metaphor recognition: chmeta, a pattern-based …of metaphorical expressions) by providing a...

38
Computational Intelligence, Volume 25, Number 4, 2009 METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED SYSTEM YUN Y ANG, 1 CHANGLE ZHOU, 1,2 XIAOJUN DING, 1,3 JIAWEI CHEN, 1 AND XIAODONG SHI 1 1 Institute of Artificial Intelligence, Xiamen University, Xiamen, P.R. China 2 Center for the Study of Language and Cognition, Zhejiang University, Hangzhou, P.R. China 3 College of Foreign Languages and Cultures, Xiamen University, Xiamen, P.R. China Metaphor recognition presents a computational challenge, in part due to metaphoric deviation from literal thinking, and also because of a metaphor’s various linguistic expressions. This article forwards a new computational method, an integrated treatment of metaphor recognition from the computational perspective, which recent related studies have not entirely addressed. The authors differentiate metaphor recognition from complex metaphor inference and interpretation employing psychological clues. To accomplish this, we have developed a formalized system of metaphorical expression in metaphor role dependency schema, which specifically defines, classifies, and quantifies metaphorical anomalies, building a computable classification system for metaphors (incorporating 32 major patterns of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through a pattern recognition method and a metaphor components’ tagging approach. This metaphor recognition and tagging system is named and implemented as “CHMeta.” Experiment results support the validity and efficiency of this metaphor recognition system. Compared with most metaphor computation systems, which work mainly on a few examples, this system classifies major metaphorical expressions from a computational perspective and is able to recognize a variety of different kinds of metaphors, including nested ones. Thus, this is the first integrated work in computable classification, recognition, and tagging of large-scale metaphors in Chinese. Key words: natural language processing, pattern recognition, metaphor recognition, metaphor role dependency schema, referential unconventionality, collocational unconventionality, CHMeta. 1. INTRODUCTION The rich expressive power of metaphor as a form of speech and its cognitive nature as a form of thought make the understanding of metaphor a focus of linguistic studies, philosophy, psychology, and cognitive science. Given that a method for computationally interpreting metaphorical language would be useful for natural language processing and natural language understanding (Zhou 2003; Mason 2004), metaphor computation has attracted substantial attention during the last few decades (Zhou, Yang, and Huang 2007). A computational model of metaphor comprehension involves at least two stages: the recognition of a metaphor and the inference of its true meaning. Most studies have been fo- cused on meaning inference of known metaphors, whereas the recognition of novel metaphors has received inadequate attention, limiting the integrated development of a metaphor com- putation system. To date, there has been no robust, broadly applicable system of metaphor computation (Mason 2004; Zhou, Yang, and Huang 2007). An accurate and comprehensive computational metaphor recognition system is required. Many recent computational methods of metaphor recognition are limited in depth and extension: they make use of semantic anomaly, but fail to abstract the internal mechanism of metaphor, including how many and what kind of metaphorical anomalies there may be, where the anomalies would lie in a metaphorical sentence, and how such anomalies attract people’s awareness. As a result, these methods do not provide a valid classification system of metaphorical expressions from a computational point of view. They only work on a few examples or a single type of metaphor and lack a large-scale treatment of metaphorical expressions. Moreover, few of the recognition methods consider further identification of metaphor components, such as “atom” as the real topic and “solar system” as the referent vehicle of the real topic when the metaphor “An atom is a solar system” is recognized. Address correspondence to Yun Yang, No.49 DongRongXian Hutong, XiCheng District, Beijing 100031, P.R. China; e-mail: [email protected] C 2009 The Authors. Journal Compilation C 2009 Wiley Periodicals, Inc.

Upload: others

Post on 24-Jun-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

Computational Intelligence, Volume 25, Number 4, 2009

METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED SYSTEM

YUN YANG,1 CHANGLE ZHOU,1,2 XIAOJUN DING,1,3 JIAWEI CHEN,1 AND XIAODONG SHI1

1Institute of Artificial Intelligence, Xiamen University, Xiamen, P.R. China2Center for the Study of Language and Cognition, Zhejiang University, Hangzhou, P.R. China

3College of Foreign Languages and Cultures, Xiamen University, Xiamen, P.R. China

Metaphor recognition presents a computational challenge, in part due to metaphoric deviation from literalthinking, and also because of a metaphor’s various linguistic expressions. This article forwards a new computationalmethod, an integrated treatment of metaphor recognition from the computational perspective, which recent relatedstudies have not entirely addressed. The authors differentiate metaphor recognition from complex metaphor inferenceand interpretation employing psychological clues. To accomplish this, we have developed a formalized system ofmetaphorical expression in metaphor role dependency schema, which specifically defines, classifies, and quantifiesmetaphorical anomalies, building a computable classification system for metaphors (incorporating 32 major patternsof metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target inputsentence through a pattern recognition method and a metaphor components’ tagging approach. This metaphorrecognition and tagging system is named and implemented as “CHMeta.” Experiment results support the validityand efficiency of this metaphor recognition system. Compared with most metaphor computation systems, whichwork mainly on a few examples, this system classifies major metaphorical expressions from a computationalperspective and is able to recognize a variety of different kinds of metaphors, including nested ones. Thus, this isthe first integrated work in computable classification, recognition, and tagging of large-scale metaphors in Chinese.

Key words: natural language processing, pattern recognition, metaphor recognition, metaphor role dependencyschema, referential unconventionality, collocational unconventionality, CHMeta.

1. INTRODUCTION

The rich expressive power of metaphor as a form of speech and its cognitive nature as aform of thought make the understanding of metaphor a focus of linguistic studies, philosophy,psychology, and cognitive science. Given that a method for computationally interpretingmetaphorical language would be useful for natural language processing and natural languageunderstanding (Zhou 2003; Mason 2004), metaphor computation has attracted substantialattention during the last few decades (Zhou, Yang, and Huang 2007).

A computational model of metaphor comprehension involves at least two stages: therecognition of a metaphor and the inference of its true meaning. Most studies have been fo-cused on meaning inference of known metaphors, whereas the recognition of novel metaphorshas received inadequate attention, limiting the integrated development of a metaphor com-putation system. To date, there has been no robust, broadly applicable system of metaphorcomputation (Mason 2004; Zhou, Yang, and Huang 2007). An accurate and comprehensivecomputational metaphor recognition system is required.

Many recent computational methods of metaphor recognition are limited in depth andextension: they make use of semantic anomaly, but fail to abstract the internal mechanismof metaphor, including how many and what kind of metaphorical anomalies there may be,where the anomalies would lie in a metaphorical sentence, and how such anomalies attractpeople’s awareness. As a result, these methods do not provide a valid classification systemof metaphorical expressions from a computational point of view. They only work on a fewexamples or a single type of metaphor and lack a large-scale treatment of metaphoricalexpressions. Moreover, few of the recognition methods consider further identification ofmetaphor components, such as “atom” as the real topic and “solar system” as the referentvehicle of the real topic when the metaphor “An atom is a solar system” is recognized.

Address correspondence to Yun Yang, No.49 DongRongXian Hutong, XiCheng District, Beijing 100031, P.R. China;e-mail: [email protected]

C© 2009 The Authors. Journal Compilation C© 2009 Wiley Periodicals, Inc.

Page 2: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

266 COMPUTATIONAL INTELLIGENCE

Practical identification and tagging for such components is essential for further metaphorinference after recognition.

