university
5.12
• output – yield
5.10
• expert – authority
5.09
• effort – attempt
5.08
• servant – maid
5.08
• soldier – warrior
5.07
26
Whitten & al. 1979 (synonymy)
cont’d
• Stimulus pairs at the low end:
• thunder – clap
• patient – invalid
• visit – chat
• suburb – neighborhood
• needle – spike
2.72
2.55
2.52
2.34
2.24
• Although instructions said that all stimuli are nouns,
some of these are more common as verbs: buy,
purchase, visit, chat
• The polysemy is obvious in many cases: fall, authority,
clap, patient, invalid
27
Whitten & al. 1979 (synonymy)
cont’d
• The main variable that they paid attention to was the order of
the two stimuli: ½ of the informants got “forward order”, ½ got
“back order”.
– In 1979 one of their main aims was to study the structurings
of the mental lexicon and lexical access.
– Example: purchase => buy 6.72
buy => purchase 6.86
– On average, perceived synonymy was affected by word
order.
– For 21 word pairs, the effect of the order was significant.
28
Whitten & al. 1979 (synonymy)
cont’d
• Some of the 21 word pairs where the order played a significant
role in the rating of the degree of synonymy:
motive => reason
quarter => fourth
mission => task
era => age
appetite => hunger
nectar => honey
aborigine => native
6.28
6.24
5.66
5.80
5.18
4.94
4.52
reason => motive
fourth => quarter
task => mission
age => era
hunger => appetite
honey => nectar
native => aborigine
5.56
5.00
4.84
4.60
4.24
3.68
3.22
• Generalization: a more specific, more academic, and
less polysemous word prompts a positive synonymy
judgement more readily than vice versa.
29
Whitten & al. 1979 (synonymy)
cont’d
• Variance (between informants)
– Mostly .50–1.20 at the end of 50 most synonymous
• Exceptionally high variance at the high synonymy end:
– murder => homicide 2.75 (cf. homicide => murder
1.03)
– Mostly 2.00–3.00 at the median of the scale
• Exceptionally low variance: province => territory 1.55
• Exceptionally high variance: congress => legislature 3.79
– Mostly 2.50–4.00 at the end of 50 least synonymous
• That is, there was little consensus at the lower end of the
30
scale.
Raukko 1994 (polysemy)
• “Decide whether the word get carries the same
meaning or two different meanings in the
sentences.”
– 0 = the same meaning
– 2 = somewhat different meaning
– 4 = very different meaning
(heuristic post hoc: 4 might mean homonymy; 0 would refer to
two instances of the same meaning type; typical polysemy
would be 1...3)
31
Raukko 1994 (polysemy)(cont’d)
• Data from my 1994 test, see handout.
32
Comparisons so far
• Whitten & al. / synonymy
– scale 1...7 (1 = very different meaning, 7 = same meaning)
– synonymy ratings ranged 2.24...6.79
– median 5.08 (most pairs were viewed at least somewhat
synonymous)
• Raukko / polysemy
– scale 0...4 (0 = same meaning, 4 = very different meaning)
– polysemy ratings ranged 0.45...3.13
– average rating 1.55, median 1.34 (most pairs were viewed as
having fairly similar but not identical meaning)
33
Comparisons so far
• Whitten & al. / synonymy
– informants saw synonymy where they were supposed to
• Raukko / polysemy
– informants did not see large meaning difference for the
most part => get is polysemous, not homonymous
– they saw some similarities, some differences, as predicted
=> they saw polysemy
• both
– differing degrees of similarity were apparent
– many ratings make sense, some don’t
– method is useful but there are skewing effects and
irreliability in several details of the setting
34
Conclusions so far
• In both synonymy and polysemy studies,
semantic intuitions vary.
• In both synonymy and polysemy studies,
finding a scale of semantic similarity is useful.
– Cf. Sandra & Rice 1995: 125
• “[researchers of prepositional polysemy] cannot propose
extremely fine-grained distinctions without bothering
about empirical data”
• “language users’ mental representation [...] is [in fact]
characterized by a high degree of granularity”
35
quantitative => qualitative
• Whitten & al’s and Raukko’s similarity rating tests did not
include informants justifying and explaining their ratings.
• E.g., Liu (this symposium) reports tests with informants
explaining their choices.
• In Raukko’s study, qualitative results come from other types of
tests
– sorting test: (1) combine stimuli into categories, (2) give
names to categories, etc.
– production test: (1) produce examples of the use of
polysemy, (2) explain links you find between them, etc.
