BEYOND THE SURFACE: A COMPUTATIONAL
EXPLORATION OF LINGUISTIC AMBIGUITY
Thesis Defense
Anmol Goel
Advisor: Prof. Ponnurangam Kumaraguru
June 27, 2023
WHY STUDY AMBIGUITY?
1
Ambiguity seems to be an essential,
indispensable element for the transfer of
information from one place to another by
words. — Lewis Thomas (1974)
WHY STUDY AMBIGUITY?
1
Ambiguity seems to be an essential,
indispensable element for the transfer of
information from one place to another by
words. — Lewis Thomas (1974)
WHY STUDY AMBIGUITY?
when a phrase, statement, or
resolution is not explicitly defined,
making several interpretations
plausible.
1
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity improves communication
Piantadosi, S. T., Tily, H., & Gibson, E. (2012). The communicative function of ambiguity in language. Cognition, 122(3), 280-291. 2
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity improves communication, makes language efficient
https://news.mit.edu/2012/ambiguity-in-language-0119 2
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity improves communication, makes language efficient, can lead to humor and wordplay
Mittal, A., Tian, Y., & Peng, N. (2022). Ambipun: Generating humorous puns with ambiguous context. NAACL. 2
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity can lead to confusion and misunderstanding
https://venturebeat.com/ai/amazon-scientist-explains-how-alexa-resolves-ambiguous-requests/ 3
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity can lead to confusion and misunderstanding, particularly in situations where
clarity is important, such as legal documents.
https://www.nbcnews.com/news/us-news/think-commas-don-t-matter-omitting-one-cost-maine-dairy-n847151
https://www.lexology.com/library/detail.aspx?g=d999319c-8125-4326-bce6-88669c654da5 3
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity can lead to strange behaviours in generative models like DALL-E
White, J. C., & Cotterell, R. (2022). Schr"{o} dinger's Bat: Diffusion Models Sometimes Generate Polysemous Words in Superposition. arXiv preprint arXiv:2211.13095. 4
AMBIGUITY: THE GOOD, THE BAD AND THE UGLY
Ambiguity can lead to strange behaviours in generative models like DALL-E and chatGPT
Ortega-Martín, M., García-Sierra, Ó., Ardoiz, A., Álvarez, J., Armenteros, J. C., & Alonso, A. (2023). Linguistic ambiguity analysis in ChatGPT. arXiv preprint arXiv:2302.06426. 4
TYPES OF AMBIGUITY
Typically, three main types of ambiguity [Fromkin et al., 2018] in language are considered -
lexical,
syntactic,
and semantic.
Fromkin, V., Rodman, R., & Hyams, N. (2018). An Introduction to Language (w/MLA9E Updates). Cengage Learning. 5
LEXICAL AMBIGUITY
Arises when a word has multiple
meanings
https://wstyler.ucsd.edu/101/ 6
SYNTACTIC AMBIGUITY
Arises when a sentence can be
parsed in multiple ways.
https://wstyler.ucsd.edu/101/ 7
SEMANTIC AMBIGUITY
Arises when a word or phrase can
be interpreted in multiple ways
based on the meaning of the words
used.
https://wstyler.ucsd.edu/101/ 8
Typically, three main types of ambiguity [Fromkin et al., 2018] in language are considered -
lexical,
syntactic,
and semantic.
TYPES OF AMBIGUITY
Foci of this thesis
Fromkin, V., Rodman, R., & Hyams, N. (2018). An Introduction to Language (w/MLA9E Updates). Cengage Learning. 9
THIS DISSERTATION
To alleviate the issues raised by linguistic ambiguities in
natural language systems, it is imperative to:
1) measure ambiguity
10
THIS DISSERTATION
To alleviate the issues raised by linguistic ambiguities in
natural language systems, it is imperative to:
1) measure ambiguity
2) analyze language models
10
THIS DISSERTATION
focuses on two manifestations of linguistic ambiguities...
11
POLYSEMY
The multiplicity of meanings (or senses) for a
specific word
A form of lexical ambiguity
THIS DISSERTATION
11
POLYSEMY
The multiplicity of meanings (or senses) for a
specific word
A form of lexical ambiguity
The word “bank” can refer to a financial
institution, the side of a river, or a place where
airplanes park, among multiple other
meanings.