Metaphor recognition works directly at the linguistic level, dealing with words andphrases, but it serves metaphor inference at the level of thought, acting as a bridge connectingtwo levels of work. To our minds, integrated metaphor recognition work should serve thefollowing purposes:

1. To identify metaphors from literal expressions at the level of natural language.2. To identify and tag all useful information about the recognized metaphor, including the

category it belongs to and what its metaphorical components are (for metaphor inferenceat the level of thought).

To achieve these purposes, we present a pattern-based view, which isolates metaphorrecognition from complex metaphor inference and interpretation. We argue that variousmetaphorical expressions can be unified into a finite set of linguistic patterns with essentialmetaphorical syntactic and semantic characteristics.

In this article, we employ Dependency Grammar and develop a Metaphor Role Depen-dency Schema Matching Theory to deal with the metaphor recognition problem. We proposea computable classification system with 32 categories of common Chinese metaphoricalexpressions. Our classification system is a finite set of essential patterns of metaphoricalexpressions. The 32 metaphor categories are normalized and stored as class schemas, whichare used for potential metaphor matching. An input sentence is first parsed and formalizedas a dependency schema and then the classifier judges if it contains any of the 32-classschemas through schema matching. If matching class schemas exist then an anomaly mea-suring procedure is evoked to compute correlative measuring parameters to see whether thepotential units are actually a metaphor. If the measurement values are satisfied then the unitsin the input sentence are recognized and tagged as metaphorical. The theory incorporatesthe Schematization Principle, a computable classification of Chinese metaphors (32 classschemas), and definition and computation of measurements. The recognition system is madeup of four computing modules: a schema formalization module, embedded schema match-ing module, semantic anomaly finding module and a metaphor role positioning and taggingmodule.

The rest of the article is organized as follows. Section 2 provides an overview and detailsof the Metaphor Role Dependency Schema Matching Theory, illustrating how metaphorrecognition is converted into pattern recognition. Section 3 discusses the implementation ofthe metaphor recognition system and describes the experiments illustrating the idea. Section 4discusses the results of the experiments. Section 5 surveys current approaches and comparesour work with these. Section 6 draws conclusions and contemplates future work.

This article derives its theory from Chinese metaphor. However, the idea and principlescan be extended to other languages.

2. METAPHOR ROLE DEPENDENCY SCHEMA THEORY

2.1. Definitions and Goals

In our research, we are more interested in identifying novel metaphors that have more ac-tive influence on daily language understanding than conventional “dead metaphors” (Lakoff1993; Lakoff and Johnson 1980; Lakoff et al. 1994). We identify all the expressions ofobject/concept A in terms of object/concept B or terms belonging to B’s category, where

Page 3: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 267

there is some similarity or correlation between A and B, as metaphor in language. Further,following Goatly’s usage, A is called the Topic, B is called the Vehicle, and their similaritiesare called Grounds (Goatly 1997). Expressions that follow the principle are all included inour metaphor recognition treatment.

The major difficulties for metaphor recognition are caused by random linguistic expres-sions of Topics, Vehicles, and Grounds. Goatly (1997) has defined Topic terms, Vehicle terms,and Ground terms to distinguish these objects/concepts from the language used to expressthem. Take for example the metaphor Life is a box of chocolates; you never know what you’regoing to get.

Life is the Topic term, which stands for the actual unconventional referent in the metaphor.A box of chocolates is the Vehicle term, which stands for the conventional referent. And younever know what you’re going to get is the Ground term, which stands for the similaritiesand/or analogies involved (Goatly 1997).

Topic terms, vehicle terms, and ground terms can be single words, phrases, clauses,or even larger units of discourse. There are no semantic rules to tell when the terms arewords, phrases or clauses. Hence, if we take terms of linguistic metaphorical elements asmathematic sets, such sets would be infinite and without any descriptive rules, one of thecentral difficulties for metaphor recognition. Further, linguistic realization of most metaphor“terms” (except for vehicle) is optional, also making metaphor recognition a difficult problem.

Metaphor Role Dependency Schema Theory deals with the following issues:

1. How to formalize a linguistic sentence and represent the syntactic and semantic relationsbetween linguistic components.

2. How to locate where potential metaphorical semantic relations would lie and how toquantify the judgment for computation requirements.

3. How to decide which class the recognized metaphor belongs to.4. Once a linguistic metaphor is discovered, how to identify and tag its topic term, vehicle

term, and other related information. Take, for example, the following Chinese metaphor:

(Chinese).Lu4-deng1-xia4 na4-ge4 mei3-li4-de ying3-zi, shi4 yi4-fu2 gan3-ren2-de hua4-mian4 (Chinese

Pinyin).Street lamp below that beautiful shadow, is a moving picture (English

equivalents).That beautiful shadow under the street lamp is a moving picture (English translation).

The topic term is the phrase “ (that beautiful shadow under the street lamp)”not the single word “ (shadow),” and the vehicle term is the phrase “ (a verymoving picture)” rather than the single word “ (picture).” The mark “ (is)” also needs tobe tagged because it implies the semantic relation of topic and vehicle is referential.

2.2. Hypothesis for Metaphor Recognition

In investigating people’s behavior in metaphor processing, it appears that different kindsof metaphors share a relatively common method of recognition. We argue that people areable to recognize a metaphor without inferring meaning or even without understanding thetrue meaning. Inference and interpretation work is more likely to be done after recognition.Thus, metaphor recognition stands in its own right, distinguishable from metaphor inferenceand interpretation. We summarize this feature of metaphor as follows:

Page 4: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

268 COMPUTATIONAL INTELLIGENCE

1. To recognize a metaphor is actually the process of discovering anomalies within syntacticstructures and shallow semantic relations between linguistic components, rather thaninvolving any complex meaning reasoning. The reasoning process is only called uponwhen people are asked to deduce the metaphor’s true meaning (Yang, Li, and Zhou 2008).

A manifestation of this hypothesis is that people can quickly make a judgment whethera sentence embeds a metaphor. On the other hand, it will take people much more time tocomprehend and interpret a metaphor. Thus, people can be aware of metaphors even thoughthey do not know their exact meaning at the time. This hypothesis separates metaphorrecognition from the more complicated process of metaphor understanding.

2.3. The Idea of Patterns for Metaphor Recognition

We are not concerned with accurate metaphor interpretation but rather the essential trig-gering attributes. Thus, recognition, inference, and interpretation can be discrete. Metaphor-ical expressions are infinite and spread throughout linguistic expressions, but there may be arelatively finite set of linguistic patterns with essential metaphorical syntactic and semanticcharacteristics. People cannot memorize all metaphorical expressions but they can masterlinguistic patterns with their basic language abilities. Within a mastered pattern, peopleare able to locate potential metaphorical counterparts and compare them with normal ortraditional usage. In order to describe the recognition process, we proposed the followingpresuppositions in constructing our patterned recognition theory:

1. Separability presupposition: Metaphor recognition can be isolated from metaphor infer-ence and studied in its own right.

2. Patterning presupposition: People cannot master and have readily available all themetaphorical expressions needed to deal with metaphor recognition, but they can masterand distinguish structures and shallow semantic relations among linguistic componentsembedded in metaphors and take advantage of this knowledge to respond to semanticanomalies within linguistic patterns.

The recognition process is manifested without drawing on numerous hand-coded rules,nor conceptual categorical knowledge and syntactic and shallow semantic relations of lin-guistic units are obtained from parsers.

