Lexical Semantics for the Semantic Web

1
Lexical Semantics for the Semantic Web

• Patrick Hanks
• Masaryk University, Brno
• Czech Republic
• hanks_at_fi,muni.cz
• UFAL, Mathematics Faculty, Charles University in
  Prague

2
Outline of the talk

• A neglected aspect of Tim Berners-Lees vision
• Introducing semantics to the semantic web
• Computing meaning and inferences in free text
• Patterns in text and how to use them
• Building a resource that encodes patterns
• linking meanings (implicatures) to patterns (not
  to words)
• A pattern dictionary
• What does the pattern dictionary look like?
• The role of an ontology in a pattern dictionary

3
Aims of the Semantic Web

• To enable computers to manipulate data
  meaningfully
• Most of the Web's content today is designed for
  humans to read, not for computer programs to
  manipulate meaningfully.
• Berners-Lee et al., Scientific American, 2001

4
A neglected aspect of Berners-Lees vision

• Web technology must not discriminate between the
  scribbled draft and the polished performance.
• T. Berners-Lee et al.,
  Scientific American, 2001
• The vision includes being able to process the
  meaning and implicatures of free text
• not just pre-processed tagged texts Wikis,
  names, addresses, appointments, and suchlike.

5
A paradox

• Traditional KR systems typically have been
  centralized, requiring everyone to share exactly
  the same definition of common concepts such as
  'parent' or 'vehicle'.
• Berners-Lee et al. 2001.
• Implying that SW is more tolerant?
• Apparently not
• Human languages thrive when using the same term
  to mean somewhat different things, but automation
  does not. --Ibid.

6
The root of the problem

• Scientists from Leibniz to the present have
  wanted word meaning to be precise and certain.
• But it isnt. Meaning in natural language is
  vague and probabilistic
• Some theoretical linguists (and CL researchers),
  not liking fuzziness in data, have preferred to
  disregard data in order to preserve theory
• Do not allow SW research to fall into this trap
• To fulfil Berners-Lees dream, we need to be able
  to compute the meaning of un-pre-processed
  documents

7
What NOT to do for the SW

• The meaning of the English noun second is vague
  a short unit of time or 1/60 of a minute.
• Wait a second.
• He looked at her for a second.
• It is also a very precisely defined technical
  term in certain scientific contexts the basic
  SI unit of time
• the duration of 9,192,631,770 cycles of
  radiation corresponding to the transition between
  two hyperfine levels of the ground state of an
  atom of caesium 133.
• If we try to stipulate a precise meaning for all
  terms in advance of using them, well never be
  able to fulfil the dream and we will invent an
  unusable language

8
Precision and vagueness

• Stipulating a precise definition for an ordinary
  word such as second removes it from ordinary
  language.
• When it is given a precise, stipulative
  definition, an ordinary word becomes a technical
  term.
• An adequate definition of a vague concept must
  aim not at precision but at vagueness it must
  aim at precisely that level of vagueness which
  characterizes the concept itself.
• Wierzbicka 1985, pp.12-13

9
The paradox of natural language

• Word meaning may be vague and fuzzy, but people
  use words to make very precise statements
• This can be done because text meaning is
  holistic, e.g.
• fire in isolation is very ambiguous
• But He fired the bullet that was recovered from
  the girl's body is not at all ambiguous
• Ithaca is ambiguous
• But Ithaca, NY is much less ambiguous.
• Even the tiniest bit of (relevant) context helps.

10
What is to be done?

• Process only the (strictly defined) mark-up of
  documents, not their linguistic content?
• And so abandon the dream of enabling computers to
  manipulate linguistic content?
• Force humans to conform to formal requirements
  when writing documents?
• Not a serious practical possibility
• Teach computers to deal with natural language in
  all its fearful fuzziness?
• Maybe this is what we need to do

11
Hypertext and relevance

• The power of hypertext is that anything can link
  to anything.
• Berners-Lee et al., 2001
• Yes, but we need procedures for determining
  (automatically) what counts as a relevant link,
  e.g.
• Firing a person is relevant to employment law.
• Firing a gun is relevant to warfare and armed
  robbery.

12
How do we know who is doing what to whom?

• Through context (a standard, uncontroversial
  answer)
• But teasing out relevant context is tricky
• Firing a person Person MUST be mentioned
• Whereas firing a gun occurs in patterns where
  neither Firearm nor Projectile are
  mentioned, e.g.
• The police fired into the crowd/over their
  heads/wide.
• Negative evidence can be important
• He fired cannot mean he dismissed someone from
  employment
• Relevant context is cumulative
• So correlations among arguments are often needed

13
How to compute meaning for the Semantic Web

• STEP 1. Identify all the normal patterns of
  normal utterances by data analysis
• STEP 2. Develop a resource that says precisely
  what the basic implicatures of each pattern are,
  e.g.
• Human fire AdvDirection
• Human causes Firearm to discharge
  Projectile
• STEP 3. Populate the semantic types in an
  ontology
• STEP 4. Develop a linguistic theory that
  distinguishes norms from exploitations
• Abandon the received theories of speculative
  linguists
• STEP 5. Develop procedures for finding best
  matches between a free text statement and a
  pattern.

