Ontological Semantics: An Overview Sergei Nirenburg University of Maryland, Baltimore County

1. Ontological Semantics: An Overview Sergei Nirenburg University of Maryland, Baltimore County

2. Ontological Semantics is • a knowledge-based linguistic theory, and • a knowledge-based NLP method of extracting, representing and manipulating meaning in natural language texts. • However, ontological-semantic applications are hybrid systems that use a variety of methods to ensure adequate coverage of language material and to enhance system capabilities.

3. Ontological-semantic analyzers take natural language texts as inputs and generate machine-tractable text meaning representations (TMRs) that form the basis of various reasoning processes. Ontological-semantic text generators take TMRs as inputs and produce natural language texts.

4. Ontological-semantic systems centrally rely on • extensive static knowledge resources: • a language-independent ontology, the model of the world that includes models of intelligent agents; • ontology-oriented lexicons (and onomasticons, or lexicons of proper names) for each natural language in the system; and • a fact repository containing instances of ontological concepts as well as remembered text meaning representations.

5. Applications of ontological semantics include • knowledge-based machine translation, • information retrieval and extraction, • text summarization, • question answering • ontological support for reasoning systems, • knowledge management and others.

6. Major processing components in ontological semantics include Tokenization Morphological Analysis Syntactic Analysis Semantic Analysis (lexical and compositional) Pragmatics and Discourse Analysis Knowledge-Based Reasoning Text generation

7. Text Meaning Representation Ontological Semantics (applied to MT) Input Text L1 Output Text L2 Analyzer Generator Ecology Morphology Syntax Lexicon Ontology Data flow Background knowledge

8. Data flow Background knowledge Extended TMR Basic TMR Connections among staticknowledgesources Lexicons Lexicons Onomasticons Ontology Fact Repository Language-Independent Static Knowledge Sources Input: Text, Query,etc. Output: Text, Filled Template, Query Answer, etc. InputTextAnalyzer Output Generator: IE, QA, MT Engines, etc. Application-Oriented InferenceEngine Language-Dependent Static Knowledge Sources Non-SemanticKnowledge: Ecology Morphology, Syntax, etc.

9. Ontological Semantics is an approach todescriptive computational linguistics. • This means that it: • is devoted to processing naturally occurring texts, • strives for broad coverage, and • expects “unexpected” inputs.

10. Several applications of ontological semantics have been implemented in the projects at CMU, NMSU and UMBC: Mikrokosmos Savona MINDS CREST MOQA and some others

11. Static Knowledge Sources: Ontology of about 6,000 concepts (about 85,000 facts) English lexicon of about 45,000 entries Spanish lexicon of about 40,000 entries Chinese lexicon of about 3,000 entries Fact repository of about 20,000 facts Ecological knowledge (including large onomasticons), morphological and syntactic grammars for many languages

12. Lexicon Ontology Fact Repository Onomasticon Object Instances Event Instances

13. Processing modules include: Preprocessors Syntactic analyzers Semantic analyzer Text generator Control architectures Knowledge acquisition tools (too many!)

14. Main Past and Current Collaborators: Steve Beale Ralf Brown Lynn Carlson Spence Koehler Kavi Mahesh Marge McShane Antonio Moreno Ortiz Boyan Onyshkevych Victor Raskin Evelyne Viegas

15. This is a broad (and shallow) overview of the ontological- semantic processing and static resources.

16. Input Text Text MeaningRepresentation (TMR) SyntacticAnalyzer SemanticAnalyzer Preprocessor Grammar: Ecology MorphologySyntax Lexicon and Onomasticon Ontology and Fact Repository Static Knowledge Resources

17. Text Meaning Representation (a fragment) proposition _1 • head %travel-1 • agent human-544“Mr. Smith” • source location-213 “London” • destination location-665 “Ankara” • tmr-time • time-end 20000702“July 2, 2000” • aspect • iteration single • phase end “arrived”

18. Deriving basic semantic dependency Input: John makes tools Syntactic Analysis: root   make cat verb tense present subject   root john cat noun-proper object   root     tool cat noun number plural

19. Relevant parts of the lexical entry for make make-v1 syn-struc root make cat v subj root $var1 cat n object root $var2 cat n sem-struc manufacturing-activity agent ^$var1 theme ^$var2