The processes for recognizing The man is a lion and The man is a trapped, helplessand anguished lion proceed the same way: using their language abilities people first capturethe idea that the semantic relation of man and lion are referential, and in succession theymove to the consciousness of the unconventionality of the man– lion reference. Whereas themodifiers trapped, helpless and anguished contribute little to the recognition of man–lionreferential metaphor, they are important for interpreting its meaning.

In sum, the principle of metaphor triggering is this: for a person to recognize a linguisticmetaphor, he must first be clear about the syntactic locations of linguistic components andthe shallow semantic relations between them, and then come to the awareness of uncon-ventionalities lying in the potential metaphorical syntactic and shallow semantic structureswhich deviate from the literal word sense using his world knowledge.

2.4. From Metaphor Recognition to Pattern Recognition: The Schema

Based on the above presuppositions, metaphors are represented as pieces of syntac-tic structures to sketch linguistic semantic relations of components in well-defined ways.

Page 5: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 269

FIGURE 1. From linguistic level to cognitive thought level.

Metaphor Role Dependency Schema Matching Theory is developed to deal with the large-scale metaphor recognition problem.

2.4.1. Metaphor Roles. Metaphor roles are the linguistic expressions of main com-ponents of a metaphor at language level, which are different from but directly related tocognitive concepts in the world (Figure 1).

Topic role: The linguistic unit denoting the topic of the metaphor in the real world. Inmany cases, the conceptual topic does not directly appear in a linguistic metaphor, instead itis part of other features expressed by Topic role.

Vehicle role: The linguistic unit denoting the vehicle of the metaphor in the real world.Ground role: The linguistic unit that implies similarities of topic and vehicle.Mark role: Metaphor tags such as is, is like, as. . .as.Take Loneliness is singing for example. Loneliness is a word in language, and it is the

topic role of the metaphor at the linguistic level and directs us to the real topic concept—theabstract concept loneliness. Its equivalent vehicle is a concrete entity—human being, whichdoes not directly appear in the linguistic sentence but was directed by some of its typicalactions, here “sing.” Thus singing is the vehicle role by our definition. Moreover, the wordsinging also directs us to the ground of the metaphor and acts both in a vehicle role andground role. This metaphor may use some attribute of the action “sing” of human beings todescribe a certain feature of the abstract concept loneliness.

Except for vehicle role, topic, ground, and mark roles do not necessarily appear in alinguistic metaphor (Goatly 1997). Roles can be extended to a single word, a single lexicalitem, a phrase or to a clause which directs us to cognitive concepts in the world.

2.4.2. Schematization. No matter how complex metaphorical roles are, there is a cen-tral or head word for each role. For example, the attributive phrase a trapped, helpless, andanguished lion has a head word lion, and modifiers trapped, helpless, and anguished dependon it. Semantic relations between metaphor roles are linked directly by their head words.Hence, The man is a lion and The man is a trapped, helpless, and anguished lion share thesame metaphor role skeleton: man is lion, see Figure 2.

Page 6: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

270 COMPUTATIONAL INTELLIGENCE

FIGURE 2. The metaphor role skeletons.

As with Metaphor Roles and Dependency Grammar (Robinson 1970), a unit of metaphoris formalized as a dependency tree and has a well-defined Metaphor Role DependencySchema. The principles are:

1. A unit of linguistic metaphor can be represented as a skeleton of metaphor roles and thedependency semantic relations between them.

2. Each metaphor role has one and only one head word called the Role Head.3. Dependency relations between metaphor roles are denoted by dependency relations be-

tween Role Heads.

A unit of metaphor considered in this article is limited within sentence level. A sentencemay embody more than one metaphorical unit. Take as an example the following:

(Chinese).Ta1 ba1-bang4-zhe bi4-ye4 jiu4-xiang4 qiu2-fan4 ba1-wang4 chong2-huo4 zi4-you2 (Chinese

Pinyin).He look-forward-to graduate just-like prisoners look-forward-to regain freedom (English

equivalents).He looks forward to graduation just as prisoners look forward to regaining freedom (English translation).

The dependency structure is as shown in Figure 3. The Chinese sentence is segmentedinto a collection of words with their part-of-speech tags beneath them. The dependency

FIGURE 3. The parsed structure of the example metaphor “ (He looks forward tograduation just as prisoners look forward to regaining freedom).” In the dependency figure, the English wordtranslations beneath correlated Chinese words are written in accordance with the original Chinese ones to reflectthe Chinese sentence constructions are not necessarily correct English sentences.

Page 7: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 271

FIGURE 4. The schematizing of a linguistic metaphorical unit.

relations between words are denoted as a directed arc from dependent word to head word.According to Dependency Grammar, the verb “ (is like)” is the head of the sentence andalso the mark role. (See Appendix A for meanings of part-of-speech tags and dependencyrelation tags.)

The dependency structure is then transformed into a dependency tree as shown inFigure 4(a). The copula “ (is like)” is the tree root and also the Mark role tagged by Met_m.The phrase “ (he looks forward to graduation)” is the Topic role and is marked bythe tag Met_t. The phrase “ (prisoners look forward to regaining freedom)” is theVehicle role and is marked by the tag Met_v. The topic role is the subject of “ (is like)” andthe vehicle role is the object. The two roles are counterpart subtrees. Their subject–predicaterelation (SBV) and object–predicate relation (VOB) with tree root are denoted through therelations between the role heads (tree roots). The Chinese verb “ (look forward to)” is thehead of the Topic role and Vehicle role and also the subtree root in each circled subtree. If weomit the concrete words in the dependency tree and only retain the semantic relations andPOS tags, then we obtain the dependency schemas as shown in Figure 4(b).

In sum, each unit of a linguistic metaphor can be formalized as a dependency tree anda dependency schema. Each metaphor role can also be transformed into a subtree. Roleheads are tree roots. Given that a tree root determines a tree, if we determine the relationsbetween tree roots, then we can trace the relations between trees. When a unit of metaphor isformalized as a dependency schema, metaphor recognition becomes a hierarchical process:i.e., to discover the unconventional referential of two concepts or semantic collocation oftwo linguistic components is to find unconventionalities between pairs of dependency nodesin a dependency schema. When an anomaly is found, the tagging process becomes specific:i.e., fix the role head node and then mark the entire subtree.

However, in the selected pair of nodes, the problem concerning which node is the topicrole head and which is the vehicle role head depends on different schema and is in the domainof computing criteria. The next task is to classify various units of metaphor and build a classbase for metaphor role dependency class schema.

2.4.3. Computable Classification and Class Schema. We extracted dependencyschemas from about 500 units of Chinese metaphorical sentences extracted from the 1998corpus of the Chinese newspaper People’s Daily (the corpus is 8.42 MB and is available

Page 8: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

272 COMPUTATIONAL INTELLIGENCE

FIGURE 5. The growth curve of dependency schemas of metaphor.

on-line). Figure 5 shows a closed set of dependency structures of metaphor roles. We makeuse of the dependency semantic relations between metaphor roles and classify them intofinite class schemas.

According to Metaphor Role Dependency Analysis, to classify metaphorical phenomenais actually to classify metaphorical or semantic anomalies and their potential locations inmetaphorical sentences. By Goatly’s definition, “metaphor occurs when a unit of discourseis used to refer unconventionally to an object, process or concept, or colligates in an uncon-ventional way” (Goatly 1997). We first put metaphorical anomalies into two major classes:Referential and Collocational. Within each major class, we subclassified them according tosyntactic locations and shallow semantic relations of semantic anomalies. In their entirety,the classes are presented as 32 schemas in this article. Each schema denotes a separate unit ofmetaphor and they can be nested and conjoined to each other to form compound metaphors.

