Lexical Semantics-Syntactic Model For Defining and Subcategorizing Attribute Noun Class
Lexical Semantics-Syntactic Model For Defining and Subcategorizing Attribute Noun Class
Abstract. Either word sense classification or word sense clustering can not avoid this
problem: what features people can get from word meaning that can be used as the tools
employed for producing word sense ontology. In this paper, we present a lexical
semantic-syntactic model to define and subcategorize a noun classattribute noun class,
by acquiring syntactic behavior and semantic restrictions from corpus. First we introduce
the idea of attribute noun class. Then we describe the model, which is a specific tool
providing a practical method to decide whether a word should be located in a specific
word sense class or not. Then we apply the model to 15 nouns to show how the model
works, and create a sample taxonomy for the 15 nouns.
The meaning of a word might impact the syntactic behavior of the word. For example, attributive
adjectives (also called non-predicative adjectives or as qu bie ci known in Chinese) in Chinese like
(snowy white) can not be syntactically used with adverb, so (very)/adv (snowy
white)/adj is not a legal phrase in Chinese. Because there is a adverbial morpheme (snowy) in
word (snowy white) which means snowy, the word merges the meaning of adverb into its
lexical meaning. Attribute noun can not be attributive in attributive-head relation and rarely be isolate
subject in subject-predicate relation which we will show in the following section.
The semantic selectional restrictions are semantic frame in which words should appear. The semantic
frame is the grammatical relation with semantic constrains. For example, there are two grammatical
slots in VO relation: verbal and objective. The meaning of object must semantically match with the
meaning of the verb and vice versa, in other word, either verb or object has semantic preferences to
the collocate words; such preferences is selectional restrictions in lexical semantics. The
selectional restrictions also exist in semantic role frame, but we do not discuss the issue in this paper.
Therefore, we have the syntactic behavior of ANC by studying its lexical meaning.
F1. The attribute noun can not be the isolate subject in SP.
F2. The attribute noun can not be the attributive in AH.
We will use F1 and F2 to determine whether a noun is an attribute noun or not in the section below.
Worth to mention that: 1) the rules of F1 and F2 is not black or white rule, the syntactic features
occur with very big probability; 2) when we test whether a noun is an attribute one, the F1 and F2
should be applied in sentence level, i.e. if a noun is an isolate subject or attributive, the SP or AH must
be semantically understandable as a sentence.
Define AN
{
syn_behvior;
sr_SP;
sr_VO;
sr_AH;
}
AttrDef(AN noun)
{
if(Be_Attr(noun, AH)==false||Be_IsoSub(noun, SP)==false)
Subcat_AN(noun);
else
printf(the noun is not an attribute noun);
}
Subcat_AN(noun)
{
while(sentence[i])
features=get_syn_features(noun, sentence[i]);
noun=SR(noun, features, AH)
store the result in AN;
noun=SR(noun, features, VO);
store the result in AN;
noun=SR(noun, features, SP);
store the result in AN;
}
end.
The attribute noun is presented in a structure AN, in which the syntactic behavior and
selectional restrictions are stored. The function AttrDef decides whether a word is an attribute
noun or not, using the syntactic features F1 and F2. Then the word is input in Subcat_AN
which assigns the word to a more specific sense class. In Subcat_AN, a noun is semantically
analyzed in SR function with 3 syntactic relations: AH, VO and SP. The result of syntactic
and semantic analysis will be presented in structure for each input.
4. Experiment
In this section, we make an experiment to evaluate the applicability of LSS model.
4.1 Methods
First we choose 15 nouns from Singapore Chinese Textbook Corpus. Then we abstract
sentences from the first page result that CCL online corpus generated for each noun and
acquire the data of syntactic features. The syntactic features are used to determine whether a
noun is attribute noun or not.
Second, we choose 15 attribute nouns from the same corpus and generate sentences for
each attribute noun from CCL online corpus, for abstracting semantic features. The semantic
features will be put in structure AN mentioned in section 3 to show how 15 nouns differ in
PACLIC 24 Proceedings 69
semantic selectional restrictions. Such semantic features help us to divide 15 attribute nouns
into several subcategories.
The features, either of syntactic or semantic, are manually abstracted from sample
sentences. We do not employ machine learning software (like Stanford Parser, SRL software)
to do the job, because the purpose of this paper is to create a model for attribute noun and
evaluate rather than evaluate machine learning algorithm. We want the data to verify the
model. In additional, the data set is quite small (15 words).
Table 1
Word Isolate sub in Head in AH Attributive in AH Object in VO
SP
/45 13 9 10 13
/50 0 0 50 0
/50 22 18 4 6
/50 2 23 23 2
/40 4 14 11 11
/43 6 18 2 17
/ 4 37 1 4
/49 5 33 1 10
/42 9 1 2 32
/51 14 14 13 10
/39 0 34 3 2
/47 0 45 0 2
/40 0 34 2 4
/42 0 41 0 1
/32 0 32 0 0
Table 1 shows the grammatical distribution of the 15 nouns. Five attribute nouns show
quiet distinguish grammatical distribution from other nouns: they can not be isolate subject in
SP and rarely be attributive in AH. This distinguishable grammatical distribution verifies that
F1 and F2 rules are valid.
