Artificial Intelligence 6001, Winter '26
Course Diary (Wareham)

---

Copyright 2026 by H.T. Wareham
All rights reserved

---

Week 1, Week 2, Week 3, (In-class Exam Notes),
Week 4, (end of diary)

Wednesday, January 28 (Lecture #1) (FS)
[Class Notes]

• Introduction: What is Natural Language?
  • A natural language is any language (be it spoken, signed, or written) used by human beings to communicate with each other, cf. artificial languages used to program computers.
  • There are ~7000 natural languages, with ~20 these languages, e.g., Mandarin, Spanish, English, Hindi, and Arabic, being spoken by more than 1% of the world's population.
  • Example: The natural languages spoken by students in this course.
  • Human language allows a complexity of expression and meaning that has to date not been seen in linguistic behaviour in animals, e.g., bird calls,, primate vocalizations, bee dances, which are more akin to lists of codewords with specific meanings that are not combined, e.g., "danger", "(this way to) food".
    • Gorilla ASL experiments suggest that it is not just the lack of an appropriate vocal apparatus -- the ability to combine words to form arbitrarily complex expressions is simply not present in animals.
    • Talking animals are for now firmly in the domain of fiction (and comedy).
  • Example: Not the Nine O'Clock News (1980): Gerald the Gorilla (Link)
  • The ability to acquire and effectively communicate with a natural language is one of the major hallmarks of human intelligence and sentience; those individuals who have for whatever reasons failed to develop this ability are often considered subhuman, e.g., feral children (the Wild Boy of Aveyron), the deaf (Children of a Lesser God, Helen Keller).
  • Example: The Miracle Worker (1961) [excerpt]
  • The ability to acquire and effectively communicate with a natural language is thus one of the major hallmarks of true artificial intelligence.
• Introduction: What is Natural Language Processing (NLP)?
  • NLP is the subfield of Artificial Intelligence (AI) concerned with the computations underlying the processing (recognition, generation, and acquisition) of natural human language (be it spoken, signed, or written).
    • NLP is distinguished from (though closely related to) the processing of artificial languages, e.g., computer languages, formal languages (regular, context-free, context-sensitive, etc).
  • NLP emerged in the early 1950's with the first commercial computers; originally focused on machine translation, but subsequently broadened to include all natural-language related tasks (J&M, Section 1.6; Kay (2003); see also BKL, Section 1.5 and Afterword).
  • Two flavors of NLP: strong and narrow
    • The focus of Strong NLP is on discovering and implementing human-level language abilities using approximately the same mechanisms as human beings when they communicate; as such, it is more closely allied with classical linguistics (and hence often goes by the name "computational linguistics").
    • The focus of Narrow NLP is on giving computers human-level language abilities by whatever means possible; as such, it is more closely allied with AI (and is often referred to by either "NLP" or more specific terms like "speech processing").
    • Narrow NLP is nonetheless influenced by Strong NLP, and vice versa (to the extent that the mechanisms proposed in Narrow NLP to get systems up and running help to revise and make more plausible the theories underlying Strong NLP).
    • In this module, we will look at both types of NLP, and distinguish them as necessary.
• Potential exam questions
• The Characteristics of Natural Language: Why Bother?
  • Natural language is the area of study of linguistics. Looking at what linguists have to say about the characteristics of natural language is very valuable in NLP for two reasons:
    1. Linguists have studied many languages and hence have described and proposed mechanisms for handling phenomena that may lie outside those handled by existing NLP systems (many of which have been developed to only handle English).
    2. Linguists have studied the full range of natural processes, from sound perception to meaning, in some cases for centuries (~2500 and 1300 years, respectively, in the case of Sanskrit and Arabic linguists), and have had insights that may be invaluable in mitigating known problems with existing NLP systems, e.g., Marcus and Davis (2019, 2021).
• The Characteristics of Natural Language: Overview
  • General model of language processing (adapted from BKL, Figure 1.5):

    

    • "A LangAct" is an utterance in spoken or signed communication or a sentence in written communication.
    • The topmost left-to-right tier encodes language recognition and the right-to-left tier immediately below encodes language production.
    • The knowledge required to perform each step (all of which must be acquired by a human individual to use language) is the third tier, and the lowest is the area or areas of linguistics devoted to each type of knowledge.
    • Three language-related processes are thus encoded in this diagram: recognition and production (explicitly) and acquisition (implicitly).
  • Whether one is modeling actual human language or implementing language abilities in machines, the mechanisms underlying each of language recognition, production, and acquisition must be able to both (1) handle actual human language inputs and outputs and (2) operate in a computationally efficient manner.
    • As we will see over the course of this module, the need to satisfy both of these criteria is one of major explanations for how research has proceeded (and diverged) in Strong and Narrow NLP.
• The Characteristics of Natural Language: Phonetics (J&M, Section 7)
  • Phonetics is the study of the actual sounds produced in human languages (phones), both from the perspectives of how they are produced by the human vocal apparatus (articulatory phonetics) and how they are characterized as sound waves (acoustic phonetics).
    • Acoustic phonetics is particularly important in artificial speech analysis and synthesis.
  • In many languages, the way words are written does not correspond to the way sounds are pronounced in those words, e.g., variation in pronouncing [t] (J&M, Table 7.9):
    • aspirated (in initial position), e.g., "toucan"
    • unaspirated (after [s]), e.g., "starfish"
    • glottal stop (after vowel and before [n], e.g., "kitten"
    • tap (between vowels), e.g., "butter"
    The most common way of representing phones is to use special sound-alphabets, e.g., International Phonetic Alphabet (IPA) [charts], ARPAbet [Wikipedia] (J&M, Table 7.1).
  • Each language has its own characteristic repertoire of sounds, and no natural language (let alone English) uses the full range of possible sounds.
  • Example: Xhosa tongue twisters (clicks)
  • Even within a language, sounds can vary depending on their context, due to constraints on the way articulators can move in the vocal tract when progressing between adjacent sounds (both within and between words) in an utterance.
    • word-final [t]/[d] palatalization (J&M, Table 7.10), e.g., "set your", "not yet", "did you"
    • Word-final [t]/[d] deletion (J&M, Table 7.10) e.g., "find him", "and we", "draft the"
  • Variation in sounds can also depend on a variety of other factors, e.g., rate of speech, word frequency, speaker's state of mind / gender / class / geographical location (J&M, Section 7.3.3).
  • As if all this wasn't bad enough, it is often very difficult to isolate individual sounds and words from acoustic speech signals

Friday, January 30 (Lecture #2) (FS)
[Class Notes]

• The Characteristics of Natural Language: Phonology (Bird (2003)) (Youtube)
  • Phonology is the study of the systematic and allowable ways in which sounds are realized and can occur together in human languages.
  • Each language has its own set of semantically-indistinguishable sound-variants, e.g., [pat] and [p^hat] are the same word in English but different words in Hindi. These semantically-indistinguishable variants of a sound in a language are grouped together as phonemes.
  • Variation in how phonemes are realized as phones is often systematic, e.g., formation of plurals in English:

    mop	mops	[s]
    pot	pots	[s]
    pick	picks	[s]
    kiss	kisses	[(e)s]
    mob	mobs	[z]
    pod	pods	[z]
    pig	pigs	[z]
    pita	pitas	[z]
    razz	razzes	[(e)z]

    Note that the phonetic form of the plural-morpheme /s/ is a function of the last sound (and in particular, the voicing of the last sound) in the word being pluralized. A similar voicing of /s/ often (but not always) occurs between vowels, e.g., "Stasi" vs. "Streisand".

  • Such systematicity may involve non-adjacent sounds, e.g., vowel harmony in the formation of plurals in Turkish (Kenstowicz (1994), p. 25):

    dal	dallar	``branch''
    kol	kollar	``arm''
    kul	kullar	``slave''
    yel	yeller	``wind''
    dis	disler	``tooth''
    g"ul	g"uller	``race''

    The form of the vowel in the plural-morpheme /lar/ is a function of the vowel in the word being pluralized.

  • It is tempting to think that such variation is encoded in the lexicon, such that each possible form of a word and its pronunciation are stored. However, such variation is typically productive wrt new words (e.g., nonsense words, loanwords), which suggests that variation is instead the result of processes that transform abstract underlying lexical forms to concrete surface pronounced forms.
    • Plural of nonsense words in English, e.g., blicket / blickets, farg / fargs, klis / klisses.
    • Unlimited concatenated morpheme vowel Harmony in Turkish (Sproat (1992), p. 44):
      c"opl"uklerimizdekilerdenmiydi =>
      c"op + l"uk + ler + imiz + de + ki + ler + den + mi + y + di
      ``was it from those that were in our garbage cans?''

      In Turkish, complete utterances can consist of a single word in which the subject of the utterance is a root-morpheme (in this case, [c"op], "garbage") and all other (usually syntactic, in languages like English) relations are indicated by suffix-morphemes. As noted above, the vowels in the suffix morphemes are all subject to vowel harmony. Given the in principle unbounded number of possible word-utterance in Turkish, it is impossible to store them (let alone their versions as modified by vowel harmony) in a lexicon.

    • Variation in English [r] according Cantonese phonology in Chinenglish (Demers and Farmer (1991), Exercise 3.10):
      • [r] deletion (before consonant or final sound in word), e.g., "tart", "party", "guitar", "bar", "sergeant" (see also Boston English)
      • [r] pronounced as [l] (before a vowel), e.g. "strawberry","brandy", "aspirin", "Listerine", "curry".

      These rules can also hold when a Cantonese speaker writes English text, e.g., "Phone Let Kloss."

    • Modification of English loanwords in Japanese (Lovins (1973)):

      dosutoru	``duster''
      sutoroberri`	`strawberry''
      kurippaa`	`clippers''
      sutoraiki	``strike''
      katsuretsu	``cutlet''
      parusu	``pulse''
      gurafu	``graph''

      Japanese has a single phoneme /r/ for the liquid-phones [l] and [r]; moreover, it also allows only very restricted types of multi-consonant clusters. Hence, when words that violate these constraints are borrowed from another language, those words are changed by the modification to [r] or deletion of [l] and the insertion of vowels to break up invalid multi-consonant clusters.

