19. What’s in CVC-like Thing? ……….Natural Language Processing Enabling machines to understand and respond to what people mean, so they can interact with a computer naturally – without having to adapt their behavior to a computer's limitations.
Resolving ambiguities in written and spoken language requires analyzing grammar, concepts, context, and human knowledge. For example:
"The company is ready to sell” is not easy for a computer to understand because the sentence is syntactically ambiguous – is the company opening for business, or does it want to be acquired?
Resolving this ambiguity requires understanding the context: is the sentence in the middle of an article on mergers and acquisitions? Or is the sentence followed by “Its shelves are stocked with all the hot products"? This succeeding sentence is helpful only if the computer understands that the possessive pronoun “its” refers to the company, and that “stocked” and “products” are more relevant to selling goods than to being acquired.
19.1
-Sound produced
-Physical construct
-Language and linguistic structure
19.1.1 Language as a formal system
-language and linguistic structure as” things
-set of sentence
19.1.2 Language as physical reality
-analogous to that of electrons in physics
-stream of speech
19.1.3language as a psychological reality
-mental code
-language process
(怡萱)
19.3.1 Experimental word games in English and Korean19.3.1 Experimental word games in English and Korean1. Forced-chioce version of the word-blending task:- English speakers: perferred onset-rime blends(e.g. SIEVE + FUZZ = SUZZ)- Korean speakers: perferred body-coda blends(e.g. THONG + SEM = THOM)2. Oral unit reduplication exercise:- English speakers: were better at the rime-copying task(Task #1: SAN-AN)- Korean speakers: were better at the body-copying task(Task #2: SA-SAN)19.3.2 Global sound-similarity judgments in English and Korean1. Global sound similarity judgments (SSJs):- Korean speakers: a shared initial CV that enhanced judged sound similarity ratings, while a shared fianl rime contributed no more to the similarity scores than a shared intial and final cosonant.- English speakers: the opposite to Korean speakers19.3.3 Concept formation in Korean1. Formation concept:- Korean speakers: a set of disyllabic words containing the common body sequence KA- as part of either syllable (KANG.CO / SIM.KAK) were significantly easier to identify a s a class than a set containing the common rime sequence -AK in either syllable (CAK.SIM / SIM.KAK).- English speakers: have not been tested yet19.3.4 A new list-recall task for non-literate participants:1.List recall task:- English participants: both readers and non-readers, were able to remember more names from the rime-sharing lists than from the body-sharing ones.- Korean participants: the readers and ono-readers performed in much the same way as English parcitipants.
(勝芬)
21. Experimental Methods in the ……..21.1 IntroductionPurposeTo provide answers about long consonants or geminates, using an experimental approach in the analysis of Hindi geminates.The specific topics in this chaptera) the duration of geminates and of the vowels preceding themb) long distance durational effectsc) the duration of geminates via-a-vis clusters and of the vowel preceding these.d)the syllabification of geminates and the issue of their integritye) the status of apparent geminates21.1.1 Some facts about geminates in Hindiinvolve the consonantal closuretwo separate consonantshave severe phonotactic restrictions.they occur only intervocalicallyalways preceded by the non-preipheral vowels, the short vowels21.2.2 Diachronic data on development of geminatesdue to cluster simplification in the development of Middle Indo-Aryan (MIA) from Sanskirt.
(威鈴)
21.1 Introductionlong consonants=geminatesPurpose:provide answers about long consonants or geminates, using an experimental approach in the analysis of Hindi geminates.The specific topics of this paper:(a)the duration of geminates and of the vowels preceding them(b) long distance durational effects(c)the duration of geminates vis-a-vis clusters and of the vowel preceding these(d)the syllabification of geminates and the issue of their integrity(e)the status of "apparent"geminates21.1.1 Some facts about geminates in Hindi:1.geminates involve the consonantal closure held for a longer period 2.geminates are not two separate consonants3.geminates occur only intervocalically4.geminates are always preceded by short vowels21.1.2 Diachronic data on development of geminates (1) Examples of geminate formation in the history of Indo-AryanSanskrit bhakta meal, food > Pali/Prakrit bhattaSanskrit sapta seven > Pali/Prakrit sattaSanskrit dugha-milk > MLA duddha-
(宜珊)