THIS DISSERTATION
11
POLYSEMY TAUTOLOGY
The multiplicity of meanings (or senses) for a
specific word
A form of lexical ambiguity
The word “bank” can refer to a financial
institution, the side of a river, or a place where
airplanes park, among multiple other
meanings.
Seemingly uninformative and ambiguous
phrases used in conversations
A form of semantic (or pragmatic) ambiguity
THIS DISSERTATION
11
POLYSEMY TAUTOLOGY
The multiplicity of meanings (or senses) for a
specific word
A form of lexical ambiguity
The word “bank” can refer to a financial
institution, the side of a river, or a place where
airplanes park, among multiple other
meanings.
Seemingly uninformative and ambiguous
phrases used in conversations
A form of semantic (or pragmatic) ambiguity
Colloquial tautologies like “boys will be boys” can
create ambiguity by relying on vague or
imprecise language that can be interpreted in
different ways depending on the context.
THIS DISSERTATION
11
THIS DISSERTATION
asks two research questions:
12
RQ1
Can we use linguistically-motivated methods
to measure the polysemy of words?
THIS DISSERTATION
asks two research questions:
12
RQ1 RQ2
Can we use linguistically-motivated methods
to measure the polysemy of words?
Can current language models like BERT and GPT
interpret tautological sentences?
THIS DISSERTATION
asks two research questions:
12
RQ1:
POLYSEMY QUANTIFICATION
13
ATTEMPTS AT POLYSEMY QUANTIFICATION
EACL 2021
-Requires carefully tuned
hyperparameters to get good results
-Dimensionality reduction can lead to
embedding distortion
+Measures polysemy by iteratively
binning the contextual embedding space.
Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings.. 14
ATTEMPTS AT POLYSEMY QUANTIFICATION
EACL 2021 EMNLP 2020
-Relies on large amount of data
-more than 10,000 sentences required
-Requires carefully tuned
hyperparameters to get good results
-Dimensionality reduction can lead to
embedding distortion
+Measures polysemy by iteratively
binning the contextual embedding space.
+ Measures polysemy using mutual
information approximations
Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings..
Pimentel, T., Maudslay, R. H., Blasi, D., & Cotterell, R. (2020). Speakers fill lexical semantic gaps with context. arXiv preprint arXiv:2010.02172. 14
THE SYNERGISTIC THEORY OF
LANGUAGE
It is hypothesized that syntax influences the polysemy of a word.
This leads us to consider incorporating syntax and networks in our approach.
Čech, R., Mačutek, J., Žabokrtský, Z., & Horák, A. (2017). Polysemy and synonymy in syntactic dependency networks. Digital Scholarship in the Humanities, 32(1), 36-49. 15
DESIDERATA
Our measure should be:
Unsupervised
Syntax-aware
Work in low-resource settings.
16
Multiple sentences
containing the
same word
HOW TO MEASURE POLYSEMY?
17
Multiple sentences
containing the
same word
Model/Pipeline
HOW TO MEASURE POLYSEMY?
17
Multiple sentences
containing the
same word
Model/Pipeline
A real number
signifying
polysemy
HOW TO MEASURE POLYSEMY?
17
Multiple sentences
containing the
same word
Model/Pipeline
A real number
signifying
polysemy
HOW TO MEASURE POLYSEMY?
17
f(mouth) = 3.2
HOW TO MEASURE POLYSEMY?
17
HOW TO MEASURE POLYSEMY?
18
HOW TO MEASURE POLYSEMY?
19
HOW TO MEASURE POLYSEMY?
Ricci curvature is a Wasserstein-based
metric to identify bridges and clusters in
a network.
20
HOW TO MEASURE POLYSEMY?
21
HOW TO MEASURE POLYSEMY?
22
HOW TO MEASURE POLYSEMY?
23
HOW TO MEASURE POLYSEMY?
Low-resource
24
HOW TO MEASURE POLYSEMY?
Low-resource
Unsupervised
25
HOW TO MEASURE POLYSEMY?