  • Each language has its own phonology, which consists of the phonemes of that language, constraints on allowable sequences of phonemes (phonotactics), and descriptions of how phonemes are instantiated as phones in particular contexts. The systematicities within a language's phonology must be consistent with (but are by no means limited to those dictated solely by) vocal-tract articulator physics, e.g., consonant clusters in Japanese.
  • Courtesy of underlying phonological representations and processes, one is never sure if an observed phonetic representation corresponds directly to the underlying representation or is some process-mediated modification of that representation, e.g., we are never sure if an observed phone corresponds to an "identical" phoneme in the underlying representation. This introduces one type of ambiguity into natural language processing.
• The Characteristics of Natural Language: Morphology (Trost (2003); Sproat (1992), Chapter 2)
  • Morphology is the study of the systematic and allowable ways in which symbol-string/meaning pieces are organized into words in human languages.
  • A symbol-string/meaning-piece is called a morpheme.
    • Two broad classes of morphemes: roots and affixes.
    • A root is a morpheme that can exist alone as word or is the meaning-core of a word, e.g., tree (noun), walk (verb), bright (adjective).
    • An affix is a morpheme that cannot exist independently as a word, and only appears in language as part of word, e.g., -s (plural), -ed (past tense; 3rd person), -ness (nominalizer).
  • A word is essentially a root combined with zero or more affixes. Depending on the type of root, the affixes perform particular functions, e.g., affixes mark plurals in nouns and subject number and tense in verbs in English.
  • Morphemes are language-specific and are stored in a language's lexicon. The morphology of a language consists of a lexicon and a specification of how morphemes are combined to form words (morphotactics).
  • Morpheme order typically matters, e.g., uncommonly, commonunly*, unlycommon* (English)
  • There are a number of ways in which roots and affixes can be combined in human languages (Trost (2003), Sections 2.4.2 and 2.4.3):
    • Prefix: An affix attached to the front of the root, e.g.,, the negative marker un- for adjectives in English (uncommon, infeasible, immature).
    • Suffix: An affix attached to the back of the root, e.g.,, the plural marker -s for nouns in English (pots, pods, dishes).
    • Circumfix: A prefix-suffix pair that must both attach to the root, e.g.,, the past participle marker ge-/-t for verbs in German (gesagt "said", gelaufen "ran").
    • Infix: An affix inserted at a specific position in a root, e.g., the -um- verbalizer for nouns and adjectives in Bontoc (Philippines):

      /fikas/	"strong"	/fumikas/	"to be strong"
      /kilad/	"red"	/kumilad/	"to be red"
      /fusl/	"enemy"	/fumusl/	"to be an enemy"

    • Template Infix: An affix consisting of a sequence of elements that are inserted at specific positions into a root (root-and-template morphology), e.g., active and passive markers -a-a- and -u-i- for the root ktb ("write") in Arabic:

      katab	kutib	"to write"
      kattab	kuttib	"cause to write"
      ka:tab	ku:tib	"correspond"
      taka:tab	tuku:tib	"write each other"

    • Reduplication: An affix consisting of a whole or partial copy of the root that can be prefix, infix, or suffix to the root, e.g., formation of the habitual-repetitive in Javanese:

      /adus/	"take a bath"	/odasadus/
      /bali/	"return"	/bolabali/
      /bozen/	"tired of"	/bozanbozen/
      /dolan/	"recreate"	/dolandolen/

  • As with phonological variation, there are several lines of evidence which suggest that morphological variation is not purely stored in the lexicon but rather the result of processes operating on underlying forms:
    • Productive morphological combination simulating complete utterances in words, e.g., Turkish example above.
    • Morphology operating over new words in a language, e.g., blicket -> blickets, television -> televise -> televised / televising, barbacoa (Spanish) -> barbecue (English) -> barbecues.
  • As if all this didn't make things difficult enough, different morphemes need not have different surface forms, e.g, variants of "book"
    "Where is my book?" (noun)
    "I will book our vacation." (verb: to arrange)
    "He doth book no argument in this matter, Milord." (verb: to tolerate)

  • Courtesy of phonological transformations operating both within and between morphemes and the non-uniqueness of surface forms noted above, one is never sure if an observed surface representation corresponds directly to the underlying representation or is a modification of that representation, e.g., is [blints] the plural of "blint" or does it refer to a traditional Jewish cheese-stuffed pancake? This introduces another type of ambiguity into natural language processing.
• The Characteristics of Natural Language: Syntax (BKL, Section 8; J&M, Chapter 12; Kaplan (2003))
  • Syntax is the study of the systematic and allowable ways in which words are organized into utterances (spoken) and sentences (written) in human languages.
  • Two fundamental facts of natural language syntax (J&M, pp. 385-386):
    • constituency: a group of one or more (most often, adjacent) words may behave as a single unit called a constituent, e.g., "fox", "dog", "jumped", "the quick brown fox", "the lazy dog", and "jumped over the lazy dog" are constituents in the utterance "the quick brown fox jumped over the lazy dog".
    • grammatical relations: constituents relate in systematic manners to other constituents in a utterance, e.g. "the quick brown fox" and "the lazy dog" are the subject and object, respectively, in the utterance "The quick brown fox jumped over the lazy dog".

  • Constituents typically have associated grammatical classes whose members may have different structures but perform equivalent grammatical functions, e.g., noun phrase, determiner, verb phrase.
    • There are typically multiple hierarchical levels of constituents in an utterance, e.g.,
      

      "[[the [quick [brown [fox]]]] [jumped over [the [lazy [dog]]]]]".

  • Some languages encode grammatical relations primarily by word order, e.g., "John visited Mary" (English); others primarily use morphology, e.g., "John-ga Mary-o tazune-ta", "Mary-o John-ga tazune-ta" (Japanese) (Kaplan (2003), p. 71).
  • Even in languages that primarily use word order and have equivalent grammatical classes, orders of grammatical classes may differ across languages, e.g, each of the six possible orderings of subject, object, and verb in basic transitive-verb utterances occur in at least one human language. though the frequencies of these orders vary greatly (Tomlin (1986), p. 22):

    SOV	"She him loves."	45%	Pashto, Latin, Japanese, Afrikaans
    SVO	"She loves him."	42%	English, Hausa, Mandarin, Russian
    VSO	"Loves she him."	9%	Hebrew, Irish, Zapotec, Tuareg
    VOS	"Loves him she."	3%	Malagasy, Baure
    OVS	"Him loves she."	1%	Apalai, Hixkaryana
    OSV	"Him she loves."	< 1%	Warao

  • Grammatical relations can hold over arbitrary distances in a utterances, e.g., "The man who knew the solder that killed the sailor who sailed as first mate on the S.S. Warsaw and visited my ex-sister-in-law Louise's best friend in Paris last month died." (garden-path utterances)
  • Grammatical relations may even be recursively embedded, e.g., "The cat the dog the rat the elephant admired bit chased likes tuna fish." (J&M, p. 536)
  • Syntax is language-specific. Building on a language's morphology, the syntax is a specification called a grammar of how words are combined to form utterances in that language.
  • Analogous to phonological and morphological variation, the fact that syntax operates productively to generate a potentially infinite number of utterances over a finite lexicon suggests that syntactic patterns are not purely stored in the grammar but are rather the result of grammatical processes operating on underlying forms.
  • As if all this didn't make things difficult enough, utterances expressing very different meanings need not have different surface forms, e.g,
    "Is your refrigerator running?"
    "Put the block in the box on the table."
    "This morning I shot an elephant in my pajamas."
    

    Hence, ambiguity inherent in grammars and lexicons adds yet another type of ambiguity into natural language processing (BKL, pp. 317-318).

Monday, February 2

• University closed; class cancelled.

Wednesday, February 4 (Lecture #3) (FS)
[Class Notes]

• The Characteristics of Natural Language: Semantics, Discourse, and Pragmatics (Lappin (2003); Leech and Weisser (2003); Ramsay (2003))
  • Semantics is the study of the manner in which meaning is associated with utterances in human language; discourse and pragmatics focus, respectively, on how meaning is maintained and modified over the course of multi-person dialogues and how these people chose different individual-utterance and dialogue styles to communicate effectively.
  • Meaning seems to be very closely related to syntactic structure in individual utterances; however, the meaning of an utterance can vary dramatically depending on the spatio-temporal nature of the discourse and the goals of the communicators, e.g., "It's cold outside." (statement of fact spoken in Hawaii; statement of fact spoken on the International Space Station; implicit order to close window).
  • Various mechanisms are used to maintain and direct focus within an ongoing discourse (Ramsay (2003), Section 6.4):
    • Different syntactic variants with subtly different meanings, e.g., "Ralph stole my bike" vs "My bike was stolen by Ralph".
    • Different utterance intonation-emphasis, e.g., "I didn't steal your bike" vs "I didn't steal your BIKE" vs "I didn't steal YOUR bike" vs "I didn't STEAL your bike".
    • Syntactic variants which presuppose a particular premise, e.g., "How long have you been beating your wife?"
    • Syntactic variants which imply something by not explicitly mentioning it, e.g., "Some people left the party at midnight" (-> and some of them didn't), "I believe that she loves me" (-> but I'm not sure that she does).
  • Another mechanism for structuring discourse is to use references (anaphora) to previously discussed entities (Mitkov (2003a)).
    • There are many kinds of anaphora (Mitkov (2003a), Section 14.1.2):
      • Pronominal anaphora, e.g., "A knee jerked between Ralph's legs and he fell sideways busying himself with his pain as the fight rolled over him."
      • Adverb anaphora, e.g., "We shall go to McDonald's and meet you there."
      • Zero anaphora, e.g., "Amy looked at her test score but was disappointed with the results."
    • Though a convenient conversational shorthand, anaphora can be (if not carefully used) ambiguous, e.g.,
      "The man stared at the male wolf. He salivated at the thought of his next meal."
      "Place part A on Assembly B. Slide it to the right."
      "Put the doohickey by the whatchamacallit over there."
      

