Russell and Norvig, Chapter 22: Agents That Communicate

22.1 Communication as Action
- speech acts (utterances), not necessarily auditory
  - inform "There is a breeze in location 2,3."
  - query "Is there a pit in location 2,4?"
  - answer "I do not know if there is a pit in location 2,4."
  - request or command "I suggest you go to location 2,4."
  - promise or offer "I will go to location 2,4."
  - acknowledge "Okay"
  - share "Breezes make me scared."
- agent must know
  - when to produce a speech act
  - which speech act to produce
  - what an external speech act means (syntax and semantics)
  - why that external speech act was produced (models and beliefs)
- language
  - formal language (e.g., Prolog)
  - natural language (e.g., English)
  - strings ("the wumpus is dead") are made up of terminal symbols (the,
    wumpus, is, dead) 
  - phrase structure: strings are composed of phrases
    - noun phrase, verb phrase, sentence
    - i.e., nonterminal symbols
  - rewrite rules: S --> NP VP
- communication: speaker S wants to convey proposition P to hearer H using
                 words W
  - [Fig22.1, p657]
  - speaker
    - intention: S wants H to believe P
    - generation: S chooses words W
    - synthesis: S utters words W
  - hearer
    - perception: H perceives W'
    - analysis: H infers W' has meanings P1,...,Pn
    - disambiguation: H infers S intended to convey Pi
    - incorporation: H decides to believe Pi (or not)

22.2 Types of Communicating Agents
- telepathic agents [Fig22.2, p660]
  - agents share same internal representation
  - any agent can use: tell(KB_A,P), ask(KB_A,Q)
    - where A is any agent
  - reconciling different symbols used by different agents
  - reconciling different knowledge bases when bandwidth is limited
- communicating agents [Fig22.3, p662]
  - different internal representations
  - common external communication language
  - communication is an action
- e.g., wumpus world slave agent [Fig22.4, p663]
  - accepts commands and produces descriptions

22.3 A Formal Grammar for a Subset of English
- formal language does not handle all issues in natural language
  - e.g., changes in a language over time and location
- but, formal languages is a good place to start
- language E0 (subset of English)
  - lexicon (words) [Fig22.5, p665]
    - open classes: nouns, verbs, adjectives, adverbs
    - closed classes: pronoun, article, preposition, conjunctions
  - grammar [Fig22.6, p665]

22.4 Syntactic Analysis (Parsing)
- simple bottom-up parser [Fig22.7, p666]
- example trace [Fig22.8, p666]

22.5 Definite Clause Grammar (DCG)
- so far, using Backus-Naur Form (BNF)
  - rewrite rules of the form: NonTerm -> {NonTerm|Term}*
  - context-free (no a^n b^n c^n)
  - nothing about semantics (meaning)
- logic grammar
  - S -> NP VP  becomes  NP(s1) & NP(s2) => S(append(s1,s2))
  - Noun -> stench | ...  becomes  (s="stench" V ...) => Noun(s)
- definite clause grammar (DCG)
  - logic grammar restricted to Horn clauses
- BNF -> DCG
  - X -> Y Z ...  becomes  Y(s1) & Z(s2) & ... => X(Append(s1,s2,...))
  - X -> word  becomes  X(["word"])
  - X -> Y | Z | ...  becomes  Y'(s) V Z'(s) V ... => X(s)
    - where Y' is the logic translation of Y
- extended BNF
  - nonterminals can be augmented with extra arguments
    - NP(sem)  becomes NP(sem,s)
  - variables can appear on the right-hand side (antecedents)
    - variables represent a single symbol in the input string
    - e.g., Double -> w w  becomes
            (s1 = [w] & s2 = [w]) => Double(Append(s1,s2)) 
  - arbitrary logical tests can appear on the right-hand side
    - enclosed in curly braces
  - example from arithmetic

    Digit(sem) ->               (s=[sem]) & sem>=0 & sem=<9
      sem {0 <= sem <= 9}         => Digit(sem,s)

    Number(sem) -> Digit(sem)   Digit(sem,s) => Number(sem,s)
    Number(sem) ->
      Number(sem1) Digit(sem2)  Number(sem1,s1) & Digit(sem2,s2)
      {sem = 10 * sem1 + sem2}    & sem is 10 * sem1 + sem2
                                  => Number(sem,Append(s1,s2))