20. Relevant Extract from Ontological Concept Specification manufacturing-activity... agent human theme artifact …

21. Relevant parts of the lexicon entries for John and tool John-n1 syn-struc root john cat noun-proper sem-struc human name john gender male tool-n1 syn-struc root tool cat n sem-struc tool

22. The lexicon entry for make establishes that the meaning of the syntactic subject of make is the main candidate to fill the agent slot in manufacturing-activity, while the meaning of the syntactic object of make is the main candidate to fill the theme slot. The lexicon entry for make refers to the ontological concept manufacturing-activity without modifying any of its constraints in the lexicon entry. This states that the meaning of its subject should be constrained to any concept in the ontological subtree with the root at the concept human; and the meaning of its object, to an element of the ontological subtree rooted at artifact. These constraints are selectional restrictions, and the lexicon entries for John and tool satisfy them.

23. The basic semantic dependency component of the TMR forJohn makes tools is as follows: … manufacturing-activity-7 agent human-3 theme set-1 element tool cardinality > 1 …

24. Other levels of basic semantic dependency John makes tools quickly John makes expensive tools John makes very expensive tools John makes power tools.

25. expensive-adj1 cat adj syn-struc root $var1 cat n mods root big sem-struc ^$var1 sem physical-object cost > 0.75

26. As meanings of nouns typically do not directly correspond to properties, processing John makes power tools is more convoluted. The problem of nominal compounding is quite confounding in English: the IBM lecturein The IBM lecture will take place tomorrow at noon can mean a lecture given by IBM employees, a lecture sponsored by IBM, a lecture about IBM, a lecture given at IBM as well as many other things, signifying the necessity of connection on different properties. We have an instance of compositional ambiguity. We’ll see examples of word-sense ambiguity later on.

27. Input Text Residual Ambiguity > 1 TMR SyntacticAnalyzer SemanticAnalyzer Preprocessor Grammar: Ecology MorphologySyntax Lexicon and Onomasticon Ontology and Fact Repository Static Knowledge Resources

28. Input Text Ambiguity and incongruity > 1 TMR 1 TMR < 1 TMR SyntacticAnalyzer SemanticAnalyzer Preprocessor Grammar: Ecology MorphologySyntax Lexicon and Onomasticon Ontology and Fact Repository Static Knowledge Resources

29. Input Text < 1 f-structure SyntacticAnalyzer 1 f-structure Preprocessor > 1 f-structure Grammar: Ecology MorphologySyntax Lexicon and Onomasticon Static Knowledge Resources

30. Input Text “Unexpected” input: Recovery and repair < 1 TMR SyntacticAnalyzer SemanticAnalyzer Preprocessor Grammar: Ecology MorphologySyntax Lexicon and Onomasticon Ontology and Fact Repository Static Knowledge Resources

31. If the above lexical entry for make is used, The gorilla makestools will violate selectional constraints; this is because gorillas are not humans and, according to the lexical and ontological definition above are unsuitable as fillers for the agent slot of manufacturing-activity. We have a case of sortal incongruity. Or a case of metonymy (if the gorilla actually refers to a person). (The system does not know this but will have to attempt repair in any case).

32. One way of dealing with repair or ambiguity is to modify the static knowledge sources.

33. The relevant part of the ontological concept manufacturing-activity should become as illustrated as follows: manufacturing-activity... agent sem human relaxable-to primate... The lexical entry for make remains unchanged—no matter that it actually means a slightly different thing now.

34. relation-slot ::= relation-name facet concept-name+ attribute-slot ::= attribute-name facet {number | literal}+ facet ::= value | sem | default | relaxable-to | … Skip facets

35. value the filler of this facet is an actual value; it may be the instance of a concept, a literal symbol, a number, or another concept. Most of the constraints in TMRs are realized as fillers of the value facet. In the ontology, in addition to ontology slots, the value facet is used to carry factual truths, e.g., that Earth has exactly one moon: earth ... number-of-moons value 1 ...

36. Sem Constraints realized through the sem facet are abductive, that is, it is expected that they might be violated in certain cases. pay definition value “to compensate somebody for goods or services rendered” agent sem human theme sem commodity patient sem human Indeed, the agent or patient of paying may be not a human but, for example, an organization; the theme of paying may be an event, as in John repaid Bill’s hospitality by giving a lecture in his class. It is important to recognize that the filler of theme cannot be “relaxed” indefinitely. To mark the boundaries of abductive relaxation, the relaxable-to facet is used (see below).