  • As demonstrated above, utterance meaning depends not only the individual utterances, but on the context in which those utterance occur (including knowledge of both the past utterances in a dialogue and the possibly unknown and dynamic goals and knowledge of all participants in the discourse), which adds yet another layer of ambiguity into natural language processing ...
• It seems then that natural language, by virtue of its structure and use, encodes both a lot of ambiguity and variation, as well as a wide variety of structures at all levels.
• There is a final set of computational characteristics of natural language that must be either accommodated (in the case of strong NLP systems) or circumvented by clever means (in the case of narrow NLP systems).
• The Characteristics of Natural Language: Natural Language Operation and Acquisition
  • Human beings (both adults and children) use language effectively with the aid of finite and ultimately limited computational devices, e.g., our brains.
  • Utterances are generated and comprehended relatively quickly in most everyday situations, cf., bad telephone connections active construction sites.
  • Generation and performance acquisition degrades more or less gracefully with the increasing complexity and length of utterances.
  • Example: Prototypical natural language acquisition device
  • Given exposure to a human language, any child can acquire the phonological, morphological, and semantic (if not total lexical) knowledge of that language within four to seven years.
    • This exposure does not need to be simplified, cf., Motherese.
    • This exposure does not need to be comprehensive or exhaustive, with many examples of each possible phenomena in the language.
    • Individual utterances in this exposure need not be either grammatical or have their grammaticality-status stated.
    • Produced utterances by the child need not be assessed for grammaticality or corrected.
  • Given exposure to any human language, any human being can acquire that language, and children under the age of 10 years are superbly good at this. In addition:
    • If children are exposed to two distinct language simultaneously, they will acquire both languages, i.e., they will become bilingual.
    • If children are exposed to a pidgin created by combining elements of two or more languages, they will regularize the unsystematic parts of the pidgin to create a proper hybrid language (creole).
• Given all the characteristics of natural language discussed in the previous lectures and their explicit and implied constraints on what an NLP system must do, what then are appropriate computational mechanisms for implementing NLP systems? It is appropriate to consider first what linguists, the folk who have been studying natural language for the longest time, have to say on this matter.
• NLP Mechanisms: The View from Linguistics
  • Three basic types of mechanisms are required to implement NLP:
    1. Sequences (to represent linguistic signals at various levels of processing).
    2. Sets (to represent sets of linguistic entities, e.g., lexicons)
    3. Processes (to implement proposed linguistic processes operating both between and within levels of processing).

  • Given that linguistic signals are expressed as temporal (acoustic and signed speech) and spatial (written sequences) sequences of elements, there must be a way of representing such sequences.
    • Sequence elements can be atomic (e.g., symbols) or have their own internal structure (e.g., feature matrices (phones / phonemes), form-meaning bundles (morphemes)); for simplicity, assume for now that elements are symbols.
    • There are at least two types of such sequences representing underlying and surface forms.
    • Where necessary, hierarchical levels of structure such as syntactic parse trees can be encoded as sequences by using appropriate interpolated and nested brackets, e.g., "[[the quick brown fox] [[jumped over] [the lazy brown dog]]]"
  • The existence of lexicons implies mechanisms for representing sets of element-sequences, as well as accessing and modifying the members of those sets.
  • The various processes operating between underlying and surface forms presuppose mechanisms implementing those processes.
    • The most popular implementation of processes is as rules that specify transformations of one form to another form, e.g., add voice to the noun-final plural morpheme /s/ if the last sound in the noun is voiced.
    • Rules are applied in a specified order to transform an underlying form to its associated surface form for utterance production, and in reverse fashion to transform a surface form to its associated underlying form(s) for utterance comprehension (ambiguity creating the possibility of multiple such associated forms).
    • Rules can be viewed as functions. Given the various types of ambiguity we have seen, these functions are at least one-many.
  • In each linguistic level, the functions corresponding to individual rules can be composed in rule application order to create "master" functions that transform between representations on adjacent linguistic levels, e.g.,

    

    In this diagram, the inverse of function X() is written as X-().
  • The above corresponds to classical appproaches to NLP; alternatively, natural language can be seen as a single function that is the composition of all of these individual level-functions or their inverses, e.g., for meaning m and utterance u, uttterance generation is u = G(m) = P2(P1(M(S(m))))) and utterance comprehension is m = C(u) = S-(M-(P2-(P1-(u))))).
    • This alternative is that taken by many neural network implementations of NLP (which will be discussed later in this module).
    • Some such systems like ChatGPT go further by going directly from utterances to replies, e.g., u' = R(u) = P2(P1(M(S(S-(M-(P2-(P1-(u))))). This raises the question if such systems ever actually compute (even internally) meanings associated with utterances at all, cf. Hinton's "Thought Vectors" (see notes on Encoders/Decoders below).
  • Great care must be taken in establishing what parameters are necessary as input to these functions and that such parameters are available. For example, a syntax function can get by with a morphological analysis of an utterance, but a semantics function would seem to require as input not only possible syntactic analyses of an utterance but also discourse context and models of discourse participant knowledge and intentions.
  • Following practice in linguistics, we can assume that all natural language processes operate in a deterministic fashion. However, given the complexity of proposed processes and how they may interact, it may be useful as an intermediate stage in analyzing linguistic complexity to allow mechanisms that operate probabilistically, e.g., rules that have a specified probability of application, composed systems of rules that rank one-many outputs by their probability of occurrence.
    • This is analogous to the use of statistical mechanics to summarize and analyze the overall behavior of large groups of particles in physics.
    • Such mechanisms may be analysis end-points in themselves if it turns out that either the interaction of deterministic linguistic processes is too complex to untangle or there are genuinely probabilistic processes underlying natural language processing in humans.
    • Beware of over-simplification for the sake of engineering convenience and equating the model with the phenomena, e.g., complex deterministic one-many -> simplified probabilistic one-many -> simplified probabilistic one-one / most probable (the last of which is what neural-network-based NLP systems do).
• Potential exam questions

Friday, February 6 (Lecture #4) (FS)
[Class Notes]

• NLP Mechanisms: Finite-state automata (FSA) (Martin-Vide (2003), Section 8.3.1; Nederhof (1996); Roche and Schabes (1997a), Section 1.2)
  • Originated in the Engineering community; inspired by the discrete finite neuron model proposed by McCullough and Pitts in 1943.
  • Basic Terminology: (start / finish) state, (symbol-labeled) transition, acceptance.
  • Operation: generation / recognition of strings
    • Generation builds a string from left to right, adding symbols to the right-hand end of the string as one progresses along a transition-path from the start state to a final state.
    • Recognition steps through a string from left to right, deleting symbols from the left-hand end of the string as one progresses along a transition-path from the start state to a final state.
    • Note that one need only keep track of one FSA-state during both operations described above, namely, the last state visited.
  • Example #1: A DFA for the set of all strings over the alphabet {a,b} that consist of one or more concatenated copies of ab.

    

  • Types of FSA:
    • Deterministic (DFA): At each state and for each symbol in the alphabet, there is at most one transition from that state labeled with that symbol.
    • Non-Deterministic (NFA): At each state, there may be more than one transition from that state labeled with a particular symbol and/or there may be transitions labeled with special symbol epsilon (denoted in our diagrams by symbol E).
      • Revised notion of string acceptance: see if there is any path through the NFA that accepts the input string.
  • Example #2: An NFA for the set of all strings over the alphabet {a,b} that start with an a and end with a b (uses multiple same-label outward transitions).

    

  • DFA can recognize strings in time linear in the length of the input string, but may not be compact; NFA are compact but may require time exponential in the length of a string to recognize that string (need to follow all possible computational paths to check string acceptance).
  • Example #3: An NFA for the set of all strings over the alphabet {a,b} that start with an a and end with a b (uses epsilon-transitions).

    

  • Oddly enough, deterministic and non-deterministic FSA are equivalent in recognition power, i.e., there is a DFA for a particular set of strings iff there is an NFA for that set of strings.
  • Example #4: A DFA for the set of all strings over the alphabet {a,b} that start with an a and end with a b.

    

  • There are a variety of operations for combining pairs of finite-state automata into a single automaton, e.g., concatenation, union, intersection.
  • Application: Building lexicons
    • The simplest types of lexicons are lists of morphemes which, for each morpheme, store the surface form and syntactic / semantic information associated with that morpheme.
    • A list of surface forms can be stored as FSA in two ways:
      • DFA encoding of a trie (= retrieval tree)
      • Deterministic acyclic finite automaton (DAFA)
      One creates a trie by compacting a word-list and a DAFA for a word-list by compacting a trie-DFA for that word-list.
    • Example #5: A trie-DFA encoding the lexicon {"in", "inline", "input", "our", "out", "outline", "output"}.

      

    • Example #6: A DAFA encoding the lexicon {"in", "inline", "input", "our", "out", "outline", "output"}.

      

  • Application: Implementing morphotactics
    • More complex lexicons need to take into account how morphemes are combined to form words.
    • Example #7: A DFA implementing the constraint that the voicing of the noun-final plural morpheme must match the voicing of the final sound in the noun being pluralized.

      

    • The simplest types of purely concatenative morphotactics can be encoded as a finite-state manner by indicating for each morpheme-type sublexicon which types of morphemes can precede and follow morphemes in that sublexicon, e.g., nouns precede pluralization-morphemes in English.
    • More complex morphotactics, e.g., circumfixes, infixes, reduplication, can be layered on top of lexicon FSA using additional mechanisms (Beesley and Karttunen (2000); Beesley and Karttunen (2003), Chapter 8; Kiraz (2000)).
• NLP Mechanisms: Finite-state transducers (FST) (J&M, Section 3.4, Mohri (1997); Roche and Schabes (1997a), Section 1.3)
  • Generalization of FSA in which each transition is labeled with a symbol-pair (either or both of which may be the special symbol epsilon).
    • The first symbol in each pair is called the lower symbol and the second is called the upper symbol.
    • The alphabet of all lower-symbols in an FST need not have a non-empty intersection with the alphabet of all upper-symbols in an FST.
  • Example #8: An FST for transforming between the set of all strings over the alphabet {a,b} that consist of n, n >= 0, concatenated copies of ab and the set of all strings over the alphabet {c,d} that consist of n, n >= 0, concatenated copies of cd.