22. Morphophonemics and the Lexicon..
22.1 introductionTry to find a way to explain the stem-final alternation in Turkish22.2 the problemSize (length) as a categorizerWedel: neighborhood density & alternation rateConclude: Wedel’s findings cannot be meaningfully evaluated for it’s done by statistics from a dictionary.(a single-speaker corpus is a better choice)22.3 methodology: TELL and a frequency corpus22.3.1 the Turkish electronic living lexicon (TELL)Maker: University of California, BerkeleyContent: 30,000 words (25,000 headwords, 5,000 place names)Voice producer: 63-year-old standard Istanbul Turkish speakerMorphological context:NOM. caseACC. case1. person predicativePossessive caseProfessional suffix22.3.2 stem-final alternations: a snapshot from TELL22.3.3 frequency corpusMaker: Kemal Oflazer, at Sabancı University in Istanbul, TurkeyContent: 12,000,000 words22.4 frequencyRhodes’ AE Flapping and Bybee’s coronal deletionGradient alternation and semi-regularFindingsIn velar deletion: more frequent, more alternationIn voicing: less frequent, more alternation22.5 neighborhood density22.5.1 neighborhood density with a single-speaker corpus22.5.2 frequency-weighted neighborhood density22.6 cohorts22.7 etymology22.8 conclusions
(晟維)
23. How Do Listeners Compensate ……
How do listeners compensate for phonology? Eurie Shin23.1 IntroductionWords vs intended formsHow do listeners plan acoustic realizationsTwo approaches-different predictionPhonological-inference accountFeature-parsing accountCross-modal repetition priming study by Gaskell & Marslen-Wilson (1996)Gow (2002) –assimilation creates lexical ambiguityBase of feature-parsing accountGoal of the studyExamine whether listeners can cope with complete assimilationUnderlying homorganic consonant sequences as stimuliThree experiments of the study1) VCCV sequences generated by cross-splicing2) non-words of the form (C)VCCV(C)3) pseudo-compounds consisting of 2 non-words of the form (C)(V)(C)(C)VC#CVC23.2 Experiment I23.2.1 Methods7 repetitions of the original stimuli in 5 different vowel environmentsWaveform editing techniquesTaskDetermine whether the stimulus be transcribed as indicatedNon-parametrical tests23.2.2 ResultListeners accept homorganic clustersNo significant differences between coronal and velar coda responses23.2.3 DiscussionPurpose of the experimentResults-no evidence in favor of the phonological-inference accountAnother possibilityStimuli did not sound like real wordsDifference between coronal and labial coda optionsSummaryFeature-parsing account predict the resultsTentative interpretation
(Aleksandra)
Part Five: Ch 23.3~6 How do Listeners Compensate for Phonology?
Part Five: Ch 23.3~6 How do Listeners Compensate for Phonology?
23.3 Experiment 2
23.3.1 Methods
Use monomorphemic non-words in (C)VCCV(V)
A linguistically naïve female native Korean speaker recorded the stimuli
310 different items occurred twice
Ten Korean native speakers as listeners
23.3.2. Results
Listeners show high rates of correct response(96.8%) to the fillers.
23.3.3 Discussion
Goal
Whether listeners can infer the underlying forms
From phonetic homorganic clusters in non-words
Interesting result
Velar stimuli
Both /s.k/ and /p.k/ assimilate to [k.k] in natural speech
Main factors of lowering YES response
Frequency of labial-to-velar assimilation
Coronal-to-noncornoal assimilation
23.4 Experiment 3
23.4.1 Methods
Stimuli is pseudo-compounds consisting of two words
Listeners were provided transcription options
Ten Korean native speakers
23.4.2. Results
Rate of YES is 96.3%
Subject understood the task
23.4.3 Discussion
Purpose
Test the expectation
If listeners identified the initial morpheme in a new compound
Response
Higher than the previous ones
23.5 General Discussion
Subject are
preferred homorganic clusters as underlying forms
preferred coronal as targets of assimilation
Response pattern in Experiment 2 &3
Support the phonological-inference account
Listeners did not recover legal heterorganic clusters as underlying form
Result of Experiment 3
Show listeners reply more on phonological inference.
23.6 Conclusion
Goal
To test whether listeners use their phonological knowledge to infer original forms in speech recognition
Result suggest
Phonological inference is a part of the speech recognition process
(惠珍)
沒有留言:
張貼留言