Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence
Figure 16.3 Vowel space. 169 Figure 16.4
Download 1.05 Mb. Pdf ko'rish
|
document (2)
|
Figure 16.3
Vowel space. 169 Figure 16.4 Pronunciation chart for Australian English vowel phonemes. 16.3 ALLOPHONES AND PHONETIC REALISATION Phonemes are idealisations. Actual speech sounds – phones – approx- imate to phonemes and may vary significantly between speakers. As language users, we are very good at detecting the distinctive contrasts in speech sounds and assimilating phones to phonemes. The first point of interest here is that typical monolingual speakers can only assimilate phones to the phonemes that are in their native language. This means that they literally can’t hear phonemes that aren’t in their language as they are not sensitive to phonetic contrasts that are not distinctive in their native language. Since natural languages di ffer in the phonetic contrasts that are semantically distinctive, this can make learning a second language particularly di fficult if the second language contains phonemes that are not in the first language. In Vietnamese and Mandarin, for instance, the tone of a speech sound is distinctive. Tone is not a dis- tinctive contrast in English, however, so English native speakers learning these languages have enormous di fficulty hearing the dis- tinction as they automatically assimilate the relevant phones to the same English phoneme. This is complicated by the fact that one and the same phoneme may be realised in distinct (but not semantically distinctive) allophones depending on the phonemic context of utterance. The phonetic di fference between allophones is very difficult for a native speaker of the language to hear as the contrast is not distinctive but predictable from the context of occurrence. Allophones occur in complementary distribution. This means that if a phoneme has more than one associated allophone, each allo- phone will always and only be produced in predictable phonemic con- texts according to phonetic realisation rules. The phonetic realisation of allophones in complementary distribution is strictly rule governed. Some examples will serve to make this strikingly clear. One phonetic realisation rule in English involves aspirating voice- less stops when they are word initial. This means that if a word begins with a voiceless stop, then the stop is produced with a little pu ff of air that is not produced in other contexts. Put the palm of your hand right in front of your mouth and utter the words ‘top’ and ‘stop’ repeatedly and you should notice the di fference. When you produce the word initial /t/ in ‘top’you should feel the extra pu ff of air on the palm of your hand. When you produce the /t/ in ‘stop’, however, this pu ff of air is absent since this allophone of /t/ is not aspirated. Try this out with other pairs of words containing 170 aspirated and unaspirated voiceless stop allophones, such as the /k/ in ‘Cate’ and ‘skate’ or the /p/ in ‘pat’ and ‘apple’. It is very di fficult to actually hear the difference between these phones, however, precisely because they are allophonic variants – aspiration is not a distinctive contrast in English. Another phonetic realisation rule of English is that vowels are nasalised always and only before nasal obstruents. Hold your nose while pronouncing ‘cat’ and ‘can’ repeatedly and you should be able to tell the di fference. Again, try this out with other pairs of words con- taining nasalised and non-nasalised vowels, such as ‘pot’ and ‘pond’ or ‘sit’ and ‘sing’. Some phonetic realisation rules are restricted to certain dialects, like that which governs the production of the glottal stop allophone of /t/ in cockney English. Speakers of cockney English still produce the standard voiceless alveolar stop allophone of /t/ when it is word initial or follows a fricative, such as in ‘tell’, ‘start’, ‘faster’ or ‘softer’, but produce the glottal stop when /t/ occurs in intervocalic contexts (between vowels) such as in ‘butter’ or ‘letter’, and also when /t/ occurs word finally, such as in ‘short’ or ‘let’. The interesting point for our purposes is that this strictly rule- governed activity occurs despite our total ignorance of these phonetic realisation rules and our complete insensitivity to the phonetic dis- tinctions between the allophones produced. In fact, these phonetic realisation rules are so deeply ingrained that we can’t help but carry them over illicitly when we learn a second lan- guage. We’re typically unable to notice that we’re actually producing di fferent sounds in different contexts but it is likely to be apparent to a native speaker of the language we are learning. An example of this is that English native speakers learning French will continue to (incorrectly) aspirate word initial voiceless stops – such as the /t/ in ‘Tour de France’ – even after this has been pointed out to them. As speakers of a language where this aspirated phoneme is in complementary distribution with other allophonic variants, we are simply unable to detect that we’re doing it. Similarly native speak- ers of German learning English have a tendency to devoice word final stops – producing the /g/ in ‘dog’ as /k/ for instance. So, not only do we tend to try and produce the sounds of our second language using only the phonemes of our first language, we also carry over our native phonetic realisation rules into our second language. These two factors account