Low-resource
Unsupervised
Syntax-aware
26
EXPERIMENTAL SETUP
DATA
Covers English, French and
Spanish
introduced by (Garí Soler and
Apidianaki, 2021)
Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 27
EXPERIMENTAL SETUP
DATA MODEL
Covers English, French and
Spanish
introduced by (Garí Soler and
Apidianaki, 2021)
Huggingface
bert-base-uncased for English,
flaubert-baseuncased for French
and bert-basespanish-wwm-
uncased for Spanish
Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 27
EXPERIMENTAL SETUP
DATA MODEL EVALUATION
Covers English, French and
Spanish
Following previous literature in
polysemy quantification
(Xypolopoulos et al., 2021), we
utilised Spearman correlation as
our evaluation metric.
introduced by (Garí Soler and
Apidianaki, 2021)
Huggingface Spearman correlation
bert-base-uncased for English,
flaubert-baseuncased for French
and bert-basespanish-wwm-
uncased for Spanish
Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL
Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings.. 27
RESULTS
Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 28
RESULTS
Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 28
DISCUSSION
The poly-bal data setting shows
consistently strong correlations as
compared to poly-rand setting which
is quite intuitive due to the carefully
controlled sense distribution in poly-
bal sentences.
Multilingual
29
DISCUSSION
The poly-bal data setting shows
consistently strong correlations as
compared to poly-rand setting which
is quite intuitive due to the carefully
controlled sense distribution in poly-
bal sentences.
Multilingual
These results suggest that studies in
ambiguity should investigate syntax
along with semantics of an utterance.
Ablation
29
LIMITATIONS
We rely on the availability of good quality language-specific language models
Any errors in the language model may propagate into our score.
Extrinsic tasks like Word Sense Disambiguation haven't been explored yet.
30
IMPLICATIONS
Our method can be used to extend existing word inventories like WordNet by
discovering new polysemantic relations.
It can be used with generative models to produce less ambiguous texts.
31
RQ2:
TAUTOLOGIES
There are two competing views of tautological constructions:
Pragmatic View
Semantic View
32
PRAGMATIC VIEW
[Grice, 1975] proposed a pragmatic model on how listeners and speakers communicate and cooperate in
conversations.
Information is implied rather than asserted.
Proposes the four maxims of conversation
Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 33
The four maxims of conversation:
Maxim of Quantity
speakers should give as much information as necessary, but no more
PRAGMATIC VIEW
Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
The four maxims of conversation:
Maxim of Quantity
speakers should give as much information as necessary, but no more
Maxim of Quality
speakers should be truthful and should provide accurate information.
PRAGMATIC VIEW
Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
The four maxims of conversation:
Maxim of Quantity
speakers should give as much information as necessary, but no more
Maxim of Quality
speakers should be truthful and should provide accurate information.
Maxim of Relation
speakers should stay on topic and should provide relevant information
PRAGMATIC VIEW
Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
The four maxims of conversation:
Maxim of Quantity
speakers should give as much information as necessary, but no more
Maxim of Quality
speakers should be truthful and should provide accurate information.
Maxim of Relation
speakers should stay on topic and should provide relevant information
Maxim of Manner
speakers should be clear and concise in their communication
PRAGMATIC VIEW
Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
SEMANTIC VIEW
Argues that the interpretation of tautologies is not solely based on their pragmatic implications, but
rather also on the syntactic patterns and nominal classifications of the phrases [Wierzbicka, 1987].
For example, tautologies of the form "N will be N" generally convey negative aspects of the topic with an
indulgent undertone.
The way that the words are arranged in a sentence can impact the interpretation of the tautology.
Wierzbicka, A. (1987). Boys will be boys:’radical semantics’ vs.’radical pragmat- ics’. Language, pages 95–114 35
PRAGMATIC VIEW SEMANTIC VIEW
The interpretation of nominal tautologies is
context-dependent. Same tautology can take
on different meanings depending on the
conversational context
Suggests that syntactic form of phrases
contribute semantic information to the
interpretation of tautologies
DICHOTOMY OF TAUTOLOGIES
36
EXPERIMENTAL SETUP
DATA
216 sentences Controlling for
noun type, syntax and context.
Methodology of [Gibbs
and McCarrell, 1990]
Gibbs, R. W. and McCarrell, N. S. (1990). Why boys will be boys and girls will be girls: Understanding colloquial tautologies. Journal of Psycholinguistic Research, 19:125–145. 37
EXPERIMENTAL SETUP
DATA MODEL
216 sentences Controlling for
noun type, syntax and context.
