Applied Linguistics: Psycholinguistics, Computational Linguistics, Second Language Acquisition and Clinical Linguistics

This chapter answers Section XH3.7, “Applied Linguistics”, of XH3 Linguistics, one of the six optional sections of the Humanities and Social Sciences (XH) paper, chosen at the time of application. It teaches how the mind acquires, produces and understands language in real time (psycholinguistics), how language relates to the brain, bilingualism and cognition, how computers model and process language — parsing, corpora, large language models, and applications such as speech recognition, morphological analysis and sentiment analysis — how a second language is learned and taught, and how language disorders are assessed and treated (clinical linguistics). The few numericals here count n-grams.

1. Psycholinguistics: language acquisition

Psycholinguistics studies the mental processes behind language. First-language acquisition follows a broadly similar path across children and languages: cooing and babbling, a one-word (holophrastic) stage, two-word combinations, then rapid growth into complex sentences. Early perception is broad and then narrows: young infants discriminate speech-sound contrasts from many languages, and over roughly the first year tune in to the contrasts of the language around them while losing sensitivity to others. Children are not simply imitating: they produce forms they have never heard, such as “goed” or “foots”. This overregularization shows that they have induced a rule (add -ed, add -s) and overapplied it, strong evidence that acquisition is rule-governed and creative rather than rote.

Why acquisition is so fast and uniform is debated. Nativist accounts (in the Chomskyan tradition) argue that children are born with a language faculty — Universal Grammar — because the input underdetermines the grammar they end up with (the “poverty of the stimulus” argument). Usage-based / constructivist accounts hold that general cognitive abilities, pattern-finding and rich interaction suffice. The critical (or sensitive) period hypothesis proposes that the capacity to acquire a first language to native levels declines with age, a claim supported by the difficulty of late first-language acquisition.

🎯 Why errors are evidence
A child’s “mistakes” are often the most informative data in psycholinguistics. Overregularized forms could not be copied from adult speech, so they reveal the rule the child has built. The same logic runs through the field: systematic errors expose the underlying system more clearly than correct output does.

2. Real-time language processing: production and comprehension

Speaking and understanding happen fast, incrementally and largely below awareness. In production, a speaker moves from message to grammatical structure to the sounds that realize it. Speech errors — slips of the tongue such as spoonerisms (“heft lemisphere”) — are patterned, not random: they show that sounds, morphemes and words are planned at distinct levels before articulation.

In comprehension, the listener accesses word meanings (lexical access), aided by priming (a related word is recognized faster), and builds syntactic structure word by word rather than waiting for the sentence to end. The clearest evidence is the garden-path sentence: in “The horse raced past the barn fell”, readers first take “raced” as the main verb, then must reanalyse when “fell” arrives. Reaction-time, self-paced-reading and eye-tracking methods measure this processing load as it unfolds, and event-related brain potentials add another window: a semantically odd word tends to elicit an N400 response, while a syntactic problem tends to elicit a P600.

💡 Reading up the garden path
A garden-path sentence is grammatical, yet it misleads the parser into a reading it must abandon — like being led up a garden path to a dead end and forced to turn back. The momentary slowdown at the disambiguating word is the measurable footprint of incremental, commitment-then-revision parsing.

3. Language and the brain, bilingualism and cognition

In most people language is lateralized to the left hemisphere. Damage there can cause aphasia, and the classical types map to the site of the lesion. Broca’s aphasia follows damage to the left frontal lobe (Broca’s area): speech is non-fluent, effortful and agrammatic (function words dropped), while comprehension is relatively preserved. Wernicke’s aphasia follows damage to the left posterior temporal lobe (Wernicke’s area): speech is fluent but often empty, with word-substitutions, and comprehension is seriously impaired. Conduction aphasia, classically linked to the pathway between the two areas, leaves comprehension and fluency fair but repetition markedly poor.

Bilingualism is the norm, not the exception, worldwide. A simultaneous bilingual acquires two languages from infancy; a sequential bilingual adds the second later. Bilinguals routinely code-switch (alternate languages within a conversation or sentence) in rule-governed ways, and both languages stay active, so managing them is thought to engage general control processes. A once-popular claim of a broad “bilingual advantage” in executive control is now regarded as at best small and inconsistent across studies, so it too should be stated with caution. On language and cognition, the strong “linguistic determinism” that language dictates thought is not supported; a weaker linguistic relativity — that habitual language can influence attention and memory in some domains — has modest support and is best stated cautiously.