37. default the filler of a default facet is the most frequent or expected constraint for a particular property in a given concept. This filler is always a subset of the filler of the sem facet. In many cases, no default filler can be determined for a property. pay, however, does have a clear default filler for its theme property: pay definition value “to compensate somebody for goods or services rendered” agent sem human theme default money sem commodity patient sem human

38. relaxable-to this facet indicates to what extent the ontology permits violations of the selectional constraints listed in the sem facet, e.g., in nonliteral usage such as a metonymy. The filler of this facet is a concept that indicates the maximal set of possible fillers beyond which the text should be considered anomalous. pay definition value “to compensate somebody for goods or services rendered” agent sem human relaxable-to organization theme default money sem commodity relaxable-to event patient sem human relaxable-to organization

39. The default, sem and relaxable-to facets are used in the procedure for matching what amounts to multivalued selectional restrictions. In cases when multiple facets are specified for a property, the program first attempts to perform the match on the selectional restrictions in default facet fillers, where available. If it fails to find a match, then the restrictions in sem facets are used and, failing that, those in relaxable-to facets.

40. The idea of multivalued selectional restrictions can be extended further, to include static probabilistic values or even probabilistic values relative to a corpus. One difficulty in acquiring such knowledge is having to count probabilities of ontological concepts and not word or phrase occurrences.

41. Selectional restrictions are not enough • individual constraints between the head of a proposition and each of its arguments typically available in static knowledge sources (lexicons) are often not strong enough or too strong for effective selection of word senses; • the real power of word sense selection lies in the ability to tighten or relax the semantic constraints on senses of a lexeme on the basis of choices made by a semantic analyzer for other words in the dynamic context. • effective sense disambiguation should be helped by the availability of rich knowledge with a high degree of cross-dependence among knowledge elements (the ontology and the fact repository)

42. Proposition Head Case Role i Case Role j WS1 WS2 WS3

43. Proposition Head WS Case Role i Case Role j WS1 WS2

44. Proposition Head WS Case Role i Case Role j WS1 WS2 skip dynamic sel. restr.

45. Dynamic tightening of selectional restrictions helps to resolve residual ambiguities. We illustrate this using the results of an experiment run in the framework of the Mikrokosmos implementation of ontological semantics (see Mahesh et al. 1997a for the original report).

46. Let us consider the sentence John prepared a cake with the range. (and disregard the PP-attachment ambiguity in it). In this sentence, several words are ambiguous, relative to the static knowledge sources. The lexical entry for prepare contains four senses, one related, respectively, to the ontological concepts prepare-food, planning-event, make-ready and training. The lexical entry for range has a number of different senses, referring to a mathematical range, a mountain range, a shooting range, a livestock grazing range as well as to a cooking device. There are two cooking senses, related to the ontological concept oven and the ontological concept stove, respectively. The lexical entry for cake is unambiguous: it has the ontological concept cake as the basis of its meaning specification.

47. The entry for John is found in the onomasticon and unambiguously recognized as a man’s name. The entry for with establishes the type of relation on which the appropriate sense of range is connected to the appropriate sense of prepare. The possibilities include agent (e.g., Bill wrote the paper with Jim), and instrument (e.g., Bill opened the door with a key). The meanings of a and the will have “fired” in the process of syntactic analysis.

48. The meaning of prepare will become the head of the proposition describing the meaning of this sentence. The selectional restriction on the direct object of prepare in the sense of prepare-food matches the candidate constraint provided by the meaning of cake while the selectional restrictions on prepare in the other senses either do not or match less precisely. This disambiguates prepare using only static selectional restrictions. The meaning of John, in fact, matches any of the senses of prepare. So, while this word does not contribute to disambiguation of prepare, at least it does not hinder it.

49. Next, we establish that the correct sense of with is the one related to instrument rather than agent because none of the senses of range are related to concepts that are descendants of human, which is a requirement for being agent of prepare-food. At this point, we can exclude all those senses of range that are not compatible with the remaining sense of with, namely all but the food-related ones, whose meanings are related to stove and oven. Static selectional restrictions already disambiguated everything but the remaining two senses of range. No static selectional restrictions are available in the lexicon to help us complete the disambiguation process. We are now at the main point of our example, namely, a demonstration of the utility of dynamic selectional restrictions.

50. The last remaining task for disambiguation is to choose either oven or stove (signaled in the input by the corresponding word senses of range) as the theme of the proposition head prepare-food-i. Without context, this determination is not possible because range can still mean either oven or stove.