Methodology of [Gibbs
and McCarrell, 1990]
Huggingface
Pretrained BERT and GPT2
37
EXPERIMENTAL SETUP
DATA MODEL EVALUATION
216 sentences Controlling for
noun type, syntax and context.
Sequence log probability
scores are a measure of how
likely a sequence of words is
according to a transformer-
based language model.
Methodology of [Gibbs
and McCarrell, 1990]
Huggingface Acceptability scores
Pretrained BERT and GPT2
37
DATA
[Gibbs and McCarrell, 1990] describes a blueprint to create datasets for
tautology acceptability studies.
We use few-shot prompting with GPT-3.5 to synthetically generate data.
Gibbs, R. W. and McCarrell, N. S. (1990). Why boys will be boys and girls will be girls: Understanding colloquial tautologies. Journal of Psycholinguistic Research, 19:125–145. 38
DATA
38
For autoregressive models like GPT:
For masked language models like BERT:
ACCEPTABILITY SCORES
39
RESULTS
Acceptability of Tautologies without context
GPT is better at handling tautologies than BERT
BERT GPT
40
RESULTS
Acceptability of Tautologies without context
GPT is better at handling tautologies than BERT
Human nouns receive higher scores than concrete and abstract nouns
BERT GPT
40
RESULTS
Acceptability of Tautologies without context
GPT is better at handling tautologies than BERT
Human nouns receive higher scores than concrete and abstract nouns
Surprisingly, LLMs seem to prefer plural tautological constructions, contrary to previous literature
on humans’ preference for modal forms [Gibbs and McCarrell, 1990]
BERT GPT
40
RESULTS
Acceptability of Tautologies with context
In general, the scores are higher for negative contexts and for plural and modal syntactic forms.
BERT GPT
41
RESULTS
Acceptability of Tautologies with context
In general, the scores are higher for negative contexts and for plural and modal syntactic forms.
This suggests that models encode negative factual connotations for tautological constructions,
similar to human behaviour [Gibbs and McCarrell, 1990].
BERT GPT
41
IMPLICATIONS
We find evidence for both - Pragmatic and Semantic View.
We hope that a syncretic approach taking into account both factors may be
necessary to fully interpret tautologies.
A more nuanced approach is needed to understand language model behaviour.
Aligning LLMs using theories like Gricean Maxims can be fruitful.
42
LIMITATIONS
Log Likelihood scores are sensitive to noise
We did not evaluate the language model acceptability scores with human
acceptability ratings.
43
SUMMARY
In this dissertation, we propose a novel and linguistically-motivated metric to quantify
polysemy.
Additionally, we investigate the pragmatic competence of large language models on
tautological sentences.
This work has important implications within linguistics and NLP to develop state-of-
the-art conversational models.
44
PUBLICATION
A pic with Sebastian Ruder (Google) who was interested in the polysemy
work at EMNLP, Abu Dhabi.
An Unsupervised, Geometric and Syntax-aware Quantification of Polysemy, EMNLP 2022
45
Prof. Ponnurangam Kumaraguru for his supervision during my Masters.
Prof. Pawan Goyal and Prof. Parameshwari for the insightful reviews.
Prof. Charu for her insights while developing this work.
Multiple collaborators - Prof. Saptarshi Ghosh, Prof. Ashutosh Modi and Prof. Ravi.
iHub Data @ IIITH for funding my Masters.
Friends and the Precog group for countless memories, encouragement and stimulating
discussions.
Family - for everything
ACKNOWLEDGEMENTS
46
THANK YOU!
Questions?
47

Beyond the Surface: A Computational Exploration of Linguistic Ambiguity

  • 1.
    BEYOND THE SURFACE:A COMPUTATIONAL EXPLORATION OF LINGUISTIC AMBIGUITY Thesis Defense Anmol Goel Advisor: Prof. Ponnurangam Kumaraguru June 27, 2023
  • 2.
  • 3.
    Ambiguity seems tobe an essential, indispensable element for the transfer of information from one place to another by words. — Lewis Thomas (1974) WHY STUDY AMBIGUITY? 1
  • 4.