⚠️ Broca versus Wernicke: fluency and comprehension
The reliable contrast is this: Broca’s (frontal) aphasia is non-fluent with relatively preserved comprehension; Wernicke’s (temporal) aphasia is fluent with impaired comprehension. Swapping fluency and comprehension between the two is the commonest exam error, as is attaching the frontal lesion to Wernicke’s.

4. Computational linguistics: modelling, parsing, corpora and LLMs

Computational linguistics models and processes language with computers. A text is first tokenized into units; a simple statistical model is the n-gram, which predicts a word from the n−1 words before it (a bigram model uses one previous word, a trigram model two). Such models need large corpora to estimate probabilities, and because many valid sequences never appear even in huge data, smoothing reserves probability for the unseen.

Parsing assigns grammatical structure to a sentence. Constituency (phrase-structure) parsing groups words into nested phrases; dependency parsing links each word to its head by grammatical relations (subject, object). Parsers are trained and tested on treebanks — corpora hand-annotated with such structures. The same pipeline supports morphological analysis, the segmenting of words into morphemes and their features, for which finite-state transducers are a standard, efficient tool. Such systems are trained and evaluated against held-out annotated data, so the quality of a corpus and its annotation sets a ceiling on what any model built from it can learn.

Large language models (LLMs) are neural networks, typically based on the transformer architecture, trained on very large text collections to predict the next token. This objective yields strikingly fluent text and strong performance on many tasks, but it is prediction, not verified knowledge: an LLM can state false things confidently (so-called hallucination), reflects biases in its training data, and does not by itself guarantee factual accuracy. Outputs therefore need checking.

  • Speech recognition (ASR) maps an acoustic signal to text, combining an acoustic component with a language model; modern systems are largely neural.
  • Sentiment analysis classifies the attitude expressed in a text (positive, negative, neutral), using either sentiment lexicons or machine-learning classifiers trained on labelled data.

5. Second language acquisition and teaching

A central distinction in SLA, associated with Krashen, is between acquisition (a subconscious process, like a child picking up a first language) and learning (conscious study of rules). On his influential but much-debated Monitor Model, learned knowledge mainly acts as a “monitor” that edits output, grammatical structures emerge in a fairly fixed natural order, progress depends on comprehensible input just beyond the current level (often summarized as “i + 1”), and a low affective filter (low anxiety, high motivation) lets input through.

A learner’s evolving system is an interlanguage (Selinker): a rule-governed grammar of its own, neither the first nor the target language, which may fossilize (stop developing) short of native proficiency. Teaching methods have shifted from the grammar-translation and drill-based audiolingual methods toward communicative language teaching and task-based approaches that make meaningful use the engine of learning. Learning also draws on cognitive resources such as working memory and attention; noticing a form in the input is widely held to aid its uptake.

ℹ️ Assessment: proficiency, achievement and washback
Language assessment distinguishes proficiency tests (overall ability, independent of a course) from achievement tests (what a specific course taught), and formative assessment (feedback during learning) from summative (a final judgement). A good test is both reliable and valid, and planners watch its washback — the way an exam reshapes what teachers teach and learners study.

6. Clinical linguistics

Clinical linguistics applies linguistic analysis to the assessment and treatment of communication disorders. Assessment combines standardized, norm-referenced tests with naturalistic language sampling and observation, describing a person’s difficulty at the level of sounds, words, grammar, meaning or use rather than by a label alone. Disorders are broadly developmental (present as language develops) or acquired (following injury or disease to a previously intact system).

  • Developmental: Developmental Language Disorder (difficulty with language despite otherwise typical development), speech sound disorders, childhood apraxia of speech, and developmental stuttering.
  • Acquired: aphasia after stroke or brain injury, dysarthria (weak or poorly controlled speech muscles), acquired apraxia of speech (impaired motor planning), and the language decline seen in some dementias.
  • Evidence-based therapy integrates the best research evidence with clinical expertise and the client’s own values and goals, and measures outcomes rather than assuming them.
Developmental versus acquired disorders
AxisWhen it arisesExamples
DevelopmentalAs language is still developingDevelopmental Language Disorder, speech sound disorder, childhood stuttering
AcquiredAfter damage to a previously intact systemAphasia, dysarthria, acquired apraxia of speech
⚠️ Developmental and acquired are not the same axis as the symptom
Whether a disorder is developmental or acquired is about its origin, not its surface symptom. The same symptom (say, difficulty producing fluent speech) can be developmental in a child or acquired after a stroke, and the assessment and therapy differ accordingly. Classifying by symptom alone is the error to avoid.

