You hear a word and know it within a fraction of a second, often before it has finished. How? This page covers the chapter with a pre-reading quiz, a cohort simulator, explainers and eight games.
Answer before you read the explainers. Guessing is fine. Each answer comes with a short explanation of what the chapter says.
The chapter assumes spoken word recognition can be split into five temporally distinct stages (following Tyler and Frauenfelder, 1987). Other processes that help us understand larger stretches of language interact with them, and different models of word recognition differ in how they treat each stage.
"Word" varies between languages. In Aymara, uta means house, but endings can be added, so that utawa means it's a house. Even in English it is unclear whether cat and cats, or foot and feet, are one word or two. The chapter also separates passive vocabulary (what we understand) from active vocabulary (what we use).
Estimates of the average vocabulary range from 20,000 to 75,000 words. Take the conservative 20,000 (Nation, 2006) and a search speed of 100 words per second. An exhaustive search would take about 3 minutes to find out that splundle is not a word. We obviously do not check every word we know.
The speech input is widely accepted to be analysed into phonemes. A phoneme is the smallest unit that, when changed, can signal a different word. The pair cat and cap differs only in /t/ versus /p/. Compared with whole words, phonemes have three advantages:
Not everyone agrees phonemes are psychologically real. Some argue they are a by-product of learning to read and write. Speakers of Chinese, whose characters stand for words, and illiterate speakers do worse at phoneme monitoring tasks (Morais et al., 1979). They still understand speech, so phonemes cannot be a necessary part of spoken word recognition.
Phonetic features, such as the voicing that separates /p/ and /b/, are an alternative. English vowels are nasalised before a nasal consonant, so in soon the nasality starts during the vowel. Gating experiments show listeners use this to tell soon from soup or suit before the final consonant (Warren and Marslen-Wilson, 1987, 1988). In Bengali and Hindi, nasal vowels contrast, so nasality tells listeners about the vowel and not about the following consonant (Lahiri and Marslen-Wilson, 1991).
Diphones and syllables are proposed too. Languages differ. A Māori syllable has one vowel, preceded by at most one consonant. English syllables can be much more complex. Stress also varies: Czech stresses the first syllable, Polish the last but one, and English has word stress that is not always in the same place (noun 'import, verb im'port).
Metrical Segmentation Strategy (MSS; Cutler and Norris, 1988). Some 90% of English content words have a stressed first syllable (Cutler and Carter, 1987). So a new word search starts each time a strong syllable is heard. In an experiment, listeners were slower to spot mint in mintave than in mintesh. Mintave has two strong syllables, so a second search starts at the /t/ and splits the word. Other languages show different patterns, which supporters explain through the rhythm of each language: Dutch, French (syllable based), Japanese (mora based), and Finnish (stress and vowel harmony).
Possible Word Constraint (PWC; Norris et al., 1997). The input should be segmented into words with no leftover sounds. It is harder to detect see in the nonword /siʃ/ than in /siʃʌb/, because the leftover /ʃ/ is not a possible English word, but shub is possible, even though it is not a current word.
The smallest unit that, when changed, can result in a change in meaning, by signalling a different word.
Two words that differ in only one sound, such as cat and cap. They help establish a language's phoneme inventory.
It is very often extremely difficult to know where one word finishes and the next begins.
A distinctive property of speech sounds, such as voicing or nasality. Phonemes can be grouped into classes by their features.
Truncated portions of a recorded word are played, with longer and longer fragments, and participants identify the word at each "gate".
After pre-lexical analysis comes contact: the output is mapped onto forms in the mental lexicon. This mapping is bottom-up processing. Some models claim contact is only bottom-up. Others also give a role to top-down processing, such as using context to pre-select words.
Models also differ in whether one word or many are considered at once. Serial search models, such as Forster's Search Model (1976), consider one word at a time. Parallel models contact several stored words at the same time. Cross-modal priming gives support for parallel contact.
Participants heard a prime and saw a visual target on screen. The study was in Dutch, but the chapter explains it with English examples.
| Item | Example | Relation |
|---|---|---|
| Prime | captain | Heard |
| Partner of the prime | capital | Starts with the same sounds, but is not heard |
| Target 1 | ship | Related to captain |
| Target 2 | money | Related to the partner, capital |
| Target 3 | lamp | Related to neither |
| Control prime | justice | Unrelated |
Early condition, targets shown part-way through the prime, before it is complete: both ship and money were responded to faster and more accurately after captain than after justice. Late condition, targets shown at the end of the prime: only ship showed facilitation. The facilitation for money had disappeared.