1. Referential Metaphor occurs when a unit of discourse is used to refer to another unitof discourse to which it does not conventionally refer. There are 21 class schemas forReferential Metaphor.For example, The man is a lion and The man is a trapped, helpless, and anguished lionare referential metaphors.(See Appendix B for details of 21 different schemas of referential metaphor.)

2. Collocational Metaphor occurs when a unit of discourse is used to collocate with aunit(s) to which it does not conventionally collocate. There are 11 class schemas forCollocational Metaphor.For example, My car drinks gasoline and Loneliness is singing are collocationalmetaphors.(See Appendix C for details of 11 different schemas of collocational metaphor.)

The 32 class-schemas (Appendix B and C) are syntactic and semantic structures forlinguistic components in possible metaphors. Class schemas only contain semantic relationsas edges and part-of-speech tags as nodes, but not concrete words. Each class schema hasits own computing criteria: the position of metaphor roles and related parameter computing.The class schemas have been preserved in a schema bank.

In sum, no matter how distinct the metaphorical expressions, they could be schematizedand did embed several class schemas. To recognize a metaphor is a process of locatingreferential or collocated pairs in a certain dependency schema and then measuring whetherthere are unconventionalities in them. The process can be quickly manipulated by schema

Page 9: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 273

matching and anomaly calculation rather than any complex reasoning work, and is thus moreappropriate for metaphor recognition.

2.5. Main Process

Based on Metaphor Role Dependency Schema Matching Theory, the metaphor recog-nition and tagging process follows the steps of schema formalization, embedded schemamatching, measuring (unconventionality finding), Metaphor Role Positioning, and Tagging.Note the recognition process in the following sentence for example. The process is illustratedin Figure 6.

(Chinese).Jiao4-shi1 shi4 ren2-lei4 ling2-hun2 de gong1-cheng2-shi1 (Chinese Pinyin).Teacher is human soul ’s engineer (English equivalents).A teacher is the engineer of a human soul (English translation).

FIGURE 6. Detailed process for metaphor recognition.

Page 10: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

274 COMPUTATIONAL INTELLIGENCE

Step 1: Parse the input sentence using the dependency parser (see Figure 6a).The parsing procedure is similar to the ability of people to recognize basic linguistic

components and their shallow semantic relationships in a sentence.Step 2: Employ a tree generation procedure to transform the parsed sentence into a

dependency tree and a dependency schema as shown in Figure 6(b) and (c).Step 3: Locate all class-schemas embedded in the sentence schema.This is to confirm the maximum number of metaphorical units the sentence may contain

and is useful for the next procedure: Selecting the final embedded metaphorical units byrelevant parameter computing. In the above example, there are three class-schemas embedded(see Figure 6d): Schemas 01, 28, and 29.

Step 4: Determine whether the embedded schemas are true metaphorical units or not(see Section 3.3).

This is done to compute correlative parameters to determine whether there are referentialor collocational unconventionalities in matched schemas (see Figure 6e). The result is that thesentence ultimately contains two metaphorical units: A referential metaphor in accordancewith class schema 01 and a collocational metaphor in accordance with class schema 28. Instep 4, class schema 29 is removed from the next steps because the value of the parameterdoes not exceed the normal range.

Step 5: Trace and tag metaphor roles using the matched class schema tagging rules.For this example, according to rules of schema 01, the left subtree and the right subtree

of the root “vx” (vx denotes copula) are two referent counterparts. The left subtree is taggedas the topic role and the entire right subtree (including the nodes in schemas 28 and 29) istagged as the vehicle role (see Figure 6f).

Two metaphors are found in the sentence. There is also an embedded relation betweenthe two metaphors: Human soul’s engineer is nested in Teacher is human soul’s engineer.This is another merit of our recognition theory.

3. IMPLEMENTATION AND EXPERIMENTS

Having presented a sample of our metaphor recognition model, we now shift to thecomputational system. For the following sections, we will call the system CHMeta.

As Figure 7 shows, the system is composed of four main computing modules: schemaformalization module, embedded schema matching module, parameter computing (uncon-ventionality finding) module, and metaphor role positioning and tagging module.

3.1. Parsing

The input sentences are initially parsed with the well-known Chinese dependency parser,HIT IR-Lab_Shared_Parser (HIT-IRLab 2006, Liu and Ma 2006). In this Chinese depen-dency parser, word sense disambiguation has received treatment. Thus, word sense disam-biguation is remained in this parsing module. The parsed result is stored in three lines:the first line is the raw sentence; the second line is the word segmentation, location, andparts-of-speech; and the third line stores the dependency pairs and relations. The parsedresult of the Chinese sentence “ (Teacher is human soul’s engineer)” is asfollows:

Line 1: .Line 2: [1] /n [2] /vx [3] /n [4] /n [5] /ue [6] /n [7] /wp

[8] <EOS>/<EOS>.

Page 11: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 275

FIGURE 7. The architecture of the metaphor recognition and tagging system CHMeta.

Line 3: [4] _[3] (ATT) [5] _[4] (DE) [2] _[1] (SBV) [6] _[5] (ATT)[2] _[6] (VOB) [8]<EOS>_[2] (HED).

Line 1: Teacher is human soul’s engineer.Line 2: [1]teacher /n [2]is /vx [3]human /n [4]soul /n [5]’s /ue [6]engineer /n

[7] /wp [8]<EOS>/<EOS>.Line 3: [4]soul_[3]human being(ATT) [5]’s_[4]soul(DE) [2]is_[1]teacher(SBV)

[6]engineer_[5]’s(ATT) [2]is_[6]engineer(VOB) [8]<EOS>_[2]is(HED).

The dependency figure and tree structure is presented in Figure 6(a) and (b).

Page 12: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

276 COMPUTATIONAL INTELLIGENCE

FIGURE 8. Examples of embedded matching and embedded matching nodes.

3.2. Embedded Schema Matching

Embedded schema matching is performed to find and mark all the class schemas em-bedded in the input dependency schema. The goal of this process is to find all the potentialmetaphorical anomalies’ locations in the input sentence. After schema matching, the impor-tant components and their dependency relations are separated and saved for unconventionalitymeasuring. Figure 8 is an example of the embedded matching between the dependency treeT1 (class schema) and T2 (input dependency schema).

3.3. Measures of Semantic Unconventionalities

To achieve recognition—as well as for locating potential metaphorical semantics andrelations provided by dependency parsing and schema matching—measurements and theirvalues for finding unconventionality in referential and collocational schemas must be avail-able. The outcomes of unconventionality calculating will help to make a decision whether theselected potential metaphorical units are real metaphors or not. If the unconventionality valueexceeds a threshold, the potential metaphorical units are considered to be a true metaphor,otherwise the potential metaphorical units are not considered metaphorical.

3.3.1. Semantic Unconventionality Measurements. Two measurements are employedfor this study: Referential unconventionality for referential schemas and collocational un-conventionality for collocational schemas.

Referential unconventionality is employed to measure how unconventional one conceptis in regard to another. There are several reasons to refer A to B. For practicality, weconsider that one concept unconventionally refers to another if either of their categories isfar removed from each other or they are hyponymy irrelevant. This can be specified by thecategory remoteness or hyponym irrelevancy of the two referential words (concepts), andthis remoteness or irrelevance can be calculated.