Key takeaways

  • First-language acquisition is rule-governed and creative: overregularization (“goed”) shows children induce rules. Nativist accounts invoke Universal Grammar and poverty of the stimulus; usage-based accounts rely on general cognition; the critical-period hypothesis says first-language capacity declines with age.
  • Production and comprehension are fast and incremental: patterned speech errors reveal planning levels, priming speeds lexical access, and garden-path sentences expose word-by-word parsing with reanalysis, measured by reaction time, self-paced reading and eye-tracking.
  • Language is usually left-lateralized: Broca’s aphasia (frontal) is non-fluent and agrammatic with preserved comprehension, Wernicke’s (temporal) is fluent with impaired comprehension. Bilingualism is normal, with code-switching; strong linguistic determinism is rejected, weak relativity is stated cautiously.
  • Computational linguistics tokenizes and models text: an n-gram predicts a word from the previous n−1, needing corpora and smoothing. Constituency and dependency parsing use treebanks; finite-state transducers do morphology; LLMs are transformer networks predicting the next token, fluent but prone to hallucination.
  • In SLA, Krashen separates subconscious acquisition from conscious learning, with the Monitor, natural order, comprehensible input (i + 1) and affective filter hypotheses; an interlanguage may fossilize. Teaching moved to communicative and task-based methods; assessment distinguishes proficiency, achievement, formative, summative and watches washback.
  • Clinical linguistics assesses communication disorders through standardized tests and language sampling, and classifies them as developmental (such as Developmental Language Disorder and stuttering) or acquired (aphasia, dysarthria, apraxia of speech); therapy should be evidence-based.

Practice questions (16)

Attempt each one before opening the answer. Every explanation names the tempting wrong option as well as the right one, because that is where marks are lost.

  1. A two-year-old who has often heard “went” starts saying “goed” instead. What does this overregularization most directly show?

    1. That the child has simply forgotten the correct form “went” because of ordinary memory decay over time
    2. That the child has induced a past-tense rule and overapplied it to an irregular verb
    3. That the child is precisely imitating a form that an adult around them must have produced
    4. That the child has no grammatical system yet and is combining sounds purely at random
    Show answer

    Answer: B — That the child has induced a past-tense rule and overapplied it to an irregular verb

    “Goed” is a form the child has not heard, so it cannot be imitation or forgetting; it is the regular -ed rule applied to an irregular verb, showing an internalized, overextended rule. That it is a systematic error, not random noise, is exactly why it counts as evidence of a grammar.

  2. Readers slow down sharply at “fell” in “The horse raced past the barn fell.” Which claim about comprehension does this best support?

    1. Comprehension is incremental: readers commit to a structure word by word and must reanalyse when it fails
    2. Readers withhold all of their interpretation right up until the final full stop, and only then build one single structure
    3. The sentence is ungrammatical, which is the only reason any slowdown is observed at all
    4. Word meaning is accessed only after the whole syntactic tree has already been completed
    Show answer

    Answer: A — Comprehension is incremental: readers commit to a structure word by word and must reanalyse when it fails

    The slowdown at the disambiguating word is the signature of a garden path: the parser commits early to “raced” as the main verb and must revise when “fell” forces the reduced-relative reading, which shows incremental processing. The sentence is in fact grammatical, and readers clearly do not wait for the end, which rules out the other options.

  3. A patient speaks in halting, effortful, telegraphic phrases that drop function words, yet follows spoken instructions fairly well. This pattern is most consistent with:

    1. Broca’s aphasia, from a lesion in the left frontal lobe, with non-fluent agrammatic speech and relatively preserved comprehension
    2. Wernicke’s aphasia, which arises from a lesion in the left posterior temporal lobe, with fluent but largely empty speech and seriously poor comprehension
    3. Developmental stuttering, a fluency disorder that typically begins in early childhood and has no lesion
    4. Dysarthria, a weakness or poor control of the speech muscles affecting articulation of all sounds
    Show answer

    Answer: A — Broca’s aphasia, from a lesion in the left frontal lobe, with non-fluent agrammatic speech and relatively preserved comprehension

    Non-fluent, effortful, agrammatic output with comprehension relatively intact is the classic Broca’s profile, tied to a left frontal lesion. Wernicke’s aphasia is the opposite — fluent speech with impaired comprehension; stuttering is a developmental fluency disorder without this agrammatism; and dysarthria is a motor-speech weakness affecting articulation, not grammar.

  4. A child raised with Tamil and English from birth, using both before age three, is best described as which kind of bilingual?

