1 / 26

Why are phoneme frequency distributions skewed?

Why are phoneme frequency distributions skewed?. Andy Martin Laboratoire de Sciences Cognitives et Psycholinguistique (EHESS, DEC-ENS, CNRS), Paris. Phoneme frequencies within a language are power-law distributed (Tambovtsev and Martindale 2007). Data: Baayen et al. 1995. Power laws.

fadhila
Download Presentation

Why are phoneme frequency distributions skewed?

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Why are phoneme frequency distributions skewed? Andy Martin Laboratoire de Sciences Cognitives et Psycholinguistique (EHESS, DEC-ENS, CNRS), Paris

  2. Phoneme frequencies within a language are power-law distributed (Tambovtsev and Martindale 2007) Data: Baayen et al. 1995

  3. Power laws • Definition: A phoneme’s frequency is proportional to • Movie actor collaborations (Barabási and Albert 1999) • Income distribution (Lorenz 1905) • City sizes (Zipf 1949) • Word frequencies (Zipf 1935)

  4. Preferential attachment • Where do power law distributions come from? • Simon (1955), Barabási and Albert (1999) • Emerge in networks with two ingredients: • Growth • Preferential attachment • “Rich-get-richer” systems

  5. NetLogo demonstration (Wilensky 2005)

  6. The proposal • New words tend to contain frequent phonemes, making them even more frequent • Goal of this talk: derive this fact from an existing model of speech production • Growth: new words entering the language • Preferential attachment: new words tend to contain common phonemes • Preferential attachment can be modeled using a spreading activation model of speech production

  7. The evolving lexicon • A speech community tends to converge on a single alternative when confronted with multiple synonyms (Lass 1997, Croft 2000, Baronchelli et al. 2006) • Words compete to enter the lexicon • The lexicon is composed of the winners of these competitions • Properties that contribute to winning will increase in frequency (Boersma 1998, 2007) • Lexical competition can be modeled as a race between two lexical entries in a speech production network

  8. Spreading activation network (Dell 1986) <dog> <cat> <cat> <rat> d a g k k æ æ t t r

  9. Adding feedback <dog> <cat> <cat> <rat> <rat> d a g k k æ æ t t r

  10. Competing synonyms <couch> <couch> <sofa> <sofa> k aʊ ʧ s oʊ f ə

  11. Competing synonyms • Evidence for the simultaneous activation of synonyms comes from speech errors • Blending errors typically formed from synonyms (Wells 1951, Fromkin 1971, Poulisse 1999) frown + scowl frowl • These could result from cases in which both lexical entries are equally active

  12. Phoneme frequency effects(Dell & Gordon 2003) <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> <word> P1 P1 P1 P2 P2 P2

  13. Simulation • Toy language: 10 phonemes • Lexicon: 50 words, each 5 phonemes long • Initial state: words contain random phonemes drawn from a uniform distribution • Implemented in a speech production network • Each “generation,” a competing synonym is introduced for each word in the lexicon • Phonemes are always equiprobable in new words • The lexical entry with the highest activation after 5 time steps is the winner • The loser is eliminated from the lexicon

  14. Simulation network concept concept concept concept concept concept concept concept concept agabj cdcdc cijhh agiij bjgge accjh fjbhb iifcc iifcc jghhe jghhe a b c d e f g h i j

  15. Simulation • Activation at each time step is determined with the following equation (Dell 1986): • Skewness of the phoneme distribution in each generation is measured using Shannon entropy: • Low entropy = more skewed

  16. Simulation results

  17. Simulation results

  18. The single-phoneme problem • The simulation approaches a state in which there is only one phoneme • One solution: a teleological pressure away from inventories with few phonemes • Another solution: give some words with low-frequency phonemes an advantage • Some non-phonological factors (e.g., sociolinguistic) surely contribute to word usage • Implemented in simulation as a weight assigned to each lexical node, which affects the probability of winning a production competition

  19. Non-phonological factors • Weights are drawn randomly from a Gaussian distribution with mean 1 and standard deviation of δ • A word’s weight does not change over the course of the simulation • Revised update algorithm: • Higher weight = more activation

  20. Results

  21. What about markedness? • The most frequent phonemes tend to be similar across languages (e.g., coronal consonants) • Giving phoneme nodes differing weights creates skewed distributions, but power laws result only if the weights are themselves power-law distributed • In my account, the nature of the distribution emerges as a natural consequence of the architecture of the model

  22. Conclusions • Phoneme frequency distributions are heavily skewed • This is a consequence of a feedback loop: new words tend to contain frequent sounds, making those sounds even more frequent • The feedback loop emerges from a spreading activation model of speech production in which words compete to be used • Differing degrees of skew can be modeled by incorporating non-phonological factors into the model

  23. Thank you Andy Martin amartin@humnet.ucla.edu

  24. References 1 Barabási, Albert-Lászlóand Réka Albert. 1999. Emergence of Scaling in Random Networks. Science 286: 509–512. Baronchelli, Andrea, Maddalena Felici, Vittorio Loreto, Emanuele Caglioti and Luc Steels. 2006. Journal of Statistical Mechanics: Theory and Experiment. Boersma, Paul. 1998. Functional phonology: Formalizing the interactions between articulatory and perceptual drives. The Hague: Holland Academic Graphics. Boersma, Paul. 2007. The evolution of phonotactic distributions in the lexicon. Talk given at the Presentation Workshop on Variation, Gradience and Frequency in Phonology, Stanford University. Croft, William. 2000. Explaining Language Change: An evolutionary approach. Harlow, Essex: Longman. Dell, Gary. 1986. A spreading-activation theory of retrieval in sentence production. Psychological Review 93: 283–321.

  25. References 2 Dell, Gary and Jean Gordon. 2003. Neighbors in the lexicon: Friends or foes? In N. O. Schiller & A. S. Meyer (eds.), Phonetics and phonology in language comprehension and production: Differences and similarities. Berlin: Mouton de Gruyter. Fromkin, Victoria. 1971. The Non-anomalous Nature of Anomalous Utterances. Language 47(1): 27–52. Lass, Roger. 1997. Historical Linguistics and Language Change. Cambridge: Cambridge University Press. Poulisse, Nanda. 1999. Slips of the Tongue: Speech Errors in First and Second Language Production. Amsterdam and Philadelphia: John Benjamins. Tambovtsev, Yuri and Colin Martindale. 2007. Phoneme Frequencies Follow a Yule Distribution. SKASE Journal of Theoretical Linguistics 4: 2. Wells, R. 1951. Predicting Slips of the Tongue. Yale Scientific Magazine, December. 9–12.

  26. References 3 Wilensky, Uri. 2005. NetLogo Preferential Attachment model. http://ccl.northwestern.edu/netlogo/models/PreferentialAttachment. Center for Connected Learning and Computer-Based Modeling, Northwestern University, Evanston, IL. Zipf, George. 1935. The Psycho-biology of Language. Boston: Houghton Mifflin. Zipf, George. 1949. Human Behavior and the Principle of Least Effort: An Introduction to Human Ecology. Addison-Wesley Press.

More Related