    

  • Operation: generation / recognition of string-pairs, reconstruction of upper (lower) string associated with given lower (upper) string.
    • Generation of a string-pair builds that pair from left to right, adding symbol-pairs to the right-hand end of the string-pair as one progresses along a transition-path from the start state to a final state.
    • Recognition of a string-pair proceeds by stepping through that pair from left to right, deleting symbol-pairs from the left-hand end of the string-pair as one progresses along a transition-path from the start state to a final state.
    • Reconstruction of a string-pair associated with a given string builds the missing string of the pair from left to right, adding missing-string symbols to the right-hand end of the missing string as one progresses along a transition-path from the start state to a final state in accordance with the given string.
      • As the given string may be either the lower or upper string, there are two types of string-pair reconstructions.
      • Each reconstructed string-pair corresponds to a particular path through the FST guided by the given string.
      • Depending on the structure of the FST, there may be more than one path through the FST consistent with the given string, and hence more than one string-pair reconstruction.
    • Note that one need only keep track of one FST-state during all operations described above, namely, the last state visited.
  • Unlike FSA, there are many types of FST and many types of FST (non-)determinism.
  • Example #9: An FST for transforming between the set of all strings over the alphabet {a,b} that consist of n, n >= 0, concatenated copies of ab and the set of all strings over the alphabet {a,b} that consist of n, n >= 0, concatenated copies of aba.

    

  • There are a variety of operations for combining pairs of finite-state transducers into a single transducer, e.g., composition, intersection.
  • Application: Implementing rule-based systems
    • Many types of linguistic rules can be implemented as FST.
    • Example #10: An FST implementing the rule that voice be added to the noun-final plural morpheme /s/ if the last sound in the noun is voiced.

      

    • The sequential application of rules can be simulated using the FST created by the sequential composition (in the specified order) of the FST corresponding to the rules.
  • Application: Implementing lexical transducers (Beesley and Karttunen (2003), Section 1.8)
    • A lexicon FSA can be converted into an FST by replacing each transition-label x with x:x, e.g., the identity FST associated an FSA. This lexicon FST can then be composed with the the underlying / surface transformation FST encoding a constraint- or rule-based morphonological system to create a lexical transducer.
    • Note as FST are bidirectional, so are lexical transducers; hence, can reconstruct surface or lexical forms from given lexical or surface forms with equal ease.
• Potential exam questions

Monday, February 9 (Lecture #5) (FS)

• Finite-state mechanisms can handle many basic linguistic phenomena like concatenation, contiguous long-distance dependency, and finite numbers of long-distance dependencies, e.g., local phonological modification, vowel harmony, Turkish morphology, garden-path utterances. However, it can be formally proved (see J&M, Section 16.2.2 and references) that finite-state mechanisms cannot handle more complex phenomena like unbounded numbers of long-distance dependencies, e.g., phonological reduplication, recursively embedded-relation utterances. This requires at least the power of context-free mechanisms.
• NLP Mechanisms: Context-free grammars (BKL, Section 8.3; HU79, Chapter 5; Martin-Vide (2003), Sections 8.2.1 and 8.3.2)
  • Originated in the Linguistics community; proposed by Chomsky in the 1950s as the second level of the Chomsky Hierarchy of grammars (Regular / Context-Free (Phrase Structure) / Context-Sensitive / Unrestricted).
  • Basic terminology: (production) rule, terminal (plus special symbol epsilon), non-terminal, rule application, termination.
    • Rules are restricted such that left-hand side is a single non-terminal and the right-hand side s of length at least one consisting of any sequence of terminals and/or non-terminals.
  • Operation: generation / recognition of strings
    • Both generation and recognition create parse trees such that each internal node and its immediate children in the tree corresponds to the application of a rule in a CFG.
  • Example #1: A CFG for the set of all strings over the alphabet {a,b} that consist of n >= 1 a's followed by n b's.

    S -> aSb | ab
    

    

  • Example #2: A CFG for a finite-size subset of all utterances:

    S -> Np Vp
    Np -> "the cat" | "the dog" | "the rat" | "the elephant"
    Vp -> V Np
    V -> "admired" | "bit" | "chased" 
    

    

  • Example #3: A CFG for an infinite-size subset of all garden-path utterances, e.g., "The cat that chased the dog that bit the rat died.":

    S -> Np Sp "died" | NP "died"
    Np -> "the cat" | "the dog" | "the rat" | "the elephant"
    Sp -> that" Vp Sp | "that" Vp
    Vp -> V Np
    V -> "admired" | "bit" | "chased" 
    

    

  • Example #4: A CFG for an infinite-size subset of the set of all recursively-embedded utterances.

    S -> Np Sp "likes tuna fish" | Np "likes tuna fish"
    Np -> "the cat" | "the dog" | "the rat" | "the elephant"
    Sp -> Np Sp V | Np V
    V -> "admired" | "bit" | "chased"
    

    

  • Example #5: A CFG for an infinite-size subset of the set of all utterances, e.g. "the dog saw a man in the park" (BKL, pp. 298-299).

    S -> Np Vp 
    Np -> "John" | "Mary" | "Bob" | Det N | Det N Pp
    Det -> "the" | "my"
    N -> "man" | "dog" | "cat" | "telescope" | "park"
    Pp -> P Np
    P -> "in" | "on" | "by" | "with"
    Vp -> V Np | V Np Pp
    

    

    Unlike the other sample grammars above, this grammar is structurally ambiguous because there may be more than one parse for certain utterances involving prepositional phrases (Pp) as it is not obvious which noun phrase (Np) a Pp is attached to, e.g., in "the dog saw the man in the park", is the man (above) or the dog (below) in the park?

    

  • Note that the parse trees encode grammatical relations between entities in the utterances, and that these relations have associated semantics; hence, one can use parse trees as encodings of basic utterance meaning!
    • As shown in Example #5 above, parse trees via structural ambiguity can nicely encode semantic ambiguity.
  • Context-free mechanisms can handle many complex linguistic phenomena like unbounded numbers of recursively-embedded long-distance dependencies. This done by parsing, which both recognizes and adds an internal hierarchical constituent phrase-structure of a sentence.
    • Parsing a sentence S with n words relative to a grammar G can be seen as a search over all possible parse trees that can be generated by grammar G with the goal of finding all parse trees whose leaves are labelled with exactly the words in S in an order that (depending on the language, is either exactly or is consistent with) that in S.
  • Though some algorithms implement the generation of one valid parse for a given sentence much quicker than others, the generation of all valid search trees (which is required in the simplest human-guided schemes for resolving sentence ambiguities) for all parsing algorithms is in the worst case at least the number of valid parse trees for the given sentence S relative to the given grammar G.
    • There are natural grammars for which this is quantity is exponential in the number of words in the given sentence (BKL, p. 317; Carpenter (2003), Section 9.4.2).
  • Dealing with very large numbers of output derivation-trees is problematic. What can be done? Can deal with this by using probability -- that is, by only computing the derivation-tree with highest probability for the given utterance!
  • This is a general way to handle the many-results problem with ambiguous linguistic processes. There are many models of probabilistic computing with context-free grammars and finite-state automata and transducers, and many of these were state-of-the-art for NLP for several decades. However, as Li (2022) points out, the more purely mathematical probabilistic modeling of natural language started over a century ago with Markov Models is actually the research tradition that leads to the final NLP implementation mechanism we shall examine -- namely, (artificial) neural networks.
• Potential exam questions
• NLP Mechanisms: Neural Networks (Kedia and Rasu (2020), Chapters 8-11)
  • The finite-state and context-free NLP mechanisms we have studied thus far in the module are all process-oriented, as they all implement rules (explicitly in the case of grammars, implicitly via the state-transitions in automata and transducers) that correspond to processes postulated by linguists on the basis of observed human utterances.
  • Neural network (NN) NLP mechanisms, in contrast, are function-oriented, in that the components of these mechanisms typically do not have linguistic-postulated correlates and the focus is instead on inferring (with the assistance of massive amounts of observed human utterances) various functions associated with NLP.
    • These functions map linguistic sequences onto categories, e.g., spam e-mail detection, sentiment analysis, or other linguistic sequences, e.g., Part-of-Speech (PoS) tagging, chatbot responses, machine translation between human languages.
    • In many cases, these linguistic sequences are recodings of human utterances in terms of multi-dimensional numerical vectors or matrices. There a variety of methods for creating these vectors and matrices, e.g., bag of words, n-grams, word / sentence / document embeddings (Kedia and Rasu (2020), Chapters 4-6; Vajjala et al (2020), Chapter 3).