• Vanhatalo 2005
36
Vanhatalo 2005 (synonymy)
• her PhD, The use of questionnaires in exploring
synonymy
• several types of tests
– choose most likely components
– rate components
– choose better alternative (cf. Liu)
– complete as sentences (only the word given)
– define typical frames
– spell out semantic differences
37
Vanhatalo 2005 (cont’d)
• several factors investigated
– 18 Finnish verbs of “nagging”, 17 Estonian verbs of
nagging
• the gender and age of the portrayed speaker (the subject
of “nag”)
• the degree of irritation of the portrayed speaker and
hearer
• the volume of the vocal act
– 2-4 Finnish adjectives ‘important, central, crucial,
significant’: open questions mainly
38
Vanhatalo 2005 (cont’d)
• main results (Vanhatalo 2005: 40-45): the
questionnaire method
– helped to trace differences in the meaning and use
of synonyms
• many differences not documented before in dictionaries
• sometimes consensus, sometimes deviation
• useful especially for large groups of semantically similar
words
• (Vanhatalo did not use the method for placing synonyms
on a scale of similarity)
• both open questions and ratings should be used
39
Vanhatalo 2005 (cont’d)
• main results (Vanhatalo) (cont’d)
– helped to find differences between related words in
Estonian and Finnish
– sociodemographic variables caused fairly little
variation
• age and education affected a bit more than gender
• answers critique
40
Vanhatalo 2005 (cont’d)
• main results (Vanhatalo) (cont’d)
– when both corpus method and questionnaire
method were applicable, they yielded similar
results
• however, justification of results was different
• questionnaire method dug up semantic properties that
corpus method could not
• in addition, can tackle low-frequency words
– results of questionnaire method can be utilized in
the production of electronic dictionaries
41
Other studies of synonymy that employ
experimental techniques
•
•
•
•
•
Arppe & Järvikivi 2002, 2007
Divjak & Gries 2008
Liu, in this symposium
Oversteegen, in this symposium
etc.
42
polysemy / qualitative
• In experimental settings (e.g., the sorting test):
– An informant gives a name to a meaning type, a
category within polysemy
– An informant spells out the semantic link between
two meanings
– An informant draws a hierarchy between
macrotypes and microtypes (more general and
more specific meaning types)
– An informant pinpoints at cases difficult to evaluate
43
And…
• to conclude…
44
Evaluating semantic similarity
• Both synonymy and polysemy operate on the scale of
semantic similarity vs. difference.
• Knowing about the degree of similarity is one useful
property of both.
• The way to find out about it is to use
elicitation/experiments.
• There is deviation in informants’ ratings.
• A simple explanation: informants use different criteria for
evaluation.
• Solutions: let them explicate the criteria.
use multiple methods.
45
Synonymy vs. polysemy
• Evaluating semantic similarity between the
meanings of two separate words (synonymy) is a
matter of evaluating the match between two
separate ”semantic events”
– There should be mismatch, but there isn’t.
• Evaluating semantic similarity/relatedness/
difference between the meanings of one word
(polysemy) is a matter of comparing the
applications of one single category.
– There should be match between the semantic events.
46
Synonymy vs. polysemy
• When you evaluate near-synonyms, you
balance between (i) the ideal of what would
constitute a perfect match and (ii) the nuances
of the near-synonyms
• When you evaluate meanings of a polysemous
word, you balance between (i) the assumption
that some meaning should be shared and (ii)
the actual semantic profile of the uses
47
Synonymy vs. polysemy
• In evaluating synonymy, the idealized
equivalence can be taken from the semantic
description of either of the two words.
• In evaluating polysemy, the common factor
(”core meaning”, ”shared meaning”) may be
hard to find, or become too abstract.
Maybe the first task is easier?
48
General relevance
• ”Insights in the equality or similarity of
meaning may shed light on meaning itself”
(Oversteegen / SKY 2010, Helsinki)
• The question of “identical meaning” is a
crucial basis for e.g. typology and language
comparisons: the problem of tertium
comparationis
– Cf. Haspelmath’s plenary on Saturday
49
References
Arppe, Antti & Juhani Järvikivi 2007.
Every method counts –
Combining corpus-based and
experimental evidence in the
study of synonymy. Corpus
Lingustics and Linguistic Theory 3:
2: 131-159.
Colombo, Lucia & Giovanni B. Flores
d’Arcais 1984. The meaning of
Dutch prepositions: a
psycholinguistic study of
polysemy. Linguistics 22: 51-98.
Divjak, Dagmar & Stefan Gries 2008:
Clusters in the mind? Converging
evidence from near-synonymy in
Russian. The Mental Lexicon 3: 2:
188-213.
Geeraerts, Dirk – in this symposium
Liu, Dilin – in this symposium
Oversteegen, Eleonore – in this
symposium
Raukko, Jarno 2003. Polysemy as
flexible meaning: experiments with
English get and Finnish pitää. In
Brigitte Nerlich & al (eds) Polysemy.
Flexible patterns of meaning in
mind and language. 161-193.
CONTINUED...
50
References
cont’d
Sandra, Dominiek & Sally Rice 1995.
Network analyses of prepositional
meaning: mirroring whose mind – the
linguist’s or the language user’s?
Cognitive Linguistics 6: 89-130.
Vanhatalo, Ulla 2005. Kyselytestit
synonymian selvittämisessä (etc.) [The
use of questionnaires in exploring
synonymy, etc.] PhD thesis, U-Helsinki.
http:/ethesis.helsinki.fi/julkaisut/hum/
suoma/vk/vanhatalo/kyselyte.pdf
Whitten, William B. II, W: Newton Suter,
and Michael L. Frank 1979.
Bidirectional Synonym Ratings of 464
Noun Pairs. Journal of Verbal Learning
and Verbal Behavior 18: 109-127.
Author’s
contact
information
• e-mail:
See handout and list of
participants.
• home postal address
See handout.
• affiliation
Department of Modern
Languages
Metsätalo (Unioninkatu 40 B)
FIN-00014 University of Helsinki
Finland
51