For example the two Chinese metaphors:

a. (Chinese).Lv4-shi1 shi4 hu2-li (Chinese Pinyin).Lawyer is fox (English equivalents).The lawyer is a fox (English translation).

b. (Chinese).Lv4-shi1 shi4 ge4 zhi2-ye4 (Chinese Pinyin).Lawyer is an occupation (English equivalents).Being a lawyer is an occupation (English translation; In Chinese, there is no morphologicaldifference between words. Thus, the Chinese word “ (lawyer)” can both refer to a personand the occupation of lawyer.)

Page 13: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 277

The two referential pairs ( lawyer, fox) and ( lawyer, occupation) are inaccord with the different categories of principles. However, the former is a metaphor, thelatter is a hyponymy reference rather than a metaphor.

Collocational unconventionality is used to measure the degree of unconventionality inthe semantics underlying a collocation of two words, when compared to usage-based expec-tations. To calculate this parameter, we propose an example-based approach: people cannotmemorize and master all word collocations. They generally hold a finite set of conventionalcollocations and semantic rules. When they encounter a new collocation they search for somesample collocations that have the same semantic relationship as the new collocation, anddetermine whether the new collocation is cognitively acceptable by comparing its similarityto the sample collocations. For example, people regard “ (drink ink)” unconventionalbecause “ (ink)” is far outside the conventional objects of the verb “ (drink)” such as“ (boiled water),” “ (tea),” “ (wine).”

Dependent upon different knowledge resources, the calculating formulas may vary. Overthe next two sections, we will introduce the knowledge resources we use and formulas wesuggest to calculate the value of these unconventionalities.

3.3.2. Resources for Calculating Semantic Unconventionality. Strictly speaking, thenecessary knowledge for a computer to discover metaphors should be no less than theknowledge of people to do the same work. We make use of two acclaimed Chinese resources:TongYiCi CiLin and HowNet.

• TongYiCi CiLin (Mei 1983; HIT-IR Lab 2005).TongYiCi CiLin is a Chinese synonym dictionary available in electronic form, which isstructured as illustrated in Figure 9. TongYiCi CiLin (Extended Edition) includes 77,343words that are structured as trees.

• HowNet (Dong and Dong 1999, 2006).

HowNet is an on-line extra-linguistic knowledge system for the computation of mean-ing in human language technology. It is a commonsense knowledge base, revealing

FIGURE 9. Structure of TongYiCi CiLin.

Page 14: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

278 COMPUTATIONAL INTELLIGENCE

FIGURE 10. Template of word concepts in HowNet.

FIGURE 11. Two concepts of the Chinese word “ ” (extracted from HowNet 2000 Version).

inter-conceptual relations and inter-attribute relations of concepts connoted in Chinese lexi-cons and their English equivalents. It includes both word entries and concept entries. Thereare more than 60,000 Chinese concept entries and approximately 70,000 English conceptentries in HowNet. Unlike WordNet (Fellbaum 1998), in which a semantic relation is a rela-tion between synsets, HowNet adopts a constructive approach to semantic representation. Itdescribes base forms of words as representing a set of one or more concepts and describeseach concept using a set of sememes, which are the smallest semantic units in HowNet andcannot be deconstructed further. The template of word concepts is organized in HowNet asshown in Figure 10.

In this article, we make use of association of sememes as stated below.A word in HowNet is defined as a set of concepts, and each concept is represented by

primitives. We express HowNet as a collection of n words: W : W = {w1, w2, . . . , wn}.Each word wi is in turn described by a set of concepts (meanings) c: wi =

{ci1, ci2, . . . , cix} and each concept ci is in turn described by a set of sememe s: ci ={si1, si2, . . . , siy}

For example, Figure 11 shows two concepts represented by the Chinese word “ (beat)”:one is “play,” the other is “fight.” The “DEF” items are sememes that interpret concepts.

Sememes are linked by a hierarchical tree to indicate the parent–child relationships asillustrated in Figure 12.

This hierarchical tree structure presents a way to link one concept with any other inHowNet. The distance between the two represents the closeness of concepts. Dependent uponinheritable characteristics, hyponym words will inherit the sememes of their hypernym se-memes. Thus, two words with little apparent similarity may have close hypernym–hyponymsignificance according to their sememes.

• Sample base.

A sample base is constructed for collocational unconventionality computing.Section 3.3.1 describes a sample-based approach. For example, people will find drink inkunconventional because ink is far removed from the conventional objects of drink, such as

Page 15: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 279

FIGURE 12. Hierarchical tree structure of sememes in HowNet.

FIGURE 13. Sample base of Chinese verbs.

boiled water, tea, wine, etc. A sample base is similar to the prior knowledge of conventionalusage people have already mastered. Take the sample base of Chinese verbs for example.The sample base for Chinese verbs consists of conventional verbal collocations. It is derivedfrom a 24.6 MB tagged dependency tree bank (HIT IR Dependency Tree Bank) and the Col-location Dictionary of Chinese Common Verbs (Li et al. 2007a). The sample base containssemantic collocations for 1,273 common Chinese verbs, and is structured as in Figure 13.

There are five lines to describe the verb and its collocated words. Line 1 is the head verb.Line 2 begins with the tag [SBV] and the conventional sample subjects of the verb are listedbehind the tag. Line 3 begins with the tag [VOB] and the conventional sample objects ofthe verb are likewise listed behind the tag. Line 4 is the complemental list and line 5 is theadverbial list of the verb.

The sample base supplies a small set of conventional collocation examples. It is not largeenough to include all the conventional sample collocations. As the sample base is designed toillustrate the semantic features of Chinese collocations, it also reflects ontological relationsamong concepts. Thus, it could be dynamically enlarged through combination with synonymbases. One way of doing this (which we have adopted in our algorithm) is to introducesemantic similarity measures between words and extend collocated samples, taking thesimilarity measures into account. For example, the sample of [ (drink wine)] in samplebase can be extended to [ (drink juice)] and [ (imbibe wine)] according to the categorysimilarity value of “wine and juice” and “drink and imbibe.” Thus, the sample base can becombined with TongYiCi CiLin to supply more conventional sample collocations.

3.3.3. Calculating Semantic Unconventionality. By making use of knowl-edge resources, semantic unconventionalities can be quantitatively computed. The

Page 16: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

280 COMPUTATIONAL INTELLIGENCE

measurements can be calculated by different methods using different knowledge resources.In this article, we have developed the following computing strategies for unconventionalitycomputing.

Referential unconventionality: UnconvRef (w1, w2) of word w1 and w2 is defined in thisarticle as the reciprocal of the maximum degree of word similarity and hyponym relevancy:

UnconvRef (w1, w2) = 1

max[Sim(w1, w2),Hypo(w1, w2)],(1)

where UnconvRef (w1, w2) ∈ R, UnconvRef (w1, w2) ≥ 1. Sim(w1, w2)(0 ≤ Sim(w1,w2) ≤ 1) is the similarity of w1 and w2 defined as two words that can be taken as thesame category. Hypo(w1, w2)(0 ≤ Hypo(w1, w2) ≤ 1) is the hyponym relevancy of w1 andw2 which is defined as the relevancy that w1 is the hyponymy of w2.

In this article, it is calculated from TongYiCi CiLin. The formula is

Sim(w1, w2) = β

Dis(w1, w2) + β, (2)

where β is an adjustable parameter with a value of 1.1 and Dis(w1, w2) is the pathlength between w1 and w2 based on the semantic tree structure used for TongYiCi CiLin(Figure 9).

The above formula does not explicitly indicate that the depth of a pair of nodes in thetree affects their similarity. For two pairs of nodes (w1, w2) and (w3, w4) with the samedistance, the deeper the depth the more commonly shared ancestors have, which means theyare semantically closer to each other.