    Ambiguity seems tobe an essential, indispensable element for the transfer of information from one place to another by words. — Lewis Thomas (1974) WHY STUDY AMBIGUITY? when a phrase, statement, or resolution is not explicitly defined, making several interpretations plausible. 1
  • 5.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity improves communication Piantadosi, S. T., Tily, H., & Gibson, E. (2012). The communicative function of ambiguity in language. Cognition, 122(3), 280-291. 2
  • 6.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity improves communication, makes language efficient https://news.mit.edu/2012/ambiguity-in-language-0119 2
  • 7.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity improves communication, makes language efficient, can lead to humor and wordplay Mittal, A., Tian, Y., & Peng, N. (2022). Ambipun: Generating humorous puns with ambiguous context. NAACL. 2
  • 8.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity can lead to confusion and misunderstanding https://venturebeat.com/ai/amazon-scientist-explains-how-alexa-resolves-ambiguous-requests/ 3
  • 9.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity can lead to confusion and misunderstanding, particularly in situations where clarity is important, such as legal documents. https://www.nbcnews.com/news/us-news/think-commas-don-t-matter-omitting-one-cost-maine-dairy-n847151 https://www.lexology.com/library/detail.aspx?g=d999319c-8125-4326-bce6-88669c654da5 3
  • 10.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity can lead to strange behaviours in generative models like DALL-E White, J. C., & Cotterell, R. (2022). Schr"{o} dinger's Bat: Diffusion Models Sometimes Generate Polysemous Words in Superposition. arXiv preprint arXiv:2211.13095. 4
  • 11.
    AMBIGUITY: THE GOOD,THE BAD AND THE UGLY Ambiguity can lead to strange behaviours in generative models like DALL-E and chatGPT Ortega-Martín, M., García-Sierra, Ó., Ardoiz, A., Álvarez, J., Armenteros, J. C., & Alonso, A. (2023). Linguistic ambiguity analysis in ChatGPT. arXiv preprint arXiv:2302.06426. 4
  • 12.
    TYPES OF AMBIGUITY Typically,three main types of ambiguity [Fromkin et al., 2018] in language are considered - lexical, syntactic, and semantic. Fromkin, V., Rodman, R., & Hyams, N. (2018). An Introduction to Language (w/MLA9E Updates). Cengage Learning. 5
  • 13.
    LEXICAL AMBIGUITY Arises whena word has multiple meanings https://wstyler.ucsd.edu/101/ 6
  • 14.
    SYNTACTIC AMBIGUITY Arises whena sentence can be parsed in multiple ways. https://wstyler.ucsd.edu/101/ 7
  • 15.
    SEMANTIC AMBIGUITY Arises whena word or phrase can be interpreted in multiple ways based on the meaning of the words used. https://wstyler.ucsd.edu/101/ 8
  • 16.
    Typically, three maintypes of ambiguity [Fromkin et al., 2018] in language are considered - lexical, syntactic, and semantic. TYPES OF AMBIGUITY Foci of this thesis Fromkin, V., Rodman, R., & Hyams, N. (2018). An Introduction to Language (w/MLA9E Updates). Cengage Learning. 9
  • 17.
    THIS DISSERTATION To alleviatethe issues raised by linguistic ambiguities in natural language systems, it is imperative to: 1) measure ambiguity 10
  • 18.
    THIS DISSERTATION To alleviatethe issues raised by linguistic ambiguities in natural language systems, it is imperative to: 1) measure ambiguity 2) analyze language models 10
  • 19.
    THIS DISSERTATION focuses ontwo manifestations of linguistic ambiguities... 11
  • 20.
    POLYSEMY The multiplicity ofmeanings (or senses) for a specific word A form of lexical ambiguity THIS DISSERTATION 11
  • 21.
    POLYSEMY The multiplicity ofmeanings (or senses) for a specific word A form of lexical ambiguity The word “bank” can refer to a financial institution, the side of a river, or a place where airplanes park, among multiple other meanings. THIS DISSERTATION 11
  • 22.