    1. A sequential bilingual, because the second language is always necessarily added only after school begins
    2. Not bilingual at all, since genuine bilingualism requires equal, perfectly balanced command of both languages
    3. A learner using code-switching, which is by definition a disorder of mixing two separate language systems
    4. A simultaneous bilingual, who acquires both languages together from infancy rather than one after the other
    Show answer

    Answer: D — A simultaneous bilingual, who acquires both languages together from infancy rather than one after the other

    Acquiring two languages together from infancy is simultaneous bilingualism; adding a second language later would be sequential. Bilingualism does not require perfectly balanced ability, and code-switching is a normal, rule-governed practice, not a disorder.

  5. In a dependency parse of a sentence, what does each link represent?

    1. The order in which the individual letters of each word were originally typed in
    2. A grammatical relation joining a word to its head, such as subject or object
    3. The acoustic similarity between two spoken words as measured by a microphone
    4. The historical sound change that connects a modern word to its ancestor form
    Show answer

    Answer: B — A grammatical relation joining a word to its head, such as subject or object

    Dependency parsing links each word to its syntactic head through a labelled grammatical relation such as subject or object; constituency parsing instead groups words into nested phrases. The other options describe typing order, acoustic measurement and sound change, none of which is a dependency relation.

  6. In a bigram language model, the probability of each word is estimated from:

    1. Every other word in the document taken together, regardless of how far away each one happens to sit
    2. Only the alphabetical rank of the word within a dictionary of the whole language
    3. The single word immediately preceding it in the sequence
    4. The frequency of the word averaged across many unrelated languages at once
    Show answer

    Answer: C — The single word immediately preceding it in the sequence

    A bigram model conditions each word on just the one word before it (an n-gram of n = 2 uses n−1 = 1 word of context). It does not use the whole document, an alphabetical rank, or cross-language averages; those would be different models or no model at all.

  7. Automatically labelling product reviews as positive, negative or neutral is an instance of which computational application?

    1. Sentiment analysis, which classifies the attitude or polarity expressed in a text
    2. Speech recognition, which converts a recorded acoustic speech signal into fully written text
    3. Morphological analysis, which segments each word into its component morphemes
    4. Dependency parsing, which assigns a head-and-relation structure to a sentence
    Show answer

    Answer: A — Sentiment analysis, which classifies the attitude or polarity expressed in a text

    Classifying the attitude in text as positive, negative or neutral is sentiment analysis. Speech recognition turns audio into text, morphological analysis breaks words into morphemes, and dependency parsing builds syntactic structure — none of which is about attitude or polarity.

  8. An LLM produces a fluent, confident paragraph that includes a non-existent citation. What does this best illustrate about such models?

    1. That the model has deliberately chosen to deceive the user by inventing a source on purpose
    2. That fluent output guarantees factual accuracy, so its citations never need to be verified
    3. That the model has direct, live access to a complete, fully up-to-date and authoritative database of every publication in existence
    4. That it predicts likely next tokens rather than retrieving verified facts, so outputs can be fluent yet false (hallucination)
    Show answer

    Answer: D — That it predicts likely next tokens rather than retrieving verified facts, so outputs can be fluent yet false (hallucination)

    An LLM is trained to predict plausible next tokens, so it can generate fluent text that is nonetheless false, including invented references — the phenomenon called hallucination. It has no intent to deceive, fluency does not guarantee truth, and it is not a live database of publications, so its outputs must be checked.

  9. On Krashen’s account, which situation best fits the term “acquisition” as opposed to “learning”?

    1. A student memorizing a table of verb conjugations from a grammar book for an upcoming written test
    2. A migrant picking up a new language subconsciously through everyday interaction without studying rules
    3. A learner consciously applying a memorized spelling rule while carefully editing a formal essay
    4. A class reciting a verb paradigm aloud in strict unison as drilled repeatedly by the teacher
    Show answer

    Answer: B — A migrant picking up a new language subconsciously through everyday interaction without studying rules

    For Krashen, acquisition is the subconscious uptake of language through meaningful exposure, as in picking up a language by living in it. Memorizing conjugations, consciously applying a rule while editing, and drilling a paradigm are all examples of conscious learning, not acquisition.

  10. A second-language learner’s speech settles into a stable, rule-governed system of its own that stops short of native grammar and no longer changes. This is best described as:

    1. A simple transfer error, meaning each mistake is only the first language showing through unchanged
    2. A grammaticality judgement, the speaker’s conscious rating of whether a sentence sounds acceptable
    3. A fossilized interlanguage, an autonomous learner grammar whose development has halted before the target
    4. The critical period, the window in early life during which a first language can be acquired natively
    Show answer

    Answer: C — A fossilized interlanguage, an autonomous learner grammar whose development has halted before the target

    An interlanguage is the learner’s own systematic grammar, and when its development halts short of the target it is said to have fossilized. A transfer error is a single first-language carry-over, a grammaticality judgement is a rating task, and the critical period is a developmental window — none names this plateaued system.