In this article, we calculate Hypo(w1, w2) by knowledge extracted from HowNet. Ac-cording to HowNet, if w2 itself or its prime sememe are one of w1’s sememes, then they arehyponym relevant. Thus, two words with low similarity may have close hypernym–hyponymrelationships according to their sememes. We define the hyponym relevancy value of twowords to represent their hyponym relation. The formula is as follows:

Hypo(w1, w2) =L − min

i=1...nlocation(c1, s2i )

L, (3)

where n is the number of sememes of w2’s concepts, s2i is the sememe of w2’s concepts, c1 isthe concept of w1, and L is the number of sememes c1 contains. The value of location(c1, s2i )is an integer and 0 ≤ location(c1, s2i ) ≤ L − 1 indicates the location where sememe s2iappears in c1’s definition sememe sequence. The earlier s2i is located in c1’s definitionsequence, the more relevant s2i is to c1. If s2i does not appear in w1 ’s definition sememeset, then location(c1, s2i ) = L. Take Hypo(lawyer, occupation) for example. The sememe of“occupation| ” of the concept “ (occupation)” is located in lawyer’s definition sequence“DEF = human| ,#occupation| ,police| ,#law| .” Thus, Hypo(lawyer, occupation) =(4-1)/4 = 0.75.

Note that Hypo(w1, w2) �= Hypo(w2, w1). Hence, UnconvRef (w1, w2) may not equalUnconvRef (w2, w1) which implies a characteristic of metaphor.

“ (A lawyer is a fox)” and “ (Lawyer is an occupation)” are bothreferential sentences. The former is a metaphor while the latter is a conventional referen-tial expression in Chinese. Merely by computing similarity of referential pairs, we cannotdistinguish the difference because both have low similarity Sim(laywer, fox) = 0.21 and

Page 17: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 281

Sim(laywer, occupation) = 0.11. But if we add hyponym relevancy Hypo(lawyer, fox) = 0and Hypo(lawyer, occupation) = 0.75, then the distinction is clear.

UnconvRef (lawyer , fox ) = 1max[Sim(lawyer ,fox ),Hypo(lawyer ,fox )]

= 1Sim(lawyer ,fox ) = 4.76 > Threshold,

UnconvRef (lawyer , occupation) = 1max[Sim(lawyer ,occupation),Hypo(lawyer ,occupation)]

= 1Hypo(lawyer ,occupation) = 1.33 < Threshold.

Here, is a special example: “ (He is my son)” is a sentence which containspronouns. Pronouns add difficulties to metaphor recognition. Pronouns have low similarityvalue with common nouns in regular similarity calculations. Using our method, at sentencelevel without context, we will consider this sentence a normal sentence. According to thereferential unconventionality the function Hypo(he, son) functions as follows:

Sim(he, son) = 0.063Hypo(he, son) = 0.5

From this we obtain UnconvRef (he, son) = 2 < Threshold.The higher the UnconvRef (w1, w2) values, the more unconventional the reference is.In fact, the hyponym irrelevancy of the two referential words (concepts) helps the system

to treat nonmetaphorical pronoun expressions as normal at the sentence level. But for truemetaphorical pronoun references in a context not within a sentence, the method introducesrecognition errors. Hence, pronoun resolution is a challenge for metaphor recognition.

Collocational unconventionality: UnconvColl (wh, wd) of headword wh and collocated(dependent) word wd is defined in this article as

UnconvColl (wh, wd) = 1

maxi=1,...,n

[Sim(wh, si ) maxj=1,...,m

Sim(esi j , wd)], (4)

where UnconvColl (wh, wd) ∈ R, UnconvColl (wh, wd) ≥ 1.Sh is the synonym set of headword wh which is defined as Swh = {si : Sim(wh, si ) > θ},

where 0 < θ ≤ 1. We set θ 0.9 based on our experiment; n is the element number of Sh ,Sim(wh, si ) is the similarity value of headword wh and its synonym word si , and Esi is thesample set of si ’s collocated words which is defined as Esi = {esi : esi is the conventionalcollocated word of si with a specific dependency semantic relation}. Esi is extracted fromthe sample base (Figure 13) and “E-C/E” terms of HowNet (see Figure 10). m is the elementnumber of Esi and Sim(esi j , wd) is the similarity value of the dependent word wd and esi .

Take “ (drink mineral water)” and “ (drink ink)” for example. The two col-locations have a predicate–object structure with the headword (drink). The process is asfollows.

By searching in the example collocation base and TongYiCi CiLin, we get S ={ (drink), (imbibe)}. Sim(drink, drink) = Sim(drink, imbibe) = 1.000. From this, we ob-tain s1 = “ (drink)” and s2 = “ (imbibe).”

Next, we get the predicate–object collocated example set:

Es1 = { (tea), (soda), (white wine), (boiled water), (milk), (coffee),(drug), (soup), (juice)};

Es2 = { (wine), (tea)}.

Page 18: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

282 COMPUTATIONAL INTELLIGENCE

The result is:

maxj=1...9

Sim(es1 j , mineral water) = 0.73, maxj=1...2

Sim(es2 j , mineral water) = 0.61,

UnconvColl (drink, mineral water) = 1.37 < Threshold,

maxj=1...9

Sim(es1 j , ink) = 0.07, maxj=1..2

Sim(es2 j , ink) = 0.01,

UnconvColl (drink, ink) = 14.29 > Threshold.

3.4. Metaphor Role Tagging

As with recognition, tagging work follows the primary process. When an anomaly isfound, the tagging process is to identify the role head node, and mark the entire subtree. Themetaphorical role head nodes are stored as rules in class schemas. Hence, once unconven-tionality is found, the role heads are also fixed.

3.5. Algorithmic Details

Algorithms 1–4 show pseudocode for the metaphor recognition and tagging process.

Algorithm 1. Metaphor_recognition_tagging (Target_Schema)

Comment: Target_Schema is the parsed dependency tree of the input sentence. The mainalgorithm is used to find all metaphorical units embedded in the input sentence and tag eachfound metaphorical unit with its class number, topic role, vehicle role, mark role, and groundrole if it has one.

Selected_Sub_Tree1 matched with class schema ASelected_Sub_Tree 2 matched with class schema B ←Embedded_Schema_

Matching. . . (Target_Tree, Class_Schema_Bank)Selected_Sub_Tree n matched with class schema N

for each Selected_Sub_Treeif Is_metaphor (Matched_Class_Schema_Num, Selected_Sub_Tree) = = truethen Tagging (Selected_Sub_Tree);else output no potential metaphorical schema is found;

Algorithm 2. Embedded_Schema_Matching (Target_Schema, Class_Schema_Bank) (Li et al. 2007b)

Comment: Find all the class-schemas embedded in Target_Schema.for each Sub_tree ∈ Target_Schema

for each Schema_Tree ∈ Class-Schema_Bankif the structure of Sub_Tree matches with the Schema_Tree

and every node (part-of-speech) of the Sub_Tree matches with the correspondingnodes of the Schema_Tree

and dependency relations between nodes of the Sub_Tree match with the corre-sponding nodes’ relations of the Schema_Tree

then return (Sub_ tree, Matched_Class_Schema_num);else return (0);

Page 19: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 283

Algorithm 3. Is_Metaphor (Matched_Class_Schema_num, Selected_Sub_Tree)Comment: Computational parameters for nodes in Selected_Sub_Tree.