for the accent that a non-native speaker will typically never quite be able to get rid of. They may even- tually – depending on the first and second languages in question – be 171 able to accurately produce the phoneme set of the second language, but overriding phonetic realisation rules is very di fficult. This is also the case with dialectical variations of the same language, such as the many dialects of English. Fortunately, processing accents is also something we are generally quite good at. Sometimes it may take some exposure before we are able to process accents with ease – I found the Glaswegian accent to be impenetrable at first – but with su fficient exposure, we implicitly learn to apply a filter to the sounds we are hearing which assimilates them to the intended phonemes. While we might have to explicitly focus on this initially, we quickly internalise these transformative principles. This indicates that we continue to internalise rule-governed lin- guistic principles long past the time at which we acquire our native language, even if we never learn a second language. Whenever we are first exposed to speakers of dialects of English that are distinct from our own, or speakers of English as a second language, we very quickly internalise rules that allow us to process their accent with ease. Of course, as far as the speaker of another dialect of the same language is concerned, it is us who has the accent – I’m sure Glaswegians find my Australian accent as initially impenetrable as I found theirs. 16.4 FIRST-LANGUAGE ACQUISITION We can also find plenty of evidence of rule-governed linguistic activ- ity in the first-language acquisition literature. When a child reaches twelve months or so of age, they enter the single word stage of language acquisition. They are able to point at objects and name them with single word utterances, and they are also able to indicate some of their desires with single word utterances – ‘ball’, ‘doll’, ‘mummy’. At this stage it is common to see both semantic overextension and semantic underextension of newly acquired terms. Overextension, as the name suggests, is when a term is applied to referents beyond its proper extension. When, for instance, a child first acquires the word ‘ball’ they may well then apply this term to other round objects, such as some fruit or the Moon. Similarly, if they have a dog and learn its name, they may well then apply this term to all dogs. Semantic underextension, as you have no doubt guessed, is the opposite phenomenon. Underextension occurs when a child restricts the application of a newly acquired term to only certain of its proper 172 referents. They might, for instance, only use ‘toy’ for a particular favourite toy. Underextension tends to be less common than overex- tension. In both cases, the child fairly rapidly learns the correct scope of application of the new term in their lexicon. This is at least prima facie evidence that the child is internalising the rules that govern the correct application of the term, based on observable features of its referents. Another typical feature of the single word stage is the systematic phonemic substitution of certain phonemes that the child is unable to produce. For instance, the alveopalatal voiceless fricative is quite a di fficult phoneme to produce as it requires fairly dextrous tongue placement so it is not uncommon to hear children systematically replace it with either the voiceless alveolar fricative or the voiceless interdental fricative – so ‘ship’ is pronounced ‘sip’ or ‘thip’. As well as making these phonemic substitutions, the child is also likely to make systematic phonological simplifications. The initial syl- lable in ‘sleep’ is also phonetically di fficult to produce, so the child is likely to produce ‘seep’ in its place. More phonologically complex words are even further simplified – often idiosyncratically. A common example is ‘sketty’ for ‘spaghetti’. The particularly interesting point here is not just that children make these systematic substitutions, but that they are also sensitive to the fact that they are doing so. They are typically perfectly able to hear the di fference between an adult’s utterances of ‘seep’ and ‘sleep’ and, depending on their age, may well realise they are being teased if you reproduce their phonetic output rather than the correct utterance. So the development of their ability to comprehend phonemes outpaces the development of their ability to produce them and they compen- sate for this by making systematic – rule-governed – simplifications and substitutions. 16.5 LANGUAGE AND RULES While we’ve mostly concentrated in this chapter on phonological processes, other areas of linguistics are also rich with examples of rule-governed behaviour. If you have some exposure to linguistics, or are planning to take an introductory course, I’d urge you to reflect on your knowledge – or to approach the subject – with a particular view to looking not just for evidence of rule-governed behaviour, but also for processes that might be troublesome to account for computationally. 173 In the preceding sections, we concentrated on identifying evidence in favour of the computational implementability of the linguistic facil- ity. In the following chapter we are going to return to philosophical material and problematise a key aspect – arguably the most crucial aspect – of the production and comprehension of language, namely the determination of meaning. 