  • Neural network research can be characterized to date in three waves:
    • First Wave (1943-1968): McCulloch and Pitts propose abstract neurons in 1943. Starting in the late 1950s, Rosenblatt explores the possibilities for representing and learning functions relative to single abstract neurons (which he calls perceptrons). The mathematical principles underlying the back propagation procedure for training neural networks are developed in the early 1960's (Section 5.5, Schmidhuber (2015)). Perceptron research is killed off by the publication in 1968 of Minsky and Papert's monograph Perceptrons, in which they show by rigorous mathematical proof that perceptrons are incapable of representing many basic mathematical functions such as Exclusive OR (XOR).
    • Second Wave (1980-1990): Rumelhart, McClelland, and colleagues propose and explore the possibilities for multi-level feed-forward neural networks incorporating hidden layers of artificial neurons. This is aided immensely by the (re)discovery of the backpropagation procedure, which allows efficient learning of arbitrary functions by multi-layer neural networks. Though it is shown that these networks are powerful (Universal Approximation Theorems state that any well-behaved function can be approximated to an arbitrarily close degree by a neural network with only one hidden layer), research remains academic as backpropagation on even small to moderate-size networks is too data- and computation-intensive.
      • It is during this period that NN-based NLP research flourishes, taking up over a third of the second volume of the 1987 summary work Parallel Distributed Processing (PDP). Alternatives to feed-forward NN (Recurrent Neural Networks (see below)) for NLP are also explored by Elman starting in 1990.
    • Third Wave (2000-now): With the availability of massive amounts of data and computational power courtesy of the Internet and Moore's Law (as instantiated in special-purpose processors like GPUs), neural network research re-ignites in a number of areas, starting with image processing and computer vision. This is aided by the development of neural networks incorporating special structures inspired by structures in the human brain, e.g., Convolutional Neural Networks, Long Short-Term Memory cells (see below). Starting around 2010, this wave reaches NLP; the results are so spectacular that by 2018, NN-based NLP techniques are state of the art in many applications.
      • This is in large part because the more complex types of neural networks have enabled the creation of pre-trained NLP models that can subsequently be customized with relatively little training data for particular applications, e.g., question answering systems (Li (2022)).

Wednesday, February 11 (Lecture #6) (FS)

• NLP Mechanisms: Neural Networks (Kedia and Rasu (2020), Chapters 8-11) (Cont'd)
  • Let us now explore this research specific to NLP by examining the various types of neural networks that have emerged during these three waves.
    • Feed-forward Multi-layer Neural Networks (FF-NN) (Kedia and Rasu (2020), Chapter 8)
      • In FF-NN, each neuron has n inputs x1, x2, ..., xn, input-specific weights w1, w2, ..., wn, a bias-term b, an activation function f(), and an output y = f((x1 * w1) + (x2 * w2) + .... + (xn * wn) + b).
        • The bias term is adjusted to make training easier.
        • There are several kinds of activation functions, e.g., sigmoid, ReLU (Rectified Linear Unit) (Kedia and Rasu (2020), pp. 181-184); all introduce necessary non-linearity into the otherwise vanilla linear-equation-system behaviour of an artificial neuron.

        

      • In a multi-layer FF-NN, there is an input layer consisting of n inputs, one or more hidden layers of possibly different numbers of artificial neurons, and an output layer consisting of one or more single artificial neurons. Each layer is fully connected to the next, and there are no connections between pairs of neurons in the same layer.

        

      • Neuron connection-weight and bias are inferred by (supervised) training.
        • Initially, all weights and bias are given random values.
        • Given a training set of input-output value-pairs specifying the wanted FF-NN behaviour, this training set is run through the FF-NN multiple times and the backpropagation procedure is applied to the FF-NN to modify the appropriate neuron weights and biases to reduce any difference between the generated and wanted outputs. This is done by exploiting greedily smooth gradients in loss functions computed from generated and wanted values.
      • Once backpropagation has converged on good weight and bias values for the FF-NN relative to the training set, the FF-NN is typically evaluated relative to a separate test set to see how well the FF-NN's behaviour generalizes.
        • FF-NN may exhibit underfitting (performance is bad on both the training and test sets) or overfitting (performance is good on the training set but bad on the test set, i.e., the FF-NN's behaviour does not generalize).
        • Underfitting is dealt with by modifications to the FF-NN's architecture, e.g., number of hidden layers, number of neurons per layer.
        • There are a variety of techniques for dealing with overfitting, e.g.,
          • more training data (add more input-output pairs to the training set).
          • regularization (add a term to penalize too good a fit between the generated and wanted output).
          • early stopping (stop training before weight- and bias-value convergence has occurred).
          • dropout (remove the effects of some percentage of randomly-selected neurons in the FF-NN in each training run).
      • Courtesy of the Universal Approximation Theorems, it is known that FF-NN are powerful in theory; however, converging on the promised network weights and bias in practice by procedures like backpropagation is difficult for several reasons:
        • Large layers and full connections between them require both lots of training data and very large (often exponential) numbers of training runs to converge, cf., so-called "few-shot" learning of natural language by human beings.
        • When layers are reduced in size by having many smaller layers, the values of the loss-function gradients exploited by backpropagation may explode or vanish, leading to wild swings in weight and bias values across training runs or no changes at all in weight and bias values in layers further from the output layer.

        These problems have been dealt with in more advanced types of neural networks by adding additional structures and modules in the network architecture such that embedded FF-NN are smaller and flatter.

    • Convolutional Neural Networks (CNN) (Kedia and Rasu (2020), Chapter 9)

      

      • First proposed for image processing and computer vision in the late 1980's, by analogy with known neural structure in the human visual cortex. CNN are designed to detect and compute over spatial patterns formed by input elements that are close together, e.g., lines / line crossings / corners within images.
      • A CNN consists of a series of feature detectors over a 2D input field, followed by "pooling" layers that summarize detected features and an FF-NN that takes the outputs of the pooling layers as inputs.
        • Patterns are encoded as small square filter matrices called kernels, and pattern detection is implemented by multiplying filter matrices by a version of the 2D input matrix that is padded with zeroes on the right and lower edges to ensure that the filter is applied to all elements of the input matrix.
        • Pooling of matrices produced by feature detection summarizes non-overlapping blocks of values by various techniques, e.g., block maximum / average / sum; the resulting downsampling not only reduces the amount of data input to the FF-NN but also helps prevent overfitting.
      • Though they have been successful in certain applications, e.g., image classification, CNN are not good at detecting long-distance sequence or temporal relationships between input elements.
    • Recurrent Neural Networks (RNN) (Kedia and Rasu (2020), Chapter 10)
      • RNN enable recognition and processing of temporal relationships in sequences by gifting neural networks with memory of past activities.
      • An individual RNN cell is a FF-NN which produces an output yt at time t using the input xt at time t (that is, the element at position t in the input sequence) and the hidden state h(t-1) of the cell at time t-1.
      • An RNN can be "unrolled" relative to its operation over time. In this unrolled version, note that the FF-NNs in all timesteps are identical.

        

      • RNN can be trained by backpropagation that moves backwards in time rather than layers from the output, where the initial hidden state consists of random values. This is best visualized relative in the unrolled version. That being said, it is important to note that the only FF-NN whose weights and biases are changed is that corresponding to timestep zero -- in all other cells, only errors and gradients are propagated.
      • There are several types of RNN, each of which is used in particular applications (shown below in their unrolled versions). These types vary in the relationship between the lengths of their input and output sequences.

        

        • One-many RNN are good for generating narratives.
        • Many-one RNN are good for implementing functions over sequences, e.g., spam e-mail detection, sentiment analysis.
        • Many-many RNN are good for implementing functions between same-length sequences, e.g., PoS tagging, or different-length sequences, e.g., machine translation.
      • Though powerful, RNN are unfortunately deep networks as input and output sequences can be long (particular in NLP). Hence, they are particularly prone to the exploding and vanishing gradient problems noted above. Additional problems arise from the necessity of large hidden states to process long sequences.
    • Long Short-Term Memory (LSTM) Cells (Kedia and Rasu (2020), Chapter 10)
      • One can get the advantages of RNN wrt temporal processing without many of the disadvantages introduced by extended recurrent processing over time by replacing the RNN cells in unrolled versions with LSTM cells.
      • LSTM cells (first proposed in 1995) allow dramatic reductions in the sizes of the hidden states by allowing both the forgetting of older irrelevant and the remembering of new relevant information.
      • Each LSTM cell consists of an input-output-chained sequence of three gates -- namely, a forget gate, an input gate, and an output gate.
      • In addition to dramatically reducing the required sizes of hidden states, LSTM-based neural networks eliminate the vanishing (and mitigate to a degree the exploding) gradient problem because backpropagation training for a particular cell only needs to take into account the cells immediately before and after it in the LSTM-based neural network.
    • Encoders and Decoders (Kedia and Rasu (2020), Chapter 11)
      • The encoder / decoder architecture, first proposed in 2014, implements sequence-to-sequence functions (Seq2Seq) by (1) building on the many-many RNN architecture, (2) replacing RNN cells with LSTM cells, and (3) allowing more structure in the hidden state passed from the final encoder cell to the first decoder cell (context vector).
        • Hinton has on at least one occasion proposed that context vectors be considered "thought vectors".
      • In this architecture, the initial encoder hidden state consists of random values and training focuses on the decoder such that the training set consists of context vector / target sequence pairs.
        • The production of a target sentence by the decoder is triggered by the input of a start token and ended by the output of an end token.

        

      • Fixed-length context vectors are problematic if target sequences that must be produced by the decoder are longer than the source sequences used to create context vectors in the encoder. This has been mitigated by incorporating an attention mechanism into the encoder / decoder architecture.
        • In this mechanism, all (not just the final) hidden states produced by the encoder are available to every cell in the decoder.
        • This set of hidden states is then weighted in a manner specific to each decoder cell to indicate the relevance of each encoder-input token to that cell's output token.
        • Essentially, this enables decoder cells to have individually-tailored variable-length context vectors.
    • Transformers (Kedia and Rasu (2020), Chapter 11)
      • The Transformer architecture, first proposed in 2017, retains the notion of encoders and decoders but discards the many-many RNN backbone. The encoders and decoders are placed in an encoder stack and a decoder stack, respectively.
      • Starting 2018, Transformers have become the state-of-the-art NLP techniques courtesy of two frameworks, BERT and GPT, both of which pre-train generic Transformer-based systems for a specific language that can subsequently by fine-tuned for particular applications.
        • BERT (Bidirectional Encoder Representation from Transformers)
          • Created at Google in 2018.
          • Two versions: BERT_BASE (which uses 12 encoders) and BERT_LARGE (which uses 24 encoders).
          • Both trained on Internet corpus totalling 3.3 billion words.
          • Code and training datasets are open-source.
        • GPT (Generative Pre-trained Transformers)
          • Created at OpenAI in 2018.
          • Four versions released so far (GPT-1, GPT-2, GPT-3, and GPT-4); though initially open-source, by GPT-3 release, full code access is only available to Microsoft employees and all others can only see APIs.
          • Also trained on massive (though proprietary) dataset; though GPT-3 architecture details not known, has 175 billion trainable parameters, which is 500 times larger than BERT's biggest version (Edwards (2021), p. 9).
      • Though very successful in some applications, e.g., question-answering systems and machine translation, and on certain benchmarks, e.g., the Large-scale Reading and Compression (RACE) tool for assessing high-school level understanding of text, current Transformer-based NLP systems produce nonsense in other situations, e.g., "a car has four wheels" vs. "a car has two round wheels" (Edwards (2021), p. 10). It is increasingly being acknowledged that attention and self-attention are not enough, and that some way must be found to integrate world knowledge to improve performance (Reis et al (2021); Edwards (2021); Marcus and Davis (2019, 2021); see also below and Haigh (2023a, 2023b, 2024a, 2024b, 2025)).