Matched_Class_Schema_num corresponds to the class the unit belongs to and thecomputing rules and parameter selected.

for referential pair of nodesif UnReferto (w1, w2) > THRESHOLD1 then return true;else return false;

for dependency collocated pair of nodesif UnColloca(wh, wd) > THRESHOLD2 then return true;else return false;

Algorithm 4. Tagging (Class_Schema_num, Selected_Sub_Tree)

Comment: Tag all the apparent metaphor roles.for the Selected_Sub_Tree

find all the metaphor role-head words in it following the rules of its schema class(Appendix C and D)

tag subtrees rooted with role heads

4. RESULTS AND DISCUSSIONS

4.1. Test Data

In order to test the performance of CHMeta, our experiments were conducted over thedata of Chinese metaphorical sentences, most of which are novel or active metaphors.

Test data consists of 400 Chinese sentences including metaphorical and nonmetaphoricalones extracted from the 23.6 MB large 1980–1997 corpus of DuZhe magazine, a popularChinese essay magazine, the articles of which cover fields including law, politics, economics,history, and everyday life. The length of each sentence is no more than 50 words as determinedby the parser we used.

Two hundred sentences were ones we had already analyzed for the study. They wereused for closed testing. The other 200 sentences were ones we had not seen before for opentesting (Table 1).

Given that test sentences were selected by our own researchers, and given that most ofthem were novel ones, we tried our best to make the test data more normative. However,it is not easy to tell exactly whether a linguistic expression is a metaphor or not. Differentpeople may hold varying opinions. All the metaphorical sentences were selected from ourChinese metaphor corpus, which was collected by three graduate students majoring in Lin-guistics. Separately, they selected metaphorical sentences from the DuZhe corpus accordingto the language definition of Goatly (1997) and the principle with which we identify activemetaphors in Section 2.1. Only metaphorical sentences selected by all of them are preserved.

TABLE 1. Number of Sentences in Test Data for Closed Testing and Open Testing

Metaphorical Normal Metaphorical Normalreferences references collocations collocations

Closed 50 50 50 50Open 50 50 50 50

Page 20: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

284 COMPUTATIONAL INTELLIGENCE

It is likewise not easy to say whether a novel metaphor is correct or not accordingto different opinions of different people. Thus, to evaluate a metaphor, it is better to useunderstandable/not understandable, acceptable/not acceptable, or appropriate/not appropriaterather than right/wrong. For metaphor recognition processing, there is no need to judgewhether the metaphor is appropriate or not, the aim is to recognize metaphors and extractthem. We leave the appropriation judgment to the meaning inference stage after metaphorrecognition.

We assume the dependency parsing results were 100% correct, thus the parsed depen-dency structures have been manually modified. Although the parser could not reach that highlevel of precision, this presupposition makes sure that we avoid parsing influence on ourmethod. Another presupposition is that the sentences under processing are considered to beformal, correct, and readable. Faulty sentences are excluded. Metaphor recognition shouldbe the work of selecting metaphorical expression from normal literal expression rather thanthe work of checking lexical or grammatical errors.

4.2. Experiment Results

The performance of metaphor recognition is mainly influenced by two main factors:precision of schema matching and validity of the values of unconventionality measurements.

Schema matching follows an embedded tree matching algorithm (Li 2007) which isan accurate rather than similar matching method to find all the class schemas that theinput sentence embeds. Thus, if the input sentence contains any of the 32 class schemas,a match will be found. The embedded matching algorithm is able to locate all the classschemas that were embedded in the target input schema. The result of schema matching wasonly constrained by the efficiency and adequacy of class schemas. However, the 32 classschemas covered the main metaphorical expressions and contributed to the precision findingof potential metaphorical units.

The validity of the values of measurements depends upon the computing strategy andan abundance of knowledge resources. HowNet and TongYiCi CiLin are adequate, efficient,and readily available Chinese knowledge resources.

Tables 2 and 3 show the performance on test data.Table 2 shows samples of values of referential unconventionality for referential pairs. We

used threshold T = 3.00. Table 3 shows samples of values of collocational unconventionalityfor collocated verbal pairs. We used threshold T = 3.30.

The thresholds are set based on the same experiment on a subset of testing corpus.The performance of unconventionality measuring is evaluated based on precision and

recall as defined below. Table 4 shows the results.

precision = number of correct recognized pairstotal number of recognized pairs

× 100%,

recall = number of correct recognized pairstotal number of actual pairs

× 100%,

F = 2 × precisionprecision + recall

.

From the results, we can see the efficiency of our metaphor recognition method. For thedifficult problem of metaphor recognition with novel metaphors, the average performance ofapproximately 65% is encouraging.

Page 21: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 285

TABLE 2. Part of Unconvf (w1, w2) Results (Threshold T = 3.00)

Referential metaphors Conventional referential expression

UnconvRef UnconvRefRank Target w1 Source w2 (w1, w2) Rank w1 w2 (w1, w2)

1 8.96 1 1.51(crabstick) (partner) (mother) (worker)

2 8.96 2 1.29(life) (book) (China) (nation)

3 9.28 3 1.00(eyebrow) (guard) (he) (gangster)

4 6.00 4 1.67(language) (mirror) (sun) (star)

5 7.92 5 1.00(time) (resource) (Beijing) (capital)

6 14.54 6 1.00(knowledge) (power) (you) (teacher)

7 4.79 7 2.25(lion) (king) (lion) (animal)

8 10.00 8 1.82(play ball) (battle) (computer) (tool)

9 7.25 9 1.62(stock) (bat) (sleep) (rest)

10 1.00 Error 10 7.25 Error(shadow) (picture) (jog) (exercise)

11 14.54(teacher) (engineer)

4.3. Discussion

CHMeta was able to distinguish different kinds of metaphors, and our methodologyprovides a measuring method for metaphor recognition.

The experiment results demonstrate the efficiency of CHMeta, for recognizingmetaphors. The success of CHMeta likely derives from the principle that the target andsource of metaphor belong to different categories. However, there still were a number oferrors. Most of the errors CHMeta made were due to measurement calculation, which isrestricted to computable knowledge resources and often inadequate definition and calcu-lation functions. By our definition, one concept is unconventionally referred to another ifeither of their categories is far from each other or they are hyponymy irrelevant. We treatedcategory and hyponym as main factors. Based upon the experiment results, the referentialunconventionality measurement needs further refinement. Resource-based computing alsolimits the accuracy.

CHMeta also failed on some of the conventional collocational metaphors because someconventional metaphorical collocations were included in the corpus data. For example,“ (save time)” which is considered to be derived from the root conceptual metaphor “Timeis money” according to Lakoff, failed to be recognized by CHMeta as a metaphor. In fact,this sort of metaphor is called a “dead metaphor” by many researchers because these types

Page 22: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

286 COMPUTATIONAL INTELLIGENCE

TABLE 3. Part of UnconvColl(wh, wd ) Results for VOB Collocations (Threshold T = 3.3)

Collocational metaphors Conventional collocational expressions

Head Dependent UnconvColl Head Dependent UnconvCollRank word wh word wd (wh, wd ) Rank word wh word wd (wh, wd )

1 32.00 1 1.82(knit) (dream) (knit) (basket)

2 3.64 2 1.00(read) (life) (read) (book)

3 14.54 3 1.00(catch) (singing) (cure) (wound)

4 20.00 4 1.36(damage) (healthy) (show) (landscape)

5 3.64 5 1.82(tear) (sky) (show) (elegance)

6 32.00 6 1.36(brewage) (darkness) (set up) (empire)

7 14.54 7 1.82(curve) (life) (lose) (goal)

8 3.64 8 1.00(open) (heart) (need) (food)

9 14.54 9 1.00(embrace) (earth) (feel) (despair)

10 1.09 Error 10 3.63 Error(chip) (ocean wave) (mine) (stone)

TABLE 4. Performance of Semantic Unconventionality Computation

Recall (%) Precision (%) F (%)

Metaphorical referential pairs (For closed testing) 76.19 80.56 78.26Metaphorical referential pairs (For open testing) 73 64.60 68.54Metaphorical collocations(verbal) (For closed testing) 68.68 70.60 69.63Metaphorical collocations(verbal) (For open testing) 55 63.22 58.82

of metaphors have already become daily language. For metaphor computation research, deadmetaphors ought to be avoided and concentration placed on new and novel metaphors.