So early in the word both captain and capital were contacted. By the end, only the word that matched the whole input was left. Later work showed the overlap has to be at the beginning: bee is primed by honey but not by money, which shares most of the sounds but not the start.
The Cohort model (Marslen-Wilson and Welsh, 1978; Marslen-Wilson, 1987) is a parallel, mainly bottom-up approach. Once the initial sounds of a word are heard, all words in the lexicon that start with the same sounds are contacted. This set is the word-initial cohort. Cohort members become activated, making further information about them available. As more input arrives, words that no longer match fall out of contention. This is called left-to-right processing, a metaphor from reading.
Early versions treated words as either available or not. It soon became clear that entries have levels of activation. The more input that supports an item, the higher its activation. Figure 8.1 in the chapter shows captain continuing to rise while captive and capital fall as the sound pattern diverges. A graded model also copes with partial matches. If a speaker says shtudent for student, the initial sound partly activates /s/ and later input keeps raising student.
Pick a word, then play it one sound at a time. Candidates that no longer match the input are crossed out.
This is a simplified illustration of Figure 8.1 and the chapter's tress, trestle and trespass example. Only three candidates are shown, and the sounds are written in simple letters, not full phonetic notation.
The set of words in the mental lexicon that share the initial sounds of the input.
Processing driven by the input speech signal.
Processing driven by context, such as using meaning to pre-select words.
Participants hear primes (spoken words or sentences) and make a decision about a visual target on a screen.
Faster and more accurate responses after a test prime than after a control prime.
Parallel models contact many stored words at once. Serial models consider one word at a time.
When many candidates have been contacted, selection must choose one. As more bottom-up information arrives, one word is left as the only match. Each word has a uniqueness point, where it stops overlapping with other words in its cohort. For captain this could be the second vowel, where it separates from captive, depending on the speaker's pronunciation. Uniqueness points can be worked out from pronunciation dictionaries and databases, and gating experiments show that actual recognition points correlate highly with them. How the decision is made is debated: perhaps activation passes a threshold, or perhaps the relative probabilities of cohort words matter.
The chapter claims nonwords are processed the same way as real words, since we cannot know in advance whether a word will be one we know. For a nonword like splundle the set of matching words shrinks until no strong candidate remains. That is the deviation point.
In an early experiment, responses to nonwords in a list took roughly 450 msec from the deviation point (Marslen-Wilson, 1980). Later studies found some variation, depending on the task and on how word-like the rest of the nonword is. A word-like ending such as -ise or -ic slows "no" answers in lexical decision. Listeners keep monitoring after the deviation point, which helps with noisy speech and with mispronunciations.
Selecting a word-form is not the same as getting its stored information. Forms are linked to lemma-based content: meaning and grammatical type. Lexical access is the point where this information becomes available. In many models, contact and access amount to the same thing.
Access happens before a unique word is selected. In Zwitserlood's study, ship and money were both responded to faster after only the fragment /kæp/ of captain. That is only explained if the meanings of both captain and capital were accessed.
We can often finish someone else's sentence, which suggests top-down prediction. But the chapter finds little evidence that context pre-selects candidates during normal fast recognition. Take the sentence The men stood around for a while and watched their captain... In cross-modal priming with such sentence primes, the target ship shown just before captain gets no facilitation. If it is shown later, after /kæp/, it does. Bottom-up input has priority in the first moments of contact.
Finishing a sentence is explained as a slower process, a field of meaning that becomes available, as when a related word primes another. A further problem for pre-selection is the huge number of words that many contexts would activate.
Context still helps. Lexical information is assessed against the developing interpretation of the sentence, which changes activation levels. So a word can be recognised before its uniqueness point if context rules out rivals. Take trespass. Alone, its recognition point equals its uniqueness point, at the /p/, where tress and trestle are eliminated. After The poacher was found guilty of tres-, tress and trestle do not fit, so recognition comes earlier, at the /s/. (Try this in the simulator above.)
Marslen-Wilson, Brown and Tyler (1988) asked listeners to press a button when they heard guitar. The context was one of four sentences, each starting The crowd was waiting eagerly... Response times were measured from the onset of the word.
Mean response times in msec. Bar length is scaled to the longest value, 320.