  11. Which pairing of a disorder with its category (developmental or acquired) is correct?

    1. Aphasia after a stroke is acquired, because it follows injury to a previously intact language system
    2. Aphasia after a stroke is developmental, because all language disorders are present from birth onward
    3. Developmental Language Disorder is acquired, since it appears only after a clear brain injury in adulthood
    4. Dysarthria from muscle weakness is developmental, because muscle control can never be lost once gained
    Show answer

    Answer: A — Aphasia after a stroke is acquired, because it follows injury to a previously intact language system

    Post-stroke aphasia is acquired: it damages a language system that had developed normally. It is not developmental, Developmental Language Disorder is by definition developmental rather than injury-caused, and dysarthria can certainly be acquired through later muscle or nerve damage.

  12. Which of the following statements about language, the brain and real-time processing are correct?

    1. In most people, language is lateralized predominantly to the left hemisphere
    2. Broca’s aphasia is typically non-fluent with comprehension relatively preserved
    3. Patterned speech errors suggest that production is planned at distinct levels before articulation
    4. Strong linguistic determinism, that language fully dictates what thoughts are possible, is well established
    Show answer

    Answer: A — In most people, language is lateralized predominantly to the left hemisphere; B — Broca’s aphasia is typically non-fluent with comprehension relatively preserved; C — Patterned speech errors suggest that production is planned at distinct levels before articulation

    Left-lateralization is typical, Broca’s aphasia is non-fluent with comprehension relatively spared, and patterned slips reveal multi-level production planning. The last is false: strong linguistic determinism is not supported; only a weak, cautiously stated relativity has modest evidence.

  13. Which of the following statements about second language acquisition and teaching are correct?

    1. Krashen distinguishes subconscious acquisition from conscious learning of rules
    2. An interlanguage is a rule-governed learner system that can fossilize short of the target
    3. Communicative language teaching foregrounds meaningful use rather than isolated rule drills
    4. A proficiency test and an achievement test are simply two names for exactly the same instrument
    Show answer

    Answer: A — Krashen distinguishes subconscious acquisition from conscious learning of rules; B — An interlanguage is a rule-governed learner system that can fossilize short of the target; C — Communicative language teaching foregrounds meaningful use rather than isolated rule drills

    The acquisition–learning distinction, the fossilizing interlanguage, and the meaning-focused nature of communicative teaching are all correct. The last is false: a proficiency test measures general ability independent of a course, while an achievement test measures what a particular course taught.

  14. Which of the following statements about clinical linguistics are correct?

    1. Assessment combines standardized tests with naturalistic language sampling and observation
    2. Disorders are broadly grouped as developmental or acquired by their origin
    3. Evidence-based therapy weighs research evidence together with clinical expertise and client values
    4. A disorder should be classified purely by its surface symptom, ignoring whether its origin is developmental or acquired
    Show answer

    Answer: A — Assessment combines standardized tests with naturalistic language sampling and observation; B — Disorders are broadly grouped as developmental or acquired by their origin; C — Evidence-based therapy weighs research evidence together with clinical expertise and client values

    Combining standardized and naturalistic assessment, grouping disorders by origin, and basing therapy on evidence plus expertise and client values are all sound. The last is wrong: the same surface symptom can be developmental or acquired, so origin matters and symptom-only classification is a known error.

  15. A tokenized sentence consists of 12 tokens. Using a sliding window, how many word bigrams (adjacent two-token sequences) can be extracted from it? Give the answer as a whole number.

    Numerical answer — type the value.

    Show answer

    Answer: 11

    A bigram is a pair of adjacent tokens, and a sequence of n tokens has n−1 adjacent pairs. With n = 12, the number of bigrams = 12 − 1 = 11. The common slip is to answer 12, forgetting that the last token begins no new pair.

  16. A bigram language model is built over a vocabulary of 6 word types. How many distinct ordered bigrams (word_i followed by word_j) are in principle possible, allowing a word to follow itself? Give the answer as a whole number.

    Numerical answer — type the value.

    Show answer

    Answer: 36

    Each ordered bigram chooses a first word and a second word independently from the 6 types, and self-pairs are allowed, so the count is 6 × 6 = 36. The trap is to compute 6 × 5 = 30 by wrongly excluding a word following itself.