Table 2
Attribute nouns Host words
(strength, physical power) (name of a man), (man), (blue
whale), (ant), (baby elephant)
(talent) (humans name), (student),
(farmer), (female colleague),
(Finland people)
(tone) (humans name), (minister of foreign
affairs), (prime ministry), (he),
(teacher)
(vitality) (microorganism), (plant),
(shrimp), (katydid), (Chinese redbud)
(instinct) (animal), (bird), (female leopard),
(human)
(voice) (humans name), (she),
(singer)
(family circumstances) (humans name), (student),
(humans name), (humans name)
(work position) (he), (government officers),
(people), (lieutenant)
(character) (teacher), (mother), (myself)
(memory) (animal), (human), (patient),
(old people), (bee)
(intention) (director), (I), (boss), (coach)
(look) (lackey), (politician),
(American), (enemy)
(garb) (women), (ladies), (she),
(guard), (general)
(diameter) (galaxy), (greenhouse),
(grenade), (pit)
(grace) (gentlemen), (scholar),
(humans name), (lady), (politician)
The hosts of attribute nouns (talent), (tone), (voice), (family
circumstances), (position), (character), (intention), (look),
(garb) and (grace) are all of meaning of human. The structure presentation of
these nouns is:
(talent)={syn_behavior[]; sr_AH[human]; sr_VO[]; sr_SP[]};
(tone) ={syn_behavior[]; sr_AH[human]; sr_VO[]; sr_SP[]};
(voice)={syn_behavior[]; sr_AH[human]; sr_VO[]; sr_SP[]};
PACLIC 24 Proceedings 71
Table 3
Attribute noun Verb
(strength, physical power) (have), (do not have), , (spend)
(talent) (have), (do not have),
(tone) (handle), (use)
(vitality) (show), (have), (maintain)
(instinct) (have), (become)
(voice) (show), (listen to)
(family circumstances) NULL
(work position) (occupy), (dismiss), (retire from)
(character) (highlight), (present)
(memory) (lost), (maintain)
(intention) (understand), (carry out), (know),
(conjecture)
(look) (describe), (explosure), (realize),
(reveal)
(garb) (wear), (look), (stare at), (see
clearly),
(diameter) NULL
(grace) (pay attention to), (show), (maintain)
72 Regular Papers
Table 4
Attribute noun Predicate
(strength, physical power) (very strong), (quite weak)
(talent) (pretty high)
(tone) (low), (gentle), (soft), (gentle)
(vitality) (vigorous), (strong)
(instinct) NULL
(voice) (loud and clear), (very charming),
(agreeable)
(family circumstances) (poor), (rich)
(work position) (pretty high), (be prompted)
(character) (optimistic), (good or bad)
(memory) (decline), (very good)
(intention) (is serious), (very clear)
(look) (completely reveal), (black)
(garb) (simple), (vulgar), (elegant)
(diameter) 0.55 (0.5-5 micrometers), 20 (less
than 20 km)
(grace) (elegant), (elegant)
PACLIC 24 Proceedings 73
Living creature
Human
Appearance
Social
Mental
Audible
Quality attribute
5. Conclusion
74 Regular Papers
In this paper, we present a practical method to define and subcategorize the attribute noun
class, and produce sample word sense taxonomy for 15 Chinese attribute nouns. Syntactic and
semantic features are employed to do the job, from which we derive a syntactic-lexical
semantic model for defining and subcategorizing the word sense class. First learn the
grammatical distributions of each word from corpus to determine a noun is whether attribute
noun or not. Then we study the selectional restrictions between the noun and its collocate
words in each grammatical relation and assign the noun to more specific sense class. Finally
produce a small word sense taxonomy. We found that:
1. we get the syntactic behavior of word by studying the meaning of the word, thus sense
classes (meanings with same concept) could be distinguished by syntactic behavior;
2. creating word sense class requires both syntactic and semantic features which we call
syntagmatic features of word;
3. general meaning of collocate words is essential for acquiring the semantic features of
word sense class, which means some coarse and solid concepts of noun, verb and adjective
are needed. Such coarse and solid concepts are the precondition for studying the selectional
restrictions between collocate words.
References:
Adrienne Lehrer, Semantic fields and lexical structure, North-Holland Press, 1974
Beth Levin, English verb classes and alternations: a preliminary investigation, University of
Chicago Press, 1993
Thomas R. Gruber, Toward Principles for the Design of Ontologies Used for Knowledge
Sharing, Formal Ontology in Conceptual Analysis and Knowledge Representation,
Kluwer Academic Publishers, 1993
Scott McDonald, Michael Ramscar, Testing the Distributional Hypothesis: The Influence of
Context on Judgements of Semantic Similarity, Proceedings of the 23rd Annual
Conference of the Cognitive Science Society, 2001
Wang Hui, A Syntagmatic Study on Noun Senses in Contemporary Chinese, Peking University
Press, 2004
ames Pustejovsky, Towards a Generative Lexical Resource: The Brandeis Semantic Ontology,
Proceedings of the Fifth Language Resource and Evaluation Conference, 2006