Thursday, February 12

• In-class Exam Notes
  I've finished making up the in-class exam. The exam will be closed-book and written in-class on paper (please bring your own pens). It will be 50 minutes long and has a total of 50 marks (this is not coincidental; I have tried to make the number of marks for a question approximately equal to the number of minutes it should take you to do it). The exam will cover material in all Module 2 lectures. There will be two questions, each with multiple parts. The distribution of question-parts and marks by topic are as follows:
  • Characteristics of natural language and NLP (3 parts / 14 marks)
  • Natural language processing mechanisms (4 parts / 26 marks)
  • NLP applications (2 parts / 10 marks)

  I hope the above helps, and I wish you all the best of luck with this exam.

Friday, February 13 (Lecture #7) (FS)

• NLP Mechanisms: Neural Networks (Kedia and Rasu (2020), Chapters 8-11) (Cont'd)
  • Working with Neural Networks (Vajjala et al (2020), Chapter 1)
    • In the eyes of some industry-oriented practitioners, rule- and NN-based components have different roles over the lifetime of modern NLP systems (Vajjala et al (2020), p. 18):
      • Rule-based are excellent for constructing initial versions of NLP systems before task-specific NLP data is available, and are useful in delineating exactly what the system must do.
      • Once (a large amount of) task-specific NLP data is available, machine learning and NN-based components can be employed to rapidly create better production-level systems.
      • Given that systems created by machine-learning and NN-based components make mistakes, these mistakes can be rectified as needed in a mature production system by using rule-based component "patches".
    • That being said, NN-based systems are not yet the silver bullet for NLP for a number of reasons (Vajjala et al (2020), pp. 28-31):
      • Overfitting small datasets (With increasing number of parameters comes increasing expressivity; hence, if insufficient data is available, NN models will overfit and not generalize well).
      • Few-shot learning and synthetic data generation (Though techniques have been developed in computer vision to allow few-shot learning, no such techniques have been developed relative to NLP).
      • Domain adaptation (NN models trained on text from one domain, e.g., internet texts and product reviews, may not generalize to domains which are syntax- and semantic-structure specific, e.g., law, healthcare. In those situations, rule-based models explicitly encoding domain knowledge will be preferable).
      • Interpretable models (Unlike rule-based NLP systems, it is often difficult to interpret why NN models behave as they do; this is a requirement made by businesses).
        • For example, though Transformer-based systems readily associate words with context, it remains far from clear what relationship is actually learned in training ("BERTology") (Edwards (2021), p. 10).
      • Common sense and world knowledge (Current NN models may perform well on standard benchmarks but are still not capable of common sense understanding and logical reasoning. Both of these require world knowledge, and attempts to integrate such knowledge into NN-based NLP systems have not yet been successful).
      • Cost (The data, specialized hardware, and lengthy computation times required to both train and maintain over time large NN-based NLP systems may be prohibitive for all but the largest businesses).
      • On-device deployment (For applications that require NLP systems to be embedded in small devices rather than the computing cloud (e.g., internet-less simultaneous machine translation), the memory- and computation-size of NN-based NLP systems may not be practical).

      Indeed, the above suggests that rule-based NLP systems may be much more applicable than is sometimes thought, and that a fusion of NN- and rule-based NLP system components (or a fundamental re-thinking of NLP system architecture that incorporates the best features of both) may be necessary to create the practical real-world NLP systems of the future.

• Potential exam questions
• Applications: Partial Morphological and Syntactic Parsing
  • As we have seen in previous lectures, complete morphological parsing of words can be done using finite-state mechanisms and complete syntactic parsing of utterances can be done using context-free mechanisms. However, there are a number of applications in which partial ("shallow") morphological and syntactic parsing are desirable.
  • Stemmers and Lemmatizers (Sub-word Morphological Parsing) (BKL, Section 3.6; J&M, Section 3.8)
    • Given a word, a stemmer removes known morphological affixes to give a basic word stem, e.g., foxes, deconstructivism, uncool => fox, construct, cool; given a word which may have multiple forms, a lemmatizer gives the canonical / basic form of that word, e.g., sing, sang, sung => sing.
    • Stemmers and lemmatizers are used to "normalize" texts as a prelude to selecting content words for a text or selecting general word-forms for searches, e.g., process, processes, processing => process, in information retrieval.
    • There are a variety of stemmers implementing different degrees of affix removal and hence different levels of generalization. It is important to chose the right stemmer for an applications, e.g., if "stocks" and "stockings" are both stemmed to "stock", financial and footwear queries will return the same answers.
  • Part-of-Speech (POS) Tagging (Single-level Syntactic Parsing) (BKL, Chapter 5; J&M, Chapter 5)
    • Given a word in an utterance, a POS tagger returns a guess for the POS of that word relative to a fixed set of POS tags.
    • The output of POS taggers are used as inputs to other partial parsing algorithms (see below), automated speech generation (to help determine pronunciation, e.g., "use" (noun) vs. "use" (verb)), and various information retrieval algorithms (see below) (to help locate important noun or verb content words).
    • Existing POS algorithms have differing performances, knowledge-base requirements, and running times and space requirements. Hence, no POS tagger is good in all applications.
  • Chunking (Few-level Syntactic Parsing) (BKL, Chapter 7; J&M, Section 13.5)
    • Given an utterance, a chunker returns a non-overlapping (but not necessarily total in terms of word coverage) breakdown of that utterance into a sequence of one or more basic phrases or chunks, e.g., "book the flight through Houston" => [NP(the flight), PN(Houston)].
    • A chunk typically corresponds to a very low-level syntactic phrase such as a non-recursive NP, VP, or PP.
    • Chunkers are used to isolate entities and relationships in texts as a pre-processing step for information retrieval.
• Applications: Language Understanding (Utterance) (BKL, Section 10; J&M, Section 17)
  • When considering an utterance in isolation from its surrounding discourse, the goal is to infer a semantic representation of the utterance giving the literal meaning of the utterance.
    • Note that this is not always appropriate, e.g., "I think it's lovely that your sister and her five kids will be staying with us for a month", "Would you like to stand in the corner?".
    • A commonly-used representation for many simple applications is First-Order Logic (FOL) (propositional logic augmented with predicates and (possibly embedded) quantification over sets of one or more variables), e.g.,
      • "Everybody loves somebody sometime." ->
        FORALL(x) person(x) => EXISTS(y,t) (person(y) and time(t) and loves(x, y, t))
      • "Every restaurant has a menu" ->
        FORALL(x) restaurant(x) => EXISTS(y) (menu(y) and hasMenu(x,y)) [All restaurants have menus]
      • "Every restaurant has a menu" ->
        EXISTS(y) and FORALL(x) (restaurant(x) => hasMenu(x,y)) hasMenu(x,y)) [Every restaurant has the same menu]

    • The simplest utterance meanings can be built up from the meanings of their various syntactic constituents, down to the level of words (Principle of Compositionality (Frege)). In this view, individual words have meanings which are stored in the lexicon, and these basic meanings are composed and built upon by the relations encoded in the syntactic constituents of the parse of the utterance until the full meaning is associated with the topmost "S"-constituent in the parse.
    • Such compositional semantics can either be done after parsing relative to the derivation-tree or during parsing in an integrated fashion (J&M, Section 18.5). The latter can be efficient for unambiguous grammars, but otherwise runs the risk of expending needless effort building semantic representations for syntactic structures considered during parsing that are not possible in a derivation-tree.
    • Though powerful, there are many shortcomings of this approach, e.g.,
      • Even in the absence of lexical and grammatical ambiguity, ambiguity in meaning can arise from ambiguous quantifier scoping, e.g., the two possible interpretations above of "every restaurant has a menu" (J&M, Section 18.3).
      • The meanings of certain phrases and utterances in natural language are not compositional, e.g., "roll of the dice", "grand slam", "I could eat a horse" (J&M, Section 18.6)

      Many of these shortcomings can be mitigated using more complex representations and representation-manipulation mechanisms; however, given the additional processing costs associated with these more complex schemes, the best ways of encoding and manipulating the semantics of utterances is still a very active area of research.

Monday, February 16 (Lecture #8) (FS)

• There are two types of multi-utterance discourses: monologues (a narrative on some topic presented by a single speaker) and dialogues (a narrative on some topic involving two or more speakers). As the NLP techniques used to handle these types of discourse are very different, we shall treat them separately.
• Applications: Language Understanding (Discourse / Monologue) (J&M, Chapter 21)
  • The individual utterances in a monologue are woven together by two types of mechanisms:
    • Coherence: This encompasses various types of relations between utterances showing how they contribute to a common topic, e.g. (J&M, Examples 21.4 and 21.5),
      • John hid Bill's car keys. He was drunk.
      • John hid Bill's car keys. He hates spinach.

      as well as use of sentence-forms that establish a common entity-focus, e.g. (J&M, Examples 21.6 and 21.7),

      • John went to his favorite musical store to buy a piano. He had frequented the store for many years. He was excited that he could finally buy a piano. He arrived just as the store was closing for the day.
      • John went to his favorite music store to buy a piano. It was a store that John had frequented for many years. He was excited that he could finally buy a piano. It was closing just as John arrived.