    POLYSEMY TAUTOLOGY The multiplicityof meanings (or senses) for a specific word A form of lexical ambiguity The word “bank” can refer to a financial institution, the side of a river, or a place where airplanes park, among multiple other meanings. Seemingly uninformative and ambiguous phrases used in conversations A form of semantic (or pragmatic) ambiguity THIS DISSERTATION 11
  • 23.
    POLYSEMY TAUTOLOGY The multiplicityof meanings (or senses) for a specific word A form of lexical ambiguity The word “bank” can refer to a financial institution, the side of a river, or a place where airplanes park, among multiple other meanings. Seemingly uninformative and ambiguous phrases used in conversations A form of semantic (or pragmatic) ambiguity Colloquial tautologies like “boys will be boys” can create ambiguity by relying on vague or imprecise language that can be interpreted in different ways depending on the context. THIS DISSERTATION 11
  • 24.
    THIS DISSERTATION asks tworesearch questions: 12
  • 25.
    RQ1 Can we uselinguistically-motivated methods to measure the polysemy of words? THIS DISSERTATION asks two research questions: 12
  • 26.
    RQ1 RQ2 Can weuse linguistically-motivated methods to measure the polysemy of words? Can current language models like BERT and GPT interpret tautological sentences? THIS DISSERTATION asks two research questions: 12
  • 27.
  • 28.
    ATTEMPTS AT POLYSEMYQUANTIFICATION EACL 2021 -Requires carefully tuned hyperparameters to get good results -Dimensionality reduction can lead to embedding distortion +Measures polysemy by iteratively binning the contextual embedding space. Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings.. 14
  • 29.
    ATTEMPTS AT POLYSEMYQUANTIFICATION EACL 2021 EMNLP 2020 -Relies on large amount of data -more than 10,000 sentences required -Requires carefully tuned hyperparameters to get good results -Dimensionality reduction can lead to embedding distortion +Measures polysemy by iteratively binning the contextual embedding space. + Measures polysemy using mutual information approximations Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings.. Pimentel, T., Maudslay, R. H., Blasi, D., & Cotterell, R. (2020). Speakers fill lexical semantic gaps with context. arXiv preprint arXiv:2010.02172. 14
  • 30.
    THE SYNERGISTIC THEORYOF LANGUAGE It is hypothesized that syntax influences the polysemy of a word. This leads us to consider incorporating syntax and networks in our approach. Čech, R., Mačutek, J., Žabokrtský, Z., & Horák, A. (2017). Polysemy and synonymy in syntactic dependency networks. Digital Scholarship in the Humanities, 32(1), 36-49. 15
  • 31.
    DESIDERATA Our measure shouldbe: Unsupervised Syntax-aware Work in low-resource settings. 16
  • 32.
    Multiple sentences containing the sameword HOW TO MEASURE POLYSEMY? 17
  • 33.
    Multiple sentences containing the sameword Model/Pipeline HOW TO MEASURE POLYSEMY? 17
  • 34.
    Multiple sentences containing the sameword Model/Pipeline A real number signifying polysemy HOW TO MEASURE POLYSEMY? 17
  • 35.
    Multiple sentences containing the sameword Model/Pipeline A real number signifying polysemy HOW TO MEASURE POLYSEMY? 17 f(mouth) = 3.2
  • 36.
    HOW TO MEASUREPOLYSEMY? 17
  • 37.
    HOW TO MEASUREPOLYSEMY? 18
  • 38.
    HOW TO MEASUREPOLYSEMY? 19
  • 39.
    HOW TO MEASUREPOLYSEMY? Ricci curvature is a Wasserstein-based metric to identify bridges and clusters in a network. 20
  • 40.
    HOW TO MEASUREPOLYSEMY? 21
  • 41.
    HOW TO MEASUREPOLYSEMY? 22
  • 42.
    HOW TO MEASUREPOLYSEMY? 23
  • 43.
    HOW TO MEASUREPOLYSEMY? Low-resource 24
  • 44.
    HOW TO MEASUREPOLYSEMY? Low-resource Unsupervised 25
  • 45.
    HOW TO MEASUREPOLYSEMY? Low-resource Unsupervised Syntax-aware 26
  • 46.
    EXPERIMENTAL SETUP DATA Covers English,French and Spanish introduced by (Garí Soler and Apidianaki, 2021) Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 27
  • 47.