14
The double helix of language norms and
exploitations

• A natural language consists of TWO kinds of
  rule-governed behaviour
• Using words normally
• Exploiting the norms
• We dont even know what the norms of any language
  are, still less the exploitation rules
• People have assumed that norms of usage are
  obvious
• But only some of the things that are obvious are
  true
• We need to identify the norms by painstaking
  empirical analysis of evidence
• There is not a sharp dividing line between norm
  and exploitation
• Todays norm is tomorrows exploitation

15
Corpus Pattern Analysis (CPA)

• Identifies normal usage patterns for each word
• Each pattern include a verb, its valencies, and
  the semantic type(s) of each argument (valence)
• Associates a meaning (implicature) with each
  pattern (NOT with each word)
• Provides a basis for matching occurrences of
  target words in unseen texts to their nearest
  pattern (norm)
• CPA is the basis for a Pattern Dictionary
  (demo)
• http//nlp.fi.muni.cz/projekty/cpa/
• Click on web access in line 1

16
Focusing arguments by semantic-type alternation

• You can calm a person, calm a horse, calm
  someones nerves, fears, or anxiety.
• These all activate the same meaning of the verb
  calm. Anxiety does not have the required
  semantic type (anxiety is not Animate)
• However, the expected animate argument is present
  but only as a possessive. And even if there is
  no possessive, being an attribute of Animate
  is part of the meaning of nerves, fear, anxiety,
  etc.
• Regular alternations such as these have a
  focusing function. They do not activate different
  senses.
• Other examples
• Repair a car, repair the engine (of a car),
  repair the damage
• Treat a person, treat her injuries, treat her
  injured arm

17
Ontologies

• The arguments of CPA patterns are expressed as
  semantic types, related to a shallow semantic
  ontology.
• The term ontology is has become highly
  ambiguous
• SW ontologies are, typically, interlinked
  networks of things like address lists, dates,
  events, and websites, with html mark-up showing
  attributes and values
• They differ from philosophical ontologies, which
  are theories about the nature of all the things
  in the universe that exist
• They also differ from lexical ontologies such as
  WordNet, which are networks of words with
  supposed conceptual relations
• The CPA shallow ontology is a device for grouping
  semantically similar words together to facilitate
  meaning processing

18
The CPA Shallow Ontology

• The CPA Shallow Ontology is a bag of bags of
  words
• Developed, bottom-up, by cluster analysis of
  corpora
• The nouns that NORMALLY occur in the same
  syntagmatic slot in relation to a given verb are
  grouped into a cluster
• A cluster of different nouns activate the same
  meaning of the verb
• The cluster is named with a semantic type, e.g.
  Human, Event, Abstract, Artefact,
  etc.
• Each cluster is compared with similar clusters
  occurring with other verbs. Each combination of
  clusters constitutes a lexical set.
• Identically named clusters contain slightly
  different members (lexical items)
• Therefore, lexical sets shimmer.

19
The Predictive Power of Lexical Sets

• EXAMPLE A noun, meeting has been classified
  with semantic type Event at both arrange and
  attend
• Suppose meeting is found in the direct object
  slot after leave or runbut not frequently enough
  to have been included in a cluster for those
  verbs in the Ontology
• However, the patterns Human leave Event
  and Human run Event will be found in
  the Pattern Dictionary
• Then there is a high probability that meeting
  belongs there (even though not listed as
  typical), activating probable implicatures
• leave "go away from
• run "organize and cause to function
  efficiently

20
Phraseology in Computational Linguistics

• Computational linguists are turning away from
  word-by-word analysis (the Lego bricks method,
  inherited from Frege) to phraseological analysis.
  E.g.
• Marine Carpuat and Dekai Wu. 2007. How phrase
  sense disambiguation outperforms word sense
  disambiguation for statistical machine
  translation. In Proceedings, Conference on
  Theoretical and Methodological Issues in Machine
  Translation (TMI 2007). Skovde, Sweden
• The Pattern Dictionary provides an inventory of
  patterns
• A benchmark for NLP researchers using patterns
• A benchmark for introducing semantics to the
  Semantic Web

21
The English Pattern Dictionary current status

• Focuses on verbs
• Specifically, the correlations among the lexical
  and semantic values of the arguments of each
  sense of each verb
• 700 verbs analysed so far
• 400 verbs complete, finalized, checked and
  released
• 300 more are work in progress, awaiting checking
• There are approximately 6000 verbs in English, so
  we have done about 10
• Shallow ontology in development
• New lexically driven theory of language, which is
  precise about the vague phenomenon of language
• Hanks (forthcoming) Analysing the Lexicon Norms
  and Exploitations. MIT Press

22
The English Pattern Dictionary the future

• 5,400 more verbs to analyse (then the adjectives)
• Develop a different procedure for nouns (noun-y
  nouns)
• Finalize the CPA shallow ontology and populate it
• Pattern dictionaries for other languages
• Czech
• German (A. Geyken, Berlin)
• Italian (E. Jezek, U. of Pavia)
• Theoretical work
• Typology of exploitations
• Implications of CPA for parsing theory
• Alternation of semantic types in arguments
• Relationship between semantic types and semantic
  roles
• Links between the Pattern Dictionary and FrameNet

23
Conclusions

• To enable computers to manipulate data
  meaningfully (the raw data itself, not just tags
  added to the data), we need
• an inventory of patterns of normal usage for each
  word
• a pattern dictionary
• a theory that distinguishes normal usage from
  exploitations of norms for rhetorical, poetic,
  and other purposes
• pattern-matching procedures text lt gt pattern
  dictionary
• a statistical, probabilistic approach to
  identifying meaning.
• Only then will computers be able to compute the
  meaning of texts, understand the implicatures,
  translate them, retrieve data from them, and
  manipulate them in other ways
• At that point, we shall be a little closer to
  realizing Berners-Lees 2001 dream