    • Coreference: This encompasses various means by which entities previously mentioned in the monologue are referenced again (often in shortened form), e.g., pronouns (see examples above).

  • Ideally, reconstructing the full meaning of a monologue requires correct recognition of all coherence and coreference relations in that monologue (creating a monologue parse graph, if you will). This is exceptionally difficult to do as the low-level indicators of coherence and coreference (for example, cue phrases (e.g., "because" "although") and pronouns, respectively) have multiple possible uses and are thus ambiguous, and resolving this ambiguity accurately can be either very computationally costly or even impossible.
  • Computational implementations of coherence and coreference recognition are often forced to rely on heuristics, e.g., using on recency of mention and basic number / gender agreement to resolve plurals (Hobb's Algorithm: J&M, Section 21.6.1).
  • Alternatively, one can weaken one's notion of what the meaning of a monologue is:
    • Summarize the overall meaning of a monologue as a vector of relevant words or phrases, where relevance is easily computable, e.g., words or phrases that are repeated frequently in the monologue (but not too frequently ("the")).
    • Summarize the most "important" meaning of a monologue in terms of the key entities in the monologue and their relations to each other, e.g., who did what to who in that news story?
    There are a very large number of narrow NLP techniques used to access these types of monologue meaning; they are the subject of study in the subdisciplines of Information Retrieval (Tzoukermann et al (2003); J&M, Section 23.1) and Information Extraction (BKL, Chapter 7; Grishman (2003); J&M, Chapter 22) respectively.
• Applications: Language Understanding (Discourse / Dialogue) (J&M, Chapters 23 and 24)
  • Even if one participant plays a more prominent role in terms of the amount speech or guiding the ongoing narrative, every dialogue is at heart a joint activity among the participants.
  • Characteristics of dialogue (J&M, Section 24.1):
    • Participants take turns speaking.
    • There may be one or more goals (initiatives) in the dialogue, each specific to one or more participants.
    • Each turn can be viewed as a speech act that furthers the goals of the dialogue in some fashion, e.g., assertion, directive, committal, expression, declaration (J&M, p. 816).
    • Participants base the dialogue on a common understanding, and part of the dialogue will be assertions or questions about or confirmations concerning that common understanding.
    • Both intention and information may be implicit in the utterances and must be inferred (conversational implicatures), e.g., "You can do that if you want" (but you shouldn't), "Sylvia does look good in that dress" (and why did you notice?).
  • The core of a Speech Dialogue System (SDS) is the Dialogue Manager (DM). An ideal DM should be able to accurately discern both explicit and implicit intentions and information in a dialogue, model and reason about both the common ground and the other participants in the dialogue, and take turns in the dialogue appropriately. This is exceptionally difficult to do, in large part because the domain knowledge and inference abilities required verge on those for full AI-style planning, which are known to be computationally very expensive (J&M, Section 24.7).

    

  • Implemented SDS deal with this by restricting their functionality and domain of interaction and retaining the primary initiative in the dialogue while allowing limited initiative on the part of human users, e.g., travel management and planning SDS (J&M, pp. 811-813).
    • The simplest single-initiative DM have finite-state implementations, e.g.,

      

      more complex mixed-initiative DM structure dialogues in terms of semantic frames whose slots can be filled in a user-directed order.

    • In addition to simplifying the DM, the restrictions above can also dramatically simplify the required language understanding and generation abilities to handle only expected interaction-types (J&M, p. 822).
  • Alternatively, SDS can focus on maintaining the illusion of dialogue while having the minimum (or even no) underlying mechanisms for modeling common ground or dialogue narrative:
    • Question Answering (QA) Systems (Harabagiou and Moldovan (2003); J&M, Chapter 23)
      • Such systems answer questions from users which require single-word or phrase answers (factoids).
      • Given the limited form of most questions, the topic of a question can typically be extracted from the question-utterance by very simple parsing (sometimes even by pattern matching).
      • Factoids can then be extracted from relevant texts, where relevance is assessed using techniques from Information Retrieval and Information Extraction.
    • Chatbots
      • Such systems engage in conversation with human beings, sometimes with some explicit purpose in mind (e.g., telephone call routing (Lloyds Banking Group, Royal Bank of Scotland, Renault Citroen)) but more often just as entertainment.
      • The first chatbots were ELIZA (Weizenbaum (1966): a simulation of a Rogerian psychotherapist) and PARRY (Colby (1981): a simulation of a paranoid personality).
        • ELIZA maintains no conversational state, instead relying on strategic pattern-matching of key phrases in user utterances and substitution of these phrases in randomly-selected utterance-frames. PARRY maintains a minimal internal state corresponding to the degree of the system's own "emotional agitation", which is then used to modify selected user utterance-phrases and their response-frames.
      • There are many chatbots today, e.g.,
        • Eliza
        • cleverbot

        (see also The Personality Forge, The Chatterbot Collection, and character.ai). Though many older chatbots rely on modifications of the mechanisms pioneered in ELIZA and PARRY, modern chatbots are often built using Seq2Seq modeling as implemented by neural network models (Vajjala et al (2020), Chapter 6).

      • Regardless of the simplicity of the mechanisms involved, chatbots can (with the co-operation of the human beings that they interact with) give a startling illusion of sentience (Epstein (2007); Tidy (2024); Weizenbaum (1967, 1979)). This is due in large part to innate human abilities to extract order from the environment, e.g., hearing voices or seeing faces in random audio and visual noise, which (in the context of interaction) assumes sentience and agency where none is present.

  • Current dialogue systems can give impressive illusions of understanding (especially recently-developed conversational AIs based on large language models such as ChatGPT and Bing's AI chatbot (Kelly (2023))) and do indeed have sentient ghosts within them -- however, for now, these ghosts come from us (Bender and Shah (2022); Greengard (2023); Videla (2023)). Hence, these systems are for all intents and purposes "glorified version[s] of the autocomplete feature on your smartphone" (Mike Wooldridge, quoted in Gorvett (2023)), conflating the most probable as the correct, and answers that are given to questions (in light of known problems of these systems wrt factual consistency and citing appropriate sources for statements made) should be treated with extreme caution (Du et al (2024); Dutta and Chakraborty (2023)).
• Applications: Automated Language Acquisition
  • Initial results in formal language learning and subsequent results in computational learning theory (see Clark and Lappin (2011) and references) suggest that it is computationally very difficult and in cases even provably impossible to acquire various types of languages. Those schemes that do work also require very large numbers of valid utterances and on occasion large pools of invalid examples (the latter being particularly useful in preventing the inference of overly general languages and/or grammars).
  • These results are at odds with the characteristics of human language acquisition (see Lecture # 5) which demonstrate that human beings acquire language efficiently with relatively few valid examples and no (or at least very few) invalid examples (Poverty of the Stimulus (POS))
  • This has lead to two schools of thought within the language acquisition community in linguistics:
    • Nativist (Internal-Driven): Human beings are born with a Universal Grammar (UG) such that, on minimal exposure to actual utterances from a particular language, that language can be efficiently acquired by inferring the appropriate settings of parameters in the UG.
    • Empiricist (Data-Driven): Human beings are born with powerful general-purpose inference abilities (e.g., probabilistic or analogical inference) which, in conjunction with restrictions in the structure of human languages (and, in some cases, minimal-grammar assumptions), allow the efficient acquisition of languages.

    The extreme versions of each school can be seen as the endpoints of a continuum, and it is generally acknowledged that the mechanisms underlying actual language acquisition probably lie somewhere in the middle.

  • These schools within linguistics have corresponding schools within the NLP language acquisition community, with the nativist school relying on hand-derived categories and grammars derived from linguistics and the empiricist school relying powerful statistical and machine learning algorithms (BKL, Chapters 6 and 9; J&M, Chapter 6; Manning and Schutze (1999)).
    • Though the pure Nativist NLP approach is tedious to implement, it has the advantage that morphosyntactic mechanisms incorporate semantic categories and thus interface seamlessly with semantic, discourse, and pragmatics processing systems.
    • The pure Empiricist NLP approach has been very successful in automatically inferring phonetic, phonological, morphological, and syntactic mechanisms. However, the basic supervised learning algorithms used to date require extremely large annotated databases of actual utterances which, while increasingly available (BKL, Chapter 11; J&M, Section 12.4), are difficult to produce. Moreover, as such databases tend to be confined to non-semantic domains, inferred categories and grammars may not be semantically meaningful and hence cannot easily interface with semantic, discourse, and pragmatics processing systems.
  • It seems likely that once the Nativist NLP community integrates insights derived within the Empiricist linguistic community (in particular, the exploitation of restrictions in the structure of human languages and minimal-grammar assumptions), further dramatic progress will be made, e.g., Yang and Piantadosi (2022). However, for the moment, NLP systems cannot automatically acquire the full range of human linguistic knowledge.
• Applications: Machine Translation (MT) (Hutchins (2003); J&M, Chapter 25; Somers (2003))
  • MT was among the first applications studied in NLP (J&M, p. 905). MT (and indeed, translation in general) is difficult because of various deep morphosyntactic and lexical differences among human languages, e.g., wall (English) vs. wand / mauer (German) (J&M, Section 25.1).
  • Three classical approaches to MT (J&M, Section 25.2):
    • Direct: A first pass on the given utterance directly translates words with the aid of a bilingual dictionary, followed by a second pass in which various simple word re-ordering rules are applied to create the translation.
    • Transfer: A syntactic parse of the given utterance is modified to create a partial parse for the translation, which is then (with the aid of a bilingual word and phrase dictionary) further modified to create the translation. More accurate translation is possible if the parse of the given utterance is augmented with basic semantic information.
    • Interlingua: The given utterance undergoes full morphosyntactic and semantic analysis to create a purely semantic form, which is then processed in reverse relative to the target language to create the translation.
    The relations between these three approaches are often summarized in the Vauquois Triangle:

    

  • The Direct and Transfer approaches require detailed (albeit shallow) processing mechanisms for each pair of source and target languages, and are hence best suited for one-one or one-many translation applications, e.g., maintaining Canadian government documents in both English and French, translating English technical manuals for world distribution. The Interlingua approach is best suited for many-many translation applications, e.g., inter-translating documents among all 25 member states of the European Union.
  • In part because of the lack of progress, the Automated Language Processing Advisory Committee (ALPAC) report recommended termination of MT funding in the 1960's. Research resumed in the late 1970's, re-invigorated in large part because of advances in semantics processing in AI as well as probabilistic techniques borrowed from automated speech recognition (J&M, p. 905).
  • Statistical MT uses inference over probabilistic translation and language models.
    • The translation models assess the probability of a given utterance being paired with a particular translation using the concept of word/phrase alignment.
    • Example #1: An alignment of an English utterance and its French translation:

      

    • To create the necessary probabilities (which are typically encoded in HMM), need large large databases of valid (often manually created) alignments.
  • Fully Automatic High-Quality Translation (FAQHT) is the ultimate goal. This currently achievable for translations relative to restricted domains, e.g., weather forecasts, basic conversational phrases. Current MT (which is largely based on the statistical and, most recently, neural network (Vajjala et al (w2020, pp. 265-268) approaches) is acceptable in larger domains if rough translations are acceptable, e.g., as quick-and-dirty first drafts for subsequent human editing (Computer-Aided Translation (CAT)). (J&M, pp. 861-862; see also BBC News (2014)) or simultaneous speech translation. It seems likely that further improvements will require a fusion of classical (in particular Interlingua) and statistical approaches.
• Potential exam questions

Wednesday, February 18

• University closed; class cancelled.

Friday, February 20

• In-class Exam

References

• BBC News (2014) "Translation tech helps firms talk business round the world." (URL: Retrieved November 14, 2014)
• Baker, J.M., Deng, L., Khuanpur, S., Lee, C.-H., Glass, J., Morgan, N., and O'Shaughnessy, D. (2009b) "Research Developments and Directions in Speech Recognition and Understanding, Part 2." IEEE Signal Processing Magazine, 26(6), 78-85.
• Beesley, K.R. and and Karttunen, L. (2000) "Finite-state Non-concatenative Morphotactics." In SIGPHON 2000. 1-12. [PDF]
• Beesley, K.R. and and Karttunen, L. (2003) Finite-State Morphology. CSLI Publications.
• Bender, E. and Shah, C. (2022) "All-knowing machines are a fantasy." IAI News. (Text)
• Bird, S. (2003) "Phonology." In R. Mitkov (ed.) (2003), pp. 3-24.
• Bird, S., Klein, E., and Loper, E. (2009) Natural Language Processing with Python. O'Reilly Media. [Abbreviated above as BKL]
• Carpenter, B. (2003) "Complexity." In R. Mitkov (ed.) (2003), pp. 178-200.
• Colby, K.M. (1981) "Modeling a paranoid mind." Behavioral and Brain Sciences, 4(4), 515-534.
• Dutoit, T. and Stylianou, Y. (2003) "Text-to-Speech Synthesis." In R. Mitkov (ed.) (2003), pp. 323-338.
• Du, M., He, F., Zou, N., Tao, D., and Hu, X. (2024) "Shortcut Learning of Large Language Models in Natural Language Understanding." Communications of the ACM, 67(1), 110-19. [Text]
• Dutta, SW. and Chakraborty, T. (2023) "Thus Spake ChatGPT: On the Reliability of AI-based Chatbots for Science Communication." Communications of the ACM, 66(12), 16-19. [Text]
• Edwards, C. (2021) "The best of NLP." Communications of the ACM, 64(4), 9-11. [Text]
• Epstein, R. (2007) From Russia With Love: How I got fooled (and somewhat humiliated) by a computer." Scientific American Mind, October, 6-17.
• Gorevtt, Z. (2023) "The AI emotions dreamed up by ChatGPT." BBC Future. Accessed February 28, 2023. [Text]
• Greengard, S. (2023) "Computational Linguistics Finds Its Voice." Communications of the ACM, 66(2), 18-20. [Text]
• Grishman, R. (2003) "Information Extraction." In R. Mitkov (ed.) (2003), pp. 545-559
• Haigh, T. (2023a) "Conjoined Twins: Artificial Intelligence and the Invention of Computer Science." Communications of the ACM, 66(6), 33-37. [HTML]
• Haigh, T. (2023b) "There was No 'First' AI Winter." Communications of the ACM, 66(12), 35-39. [HTML]
• Haigh, T. (2024a) "How the AI Boom Went Bust." Communications of the ACM, 67(2), 22-26. [HTML]
• Haigh, T. (2024b) "Between the Booms: AI in Winter." Communications of the ACM, 67(11), 19-23. [HTML]
• Haigh, T. (2025) "Artificial Intelligence: Then and Now." Communications of the ACM, 68(2), 24-29. [HTML]
• Harabagiou, S. and Moldovan, D. (2003) "Question Answering." In R. Mitkov (ed.) (2003), pp. 560-582
• Hutchins, J. (2003) "Machine Translation: General Overview." In R. Mitkov (ed.) (2003), pp. 501-511.
• Jurafsky, D. and Martin, J.H. (2008) Speech and Natural Language Processing (2nd Edition). Prentice-Hall. [Abbreviated above as J&M]
• Jurafsky, D. and Martin, J.H. (2022) Speech and Natural Language Processing (3rd Edition). (Book Website) [Abbreviated above as J&M2]
• Kaplan, R. (2003) "Syntax." In R. Mitkov (ed.) (2003), pp. 70-90.
• Kay, M. (2003) "Introduction." In R. Mitkov (ed.) (2003), pp. xvii-xx.
• Kedia, A, and Rasu, M. (2020) Hands-On Python Natural Language Processing. Packt Publishing; Birmingham, UK.
• Kelly, S.M. (2023) "The dark side of Bing's new AI chatbot." CNN Business. Accessed February 17, 2023. [Text]
• Kenstowicz, M.J. (1994) Phonology in Generative Grammar. Basil Blackwell.
• Kiraz, G.A. (2000) "Multi-tiered nonlinear morphology using multitape finite automata: A case study on Syriac and Arabic." Computational Linguistics, 26(1), 77-105. [PDF]
• Lamel, L. and Gauvin, J.-L. R. (2003) "Speech Recognition." In R. Mitkov (ed.) (2003), pp. 305-322.
• Lappin, S. (2003) "Semantics." In R. Mitkov (ed.) (2003), pp. 91-111.
• Leech, G. and Weisser, M. (2003) "Pragmatics and Dialogue." In R. Mitkov (ed.) (2003), pp. 136-156.
• Li, H. (2022) "Language models: past, present, and future." Communications of the ACM, 65(7), 56-63. (HTML)
• Lovins, J.B. (1973) Loanwords and the Phonological Structure of Japanese. PhD thesis, University of Chicago.
• Marcus, G. and Davis, E. (2019) Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon Books; New York.
• Marcus, G. and Davis, E. (2021) "Insights for AI from the Human Mind", Communications of the ACM, 64(1), 38-41. [Text]
• Martin-Vide, C. (2003) "Formal Grammars and Languages." In R. Mitkov (ed.) (2003), pp. 157-177.
• Mitkov, R. (ed.) (2003) The Oxford Handbook of Computational Linguistics. Oxford University Press.
• Mitkov, R. (2003a) "Anaphora Resolution." In R. Mitkov (ed.) (2003), pp. 266-283.
• Mohri, M. (1997) "Finite-state transducers in language and speech processing." Computational Linguistics, 23(2), 269-311. [PDF]
• Nederhof, M.-J. (1996) "Introduction to Finite-State Techniques." Lecture notes. [PDF]
• Ramsay, A. (2003) "Discourse." In R. Mitkov (ed.) (2003), pp. 112-135.
• Reis, E. S. D., Costa, C. A. D., Silveira, D. E. D., Bavaresco, R. S., Righi, R. D. R., Barbosa, J. L. V., and Federizzi, G. (2021) "Transformers aftermath: current research and rising trends." Communications of the ACM, 64(4), 154-163. [Text]
• Roche, E. and Schabes, Y. (eds.) (1997) Finite-state Natural Language Processing. The MIT Press.
• Roche, E. and Schabes, Y. (1997a) "Introduction." In E. Roche and Y. Schabes (Eds.) (1997), pp. 1-66.
• Schmidhuber, J. (2015) "Deep learning in neural networks: An overview." Neural Networks, 61, 85-117.
• Somers, H. (2003) "Machine Translation: Latest Developments." In R. Mitkov (ed.) (2003), pp. 512-528.
• Sproat, R. (1992) Morphology and Computation. The MIT Press.
• Tidy, J. (2024) "Character.ai: Young people turning to AI therapist bots." BBC Technology. Accessed January 8, 2024. [Text]
• Trost, H. (2003) "Morphology." In R. Mitkov (ed.) (2003), pp. 25-47.
• Tzoukermann, E., Klavans, J.L., and Strazalkowski, T. (2003) "Information Retrieval." In R. Mitkov (ed.) (2003), pp. 529-544.
• Vajjala, S., Majumder, B., Gupta, A., and Surana, H. (2019) Practical Natural Language Processing: A Comprehensive Guide to Building Real-world NLP Systems. O'Reilly; Boston, MA.
• Videla, A. (2023) "Echoes of Intelliogence: Textual Interpretation and Large Language Models." Communications of the ACM, 66(11), 38-43. [Text]
• Weizenbaum, J. (1966) "ELIZA - A computer program for the study of natural language communication between man and machine." Communications of the ACM, 9(1), 36-45.
• Weizenbaum, J. (1967) "Contextual understanding by computers." Communications of the ACM, 10(8), 474-480.
• Weizenbaum, J. (1979) Computer power and human reason. W.H. Freeman.

Created: January 26, 2026
Last Modified: February 18, 2026