    EXPERIMENTAL SETUP DATA MODEL CoversEnglish, French and Spanish introduced by (Garí Soler and Apidianaki, 2021) Huggingface bert-base-uncased for English, flaubert-baseuncased for French and bert-basespanish-wwm- uncased for Spanish Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 27
  • 48.
    EXPERIMENTAL SETUP DATA MODELEVALUATION Covers English, French and Spanish Following previous literature in polysemy quantification (Xypolopoulos et al., 2021), we utilised Spearman correlation as our evaluation metric. introduced by (Garí Soler and Apidianaki, 2021) Huggingface Spearman correlation bert-base-uncased for English, flaubert-baseuncased for French and bert-basespanish-wwm- uncased for Spanish Garí Soler, A., & Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL Xypolopoulos, C., Tixier, A. J. P., & Vazirgiannis, M. (2020). Unsupervised Word Polysemy Quantification with Multiresolution Grids of Contextual Embeddings.. 27
  • 49.
    RESULTS Garí Soler, A.,& Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 28
  • 50.
    RESULTS Garí Soler, A.,& Apidianaki, M. (2021). Let’s play mono-poly: BERT can reveal words’ polysemy level and partitionability into senses. TACL 28
  • 51.
    DISCUSSION The poly-bal datasetting shows consistently strong correlations as compared to poly-rand setting which is quite intuitive due to the carefully controlled sense distribution in poly- bal sentences. Multilingual 29
  • 52.
    DISCUSSION The poly-bal datasetting shows consistently strong correlations as compared to poly-rand setting which is quite intuitive due to the carefully controlled sense distribution in poly- bal sentences. Multilingual These results suggest that studies in ambiguity should investigate syntax along with semantics of an utterance. Ablation 29
  • 53.
    LIMITATIONS We rely onthe availability of good quality language-specific language models Any errors in the language model may propagate into our score. Extrinsic tasks like Word Sense Disambiguation haven't been explored yet. 30
  • 54.
    IMPLICATIONS Our method canbe used to extend existing word inventories like WordNet by discovering new polysemantic relations. It can be used with generative models to produce less ambiguous texts. 31
  • 55.
    RQ2: TAUTOLOGIES There are twocompeting views of tautological constructions: Pragmatic View Semantic View 32
  • 56.
    PRAGMATIC VIEW [Grice, 1975]proposed a pragmatic model on how listeners and speakers communicate and cooperate in conversations. Information is implied rather than asserted. Proposes the four maxims of conversation Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 33
  • 57.
    The four maximsof conversation: Maxim of Quantity speakers should give as much information as necessary, but no more PRAGMATIC VIEW Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
  • 58.
    The four maximsof conversation: Maxim of Quantity speakers should give as much information as necessary, but no more Maxim of Quality speakers should be truthful and should provide accurate information. PRAGMATIC VIEW Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
  • 59.
    The four maximsof conversation: Maxim of Quantity speakers should give as much information as necessary, but no more Maxim of Quality speakers should be truthful and should provide accurate information. Maxim of Relation speakers should stay on topic and should provide relevant information PRAGMATIC VIEW Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
  • 60.
    The four maximsof conversation: Maxim of Quantity speakers should give as much information as necessary, but no more Maxim of Quality speakers should be truthful and should provide accurate information. Maxim of Relation speakers should stay on topic and should provide relevant information Maxim of Manner speakers should be clear and concise in their communication PRAGMATIC VIEW Grice, H. P. (1975). Logic and conversation. In Speech acts, pages 41–58. Brill. 34
  • 61.
    SEMANTIC VIEW Argues thatthe interpretation of tautologies is not solely based on their pragmatic implications, but rather also on the syntactic patterns and nominal classifications of the phrases [Wierzbicka, 1987]. For example, tautologies of the form "N will be N" generally convey negative aspects of the topic with an indulgent undertone. The way that the words are arranged in a sentence can impact the interpretation of the tautology. Wierzbicka, A. (1987). Boys will be boys:’radical semantics’ vs.’radical pragmat- ics’. Language, pages 95–114 35
  • 62.
    PRAGMATIC VIEW SEMANTICVIEW The interpretation of nominal tautologies is context-dependent. Same tautology can take on different meanings depending on the conversational context Suggests that syntactic form of phrases contribute semantic information to the interpretation of tautologies DICHOTOMY OF TAUTOLOGIES 36
  • 63.