174 C H A P T E R 1 7 MEANING This chapter marks a return to philosophical material after six chap- ters of technical material. We’ve seen how computers can be programmed to strategically play complex games and we compared this to our intuitive under- standing of how humans play these games. We’ve looked at expert systems as an example of machine reasoning and we’ve considered typical human performance on certain reasoning tasks in the context of determining the scope of the explanatory burden for the compu- tationalist in accounting for this typical performance. We’ve also examined how we might employ formal systems and search procedures to facilitate one of the mechanisms implicated in the linguistic facility – ruling on the grammaticality of strings. In addition, We’ve drawn out evidence from linguistics – mostly per- taining to phonological processes – which supports the notion that language behaviour is rule governed and, hence, computationally implementable. Next we’re going to consider a thought experiment which targets computationalism and seeks to show that there is a crucial facet of mental life that the computationalist cannot account for – the fact that our mental states are meaningful. 17.1 THE CHINESE ROOM It is a crucial and defining feature of our mental states that they have semantic content – that they are meaningful states. Any adequate theory of mind must be able to account for the semantic contents of mental states. Computation is an entirely syntactic process. The operations of formal systems are syntactically specified symbol manipulations. We’ve seen the explanatory e fficacy of formal systems in accounting for a number of cognitive mechanisms. The crucial question for 175 present purposes is whether or not we can account for semantics in terms of syntactic operations. The thought experiment I want to entertain here was originally described by John Searle and seeks to establish that syntax alone is not su fficient for semantics. Imagine that you are asked to spend several hours carrying out a certain task for experimental purposes. You are introduced to an enormous room containing thousands of shelves of numbered books. There is a table in the centre of the room with one of the books upon it. You flick through the book and see that it contains nothing but rewrite rules for symbols that you’ve never seen before, with a numer- ical notation alongside each rewrite rule. You are told that you will be left alone in the room, at which point a piece of paper with a string of symbols on it will be passed through a slot in the door. Your task is to find the rule in the book whose input side is exactly that string of symbols and to copy out the output string of symbols onto the other side of the piece of paper and pass it back through the slot. You are then to find the book on the shelves whose number corresponds with the numerical notation alongside the rule you just followed and replace the book on the table with this new book. You are left alone in the room and things proceed exactly as described. A piece of paper is passed through the slot in the door, you trawl through the book on the table to find this string on the input side of a rule and copy out the output string of the rule, then you replace the book on the table with the book from the shelves whose number was given by the notation next to the rule. Another piece of paper with a new string of symbols is passed through the slot in the door and you repeat the procedure. After doing this for several hours, you are told that the strings of symbols were actually sentences in Chinese script. The books in the room encode all the possible conversations you might have in Chinese in several hours. Each book represents a conversation state and pro- vides reasonable responses for possible inputs. It turns out that you have been having a conversation with a Chinese native speaker for several hours and, on the basis of merely following the rewrite rules encoded in the books, have passed a Chinese Turing test. Clearly, however, you do not thereby understand Chinese. For one thing, the conversational replies you were making to the Chinese questions did not accord with your beliefs and desires, but with arbit- rary responses encoded in the books. For instance, one of the ques- tions might have been ‘do you like strawberry ice cream?’ and your 176 scripted response was ‘yes, it’s delicious’, despite the fact that you can’t abide strawberry ice cream. Or perhaps one of the questions was ‘are you getting hungry?’ and your scripted response was ‘no, I’m fine for the moment, thanks’, despite the fact that you were ravenous and wondering when lunch was. For another thing, your capacity to converse in Chinese does not extend beyond the Chinese room. If a Chinese native speaker were to pass you a written Chinese query once you have left the room, the strings of symbols would still be meaningless to you. It is only through recourse to the encoded conversation states in the books of the Chinese room that you are able to give the appearance of under- standing and to pass the relevant Turing test. The situation described in the thought experiment is one in which the processes of a formal system – the rewriting of symbols accord- ing to formal rules – su ffice to pass a Turing test. The intuition that is primed by the thought experiment, however, is that even though the Download 1.05 Mb. Do'stlaringiz bilan baham: 
|