Times grew longer as the verb constrained the noun less. Monitoring times also tended to fall the later a word sat in a grammatically well-formed sentence, because the interpretation becomes more constraining. A later neuroimaging study (Kuperberg et al., 2000) used the same kind of sentences. All three violation conditions raised activation in a brain area linked to comprehension problems in brain-damaged patients, which may be the area that integrates information.
The point in a word where it no longer overlaps with other words in its initial cohort.
The point where the decision is reached that a particular word has been heard.
The point in a nonword where it diverges from known words.
The point at which the information stored with a word, such as meaning and grammatical type, becomes available.
Choosing candidate words in advance from the context, before the sounds arrive.
Listeners press a button as soon as they hear a target word in a sentence.
Participants decide whether a spoken or written string is a real word.
Frequency effects are a regular and robust finding (Forster and Chambers, 1973; Monsell, 1991; Murray and Forster, 2004). Common words give faster response times and greater accuracy. In the cohort illustration, captive and capital rise more slowly than the more frequent captain (Figure 8.2). Priming effects in the cross-modal experiments were also greater for frequent candidates.
An alternative account gives frequent words a higher resting level of activation. With the same input, they cross a recognition threshold sooner. Recency (Oliphant, 1983) may work similarly. Using a word raises its activation, and if the return to the resting level takes a while, hearing the word a second time can fall at a moment when activation is still raised. ERP studies show frequency in the size and speed of electrical responses, which suggests more neural stimulation for high-frequency words.
Recognising a word depends on the properties of other words that compete with it. Nasalisation on the vowel of soon helps, because it separates soon from suit and soup. The same cue on the vowel of fruit tells us little about the word, because no English word has a nasal after that vowel. It may say something about the speaker. This is contingency of choice: knowing you heard cat depends on the sounds of cat and also on knowing you did not hear cap, can or cash (Warren and Marslen-Wilson, 1987, 1988).
One popular idea is that competing words inhibit each other. Given an ambiguous input like /kæpt/, the more frequent captain has higher activation, so it inhibits captive more than the reverse. This boosts the difference between the two and makes recognition of captain more certain when nothing else tells them apart.
A word's neighbourhood is the set of words that share similar properties (Luce and Pisoni, 1998). A word-initial cohort is one kind. A wider kind is all words that differ from the target by one phoneme, wherever that phoneme is. The chapter's example is monkey, whose neighbours include month, muddy, mummy, many, honey, funny and sunny.
Recognition (in tasks such as lexical decision) and naming (reading aloud) are affected by how many neighbours a word has, how similar they are, and how frequent they are. A word with many high-frequency neighbours is identified more slowly and less accurately than a word with few neighbours or only low-frequency ones. These factors make up neighbourhood density, which also influences word production (Vitevitch, 2002).
Frequently encountered words are recognised faster and more accurately.
A word's baseline activation. Frequent words may start higher.
A word heard recently has a temporary advantage over similar words not heard for a while.
How informative a cue is depends on which alternative words exist.
Competition in which words reduce each other's activation, more strongly when the competitor is more active.
How many similar words a word has, how similar they are and how frequent they are.
Morphologically complex words, such as cats, children or predisposition, have more than one meaningful element. Inflection adds affixes for grammatical information (cat to cats, jump to jumped). Derivation makes a different kind of word (dispose to disposition). Inflections tend to be more productive than derivations, and the more regular patterns are the ones children identify first, which is why they say foots or feets.
The matching process in recognition would be morphological decomposition: look for cats under the entry for cat. This is simple for left-to-right processing, because cat is contacted from the first three phonemes anyway. Inflection is mostly done through suffixes across the world's languages.
Two systems are contrasted. A rule-based system analyses the input word into meaningful parts. A full-listing system lists every word, simple or complex. The differences between regular and irregular forms, and between transparent and opaque derivations, suggest the mental lexicon probably relies on both. Rule-based analysis of productive, regular affixation exists alongside stored representations of complex forms that do not suit easy analysis.
8.1 Based on the first two phonemes of monkey and swing, what words are in the word-initial cohort of each?
8.2 What are the uniqueness points of monkey and swing? In the sentence All the children loved watching the long-tailed monkey as it started to swing through the trees, is the recognition point of each word likely to be the same as its uniqueness point? What information might be used?
8.3 What are the deviation points of these nonwords: blorsilize, breganist, brondation, munart, pytel, shrortile?
8.4 In a lexical decision experiment, correct "no" times, measured from the deviation point, were longer for three-syllable than for two-syllable nonwords. Is this what the Cohort model predicts? Why or why not? Measured from the end of the nonword, the three-syllable items were faster.