    EXPERIMENTAL SETUP DATA 216 sentencesControlling for noun type, syntax and context. Methodology of [Gibbs and McCarrell, 1990] Gibbs, R. W. and McCarrell, N. S. (1990). Why boys will be boys and girls will be girls: Understanding colloquial tautologies. Journal of Psycholinguistic Research, 19:125–145. 37
  • 64.
    EXPERIMENTAL SETUP DATA MODEL 216sentences Controlling for noun type, syntax and context. Methodology of [Gibbs and McCarrell, 1990] Huggingface Pretrained BERT and GPT2 37
  • 65.
    EXPERIMENTAL SETUP DATA MODELEVALUATION 216 sentences Controlling for noun type, syntax and context. Sequence log probability scores are a measure of how likely a sequence of words is according to a transformer- based language model. Methodology of [Gibbs and McCarrell, 1990] Huggingface Acceptability scores Pretrained BERT and GPT2 37
  • 66.
    DATA [Gibbs and McCarrell,1990] describes a blueprint to create datasets for tautology acceptability studies. We use few-shot prompting with GPT-3.5 to synthetically generate data. Gibbs, R. W. and McCarrell, N. S. (1990). Why boys will be boys and girls will be girls: Understanding colloquial tautologies. Journal of Psycholinguistic Research, 19:125–145. 38
  • 67.
  • 68.
    For autoregressive modelslike GPT: For masked language models like BERT: ACCEPTABILITY SCORES 39
  • 69.
    RESULTS Acceptability of Tautologieswithout context GPT is better at handling tautologies than BERT BERT GPT 40
  • 70.
    RESULTS Acceptability of Tautologieswithout context GPT is better at handling tautologies than BERT Human nouns receive higher scores than concrete and abstract nouns BERT GPT 40
  • 71.
    RESULTS Acceptability of Tautologieswithout context GPT is better at handling tautologies than BERT Human nouns receive higher scores than concrete and abstract nouns Surprisingly, LLMs seem to prefer plural tautological constructions, contrary to previous literature on humans’ preference for modal forms [Gibbs and McCarrell, 1990] BERT GPT 40
  • 72.
    RESULTS Acceptability of Tautologieswith context In general, the scores are higher for negative contexts and for plural and modal syntactic forms. BERT GPT 41
  • 73.
    RESULTS Acceptability of Tautologieswith context In general, the scores are higher for negative contexts and for plural and modal syntactic forms. This suggests that models encode negative factual connotations for tautological constructions, similar to human behaviour [Gibbs and McCarrell, 1990]. BERT GPT 41
  • 74.
    IMPLICATIONS We find evidencefor both - Pragmatic and Semantic View. We hope that a syncretic approach taking into account both factors may be necessary to fully interpret tautologies. A more nuanced approach is needed to understand language model behaviour. Aligning LLMs using theories like Gricean Maxims can be fruitful. 42
  • 75.
    LIMITATIONS Log Likelihood scoresare sensitive to noise We did not evaluate the language model acceptability scores with human acceptability ratings. 43
  • 76.
    SUMMARY In this dissertation,we propose a novel and linguistically-motivated metric to quantify polysemy. Additionally, we investigate the pragmatic competence of large language models on tautological sentences. This work has important implications within linguistics and NLP to develop state-of- the-art conversational models. 44
  • 77.
    PUBLICATION A pic withSebastian Ruder (Google) who was interested in the polysemy work at EMNLP, Abu Dhabi. An Unsupervised, Geometric and Syntax-aware Quantification of Polysemy, EMNLP 2022 45
  • 78.
    Prof. Ponnurangam Kumaragurufor his supervision during my Masters. Prof. Pawan Goyal and Prof. Parameshwari for the insightful reviews. Prof. Charu for her insights while developing this work. Multiple collaborators - Prof. Saptarshi Ghosh, Prof. Ashutosh Modi and Prof. Ravi. iHub Data @ IIITH for funding my Masters. Friends and the Precog group for countless memories, encouragement and stimulating discussions. Family - for everything ACKNOWLEDGEMENTS 46
  • 79.