22.6 Augmenting a Grammar
- E0 can generate grammatically incorrect sentences
  - e.g., "Me smell a wumpus"
  - need a subjective pronoun (like "I"), not objective ("me")
  - could replace NP rule with NP_subj and NP_obj [Fig22.9, p669]
    - but number of rules grows as we duplicate rules
      - e.g., subject-verb agreement would result in 4 NP rules
- use augmentations (grammar E1) [Fig22.10, p670]
  - NP(case) -> Pronoun(case) | ...
    Pronoun(Subjective) -> I | ...
    Pronoun(Objective) -> me | ...
  - subject-verb agreement can also be handled this way (Ch23)
- verb subcategorization
  - E1 allows "give me the gold", but also "go me the gold"
  - need to specify what categories can follow a verb
    - verb subcategorization list of complements
    - e.g., "give" followed by [NP,NP] or [NP,PP]
            "go" followed by [PP]
  - can be enforced by augmenting grammar
    - VP(subcat) -> VP([NP|subcat]) NP(Objective) |
                    VP([Adjective|subcat]) Adjective |
                    VP([PP|subcat]) PP |
                    Verb(subcat)
    - S -> NP(Subjective) VP([])
    - Verb([NP,NP]) -> give | ...
    - example parse tree in [Fig22.12, p672]
- generation from augmented grammars
  - depends on number of values augmentations can take on
  - can be infinite (e.g., a^n b^n c^n)

22.7 Semantic Interpretation
- compositional semantics
  - meaning of phrase determined only by meaning of its subphrases
- attaching logical (lambda) expression to a verb
  - e.g., Verb(lambda(x) lambda(y) Loves(x,y)) -> loves
  - see [Fig22.15-16, p675]
- intermediate (quasi-logical) form
  - compositional form of logic
  - avoids problems of sharing arguments across phrases
    - e.g., "the wumpus is dead"
      - logic: exists(w) wumpus(w) & ...  AND  dead(...)
      - quasi-logic: dead(exists(w) wumpus(w))
    - add quantified terms
      - e.g., exists(w) wumpus(w) is a qualified term above
- constructing grammar semantics [Fig22.17, p678]
  - (1) Decide on logical or quasi-logical forms to generate.
        - sample sentences and their logical forms
  - (2) Study one-word modifications (e.g., "every" to "a")
  - (3) Write down logical types of each category (e.g., noun, verb, ...)
  - (4) Study one-phrase modifications
        - e.g., "ate the agent" to "shot the wumpus"
  - (5) Attach semantic augmentations to grammar rules
        - simple: NP(sem) -> Pronoun(sem)
        - function application: S(rel(obj)) -> NP(obj) VP(rel)
        - concatenation: NP([sem1,sem2]) -> Digit(sem1) Digit(sem2)
        - complex: VP(L(x) rel1(x) & rel2(EventVar(rel1))) ->
                     VP(rel1) Adverb(rel2)
  - example parse [Fig22.18, p679]
- grammar E2 [Fig22.19, p679]
- converting quasi-logic to logic
  - handle quantified terms Q(x) P(x), Q = exists or forall
  - for quantified term Q(x) P(x) in quasi-logic form QLF
    - replace Q(x) P(x) with x
    - replace QLF with Q(x) P(x) op QLF
      - op is => when Q is forall
      - op is &  when Q is exists
  - e.g., QLF dead(exists(w) wumpus(w))
          logic: exists(w) wumpus(w) & dead(w)
  - ambiguity when more than one quantified term
    - extract left to right usually works
- pragmatic interpretation
  - indexicals: e.g., "I" in "I saw the wumpus"
  - anaphora: e.g., "it" in "the wumpus is in 3 3" "it is dead"
  - resolved by disambiguation

22.8 Ambiguity and Disambiguation
- types of ambiguity
  - lexical ambiguity
    - words with more than one meaning or category
  - syntactic ambiguity leads to semantic ambiguity
    - "I smelled the wumpus in 2,2."
  - referential ambiguity: finding pronoun referents
  - pragmatic ambiguity
    - speaker and hearer disagree on context
  - local ambiguity
    - local meaning depends on external context
- disambiguation
  - derive possible interpretations as hypotheses
  - use models and uncertainty reasoning to choose
    - world model: facts and their probabilities
    - mental model: other agents and their beliefs
    - language model: P(string | fact to be communicated)
    - acoustic model: P(sounds | string to be uttered)
  - probabilistic context-free grammar (PCFG)
    - each rule has a probability
      - probabilities on all rules for one NonTerm sum to 1
    - context-free
      - P(pit) independent of its context "walk into the pit"

22.9 A Communicating Agent
- wumpus world slave agent [Fig22.4, p663]
- grammar E3 = E2 + commands + acknowledgements
  - commands (e.g., "shoot north")
    - implied noun object "Hearer"
  - acknowledgements (e.g., "ok")
- command grammar rules
  - S(Command(rel(Hearer)) -> VP(rel)
    S(Statement(rel(obj)) -> NP(obj) VP(rel)
  - e.g., "shoot north" becomes Command(shoot(Hearer,north))
- acknowledgement grammar rules
  - S(Acknowledge(sem)) -> Ack(sem)
    Ack(True) -> "yes"
    Ack(True) -> "ok"
    Ack(False) -> "no"