In conclusion, the experiment results support the validity and efficiency of our pattern-based idea in accomplishing metaphor recognition.

5. RELATED WORK

Recent computational studies of metaphor can be categorized into three major approachesfor detecting metaphors: a rules-based approach, statistical approach, and a structure-basedapproach. Each approach has significant limitations that require further investigation.

Page 23: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 287

Representing the rule-based perspective, Wilks (1975) advances Preference SemanticsTheory while Fass (1991) forwards a Collative Semantics system to identify transitive verbalmetaphors such as my car drinks gasoline by hand-coded preference restriction rules (animal,drink, liquid). One limitation to this is that the preference restriction rules are fixed and lackadaptability. We would argue that a metaphor like The young man has drunk years of ink,which means “the young man has learned for years,” conforms to (animal, drink, liquid) andyet, according to Wilks and Fass, it is not recognized as a metaphor. But it is recognized asmetaphor by our recognition system. Another limitation of their approach is that it is ratherdifficult to write out all the selectional preference restriction rules with precision for all theverbs. In fact, metaphors are the results of human creativity, thus they are not easily fixed byrules.

Zhang’s (2003) logic inference and similarity computation method to recognize metaphoris based on finding exact similarity terms between two concepts. In his opinion, if termswhich represent the similarity of two concepts are found through logic inference, thenthe sentence is a metaphor. In that case, however, if the characteristic “sly” for lawyerand fox is not included in both of characteristic sets of the two concepts, the sentenceMy lawyer is a fox will not be recognized as a metaphor. As we have pointed out,metaphor recognition can be successfully done without inferring the true metaphoricalmeaning.

Rule-based and logic-based approaches work well on specific examples but the fixedrules are far from satisfactory for a large-scale approach for metaphor recognition.

Along with rule-based approaches, statistical methods have also been tried for metaphorrecognition. Mason (2002, 2004) presents a corpus-based approach to extract conven-tional metaphorical mappings between concepts by finding systematic variations in domain-specific selectional preferences, which are inferred from large, dynamically mined Inter-net corpora. Mixed use of domain keywords triggers recognition of verbal metaphors.For example, the verb dissolve belongs to the LAB domain while the noun company be-longs to the FINANCE domain, thus The company dissolved is extracted as metaphor.Metaphors considered in Mason’s method are simply-structured sentences without modi-fiers and the predicates and concepts are restricted to fairly specific domains. Thus, theCorMet system presented by Mason only detects higher-order conceptual metaphors by find-ing some of the sentences embodying some of the inter-concept mappings constituting themetaphor of interest, but cannot be a tool for reliably detecting all instances of a particularmetaphor.

Wang (2006) and Xu (2007) propose a Maximum Entropy (ME)-based model for Chinese“noun + noun” noun phrase metaphor recognition. The method functioned in the window(−2, +2) (e.g., the ocean of knowledge) but could do nothing for noun metaphorssuch as the following:

(Chinese).Jia1-ting2 cheng2-le ta1-men2 yi1-wei1-qi1-xi1, chu3-xu4-li4-liang4 de gang3-wan1 (Chinese

Pinyin).Family become their rest brew energy DE harbor (English

equivalents).Family becomes their resting and energy-brewing harbor (English translation).

The real topic, family, is located far from its vehicle, harbor. Our Metaphor Role De-pendency Schema Matching Theory addresses this problem.

Statistical approaches may function adequately for conventional metaphor recogni-tion. However, in addition to conventional metaphors that people are familiar with, novelmetaphors with low frequency, and fuzzy domain difference are created all the time. In com-bination with a variety of linguistic expressions, they are unavoidable factors in automatic

Page 24: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

288 COMPUTATIONAL INTELLIGENCE

metaphor recognition. Hence, recent statistical approaches have generally been limited indepth and extension as to metaphor recognition.

Martin (1990, 2000) and Agerri (2007a, 2007b) are representative of structure-basedapproaches. The basic idea of such approaches is to make use of invariant mappings oftarget and source in metaphors to interpret conventional metaphors. Kriahnakumaran andZhu (2007) propose algorithms to automatically classify sentences into metaphorical ornormal usages, which also make use of sentence structure. They classify metaphoricalexpressions into three main types: subject-object, verb-noun, and adjective-noun. Like us,they also produced good performance levels in metaphor identification. However, they do notfurther subclassify their types, nor supply a formalization strategy for metaphor expressions.The advantage of their work is that they detail several important problems with metaphorrecognition, especially word sense disambiguation, pronoun resolution, prepositional phrasemodifications for a noun, and conventional metaphorical usages that are already covered bylexical resources. Their work is constructive and highly suggestive, and will be of use in ourfuture work to improve our CHMeta system.

Besides Kriahnakumaran and Zhu’s computational classification of metaphorical ex-pressions, Yang et al. (2004) have also proposed a preliminary computational classificationsystem of Chinese metaphors according to the cognitive structure of topic, vehicle and theirsimilarities. The classification system includes nine categories. But ultimately this classifi-cation system has not been calculated to achieve metaphor computation.

Significantly, none of the above methods have considered the importance of furtheridentification of metaphor components, yet practical identification and tagging for metaphorroles is essential for further metaphor inference after recognition. While Sardinha (2006) haspresented a metaphor tagger, not a metaphor recognizer, they only identify words that are usedmetaphorically in corpora based on databases, such as a target database or vehicles database.Their system functions for simple referential metaphors, but their tagging performance levelonly functions at 30%. As we have discussed previously, topic terms, vehicle terms, andground terms of metaphor can be single words, phrases, clauses, or even larger units ofdiscourse. There are no semantic rules to indicate when the terms are words, phrases orclauses. Hence, if we take the terms of linguistic metaphorical elements as mathematic sets,such sets would be infinite and without any descriptive rules. In sum, setting up target orvehicle databases has severe limitations.

Compared with these approaches, our schematized idea overcomes many of their limita-tions. Its merit is that Metaphor Role Dependency Schema Matching Theory transformscomplex metaphor expressions into combinations of semantic dependency schemas ofmetaphor roles and converts metaphor recognition into pattern recognition. Unlike othermethods, the input testing sentences in our metaphor recognition system need not be re-stricted to a certain syntax structure. Not only is CHMeta able to locate potential semanticanomaly, but also able to quantify metaphorical anomaly, which is important for metaphoricalanomaly finding. As a result, CHMeta demonstrates a much more practical and applicablequality.

6. CONCLUSIONS AND FUTURE WORK

In this article, we have presented an integrated Metaphor Role Dependency SchemaTheory and a computational system CHMeta for Chinese metaphor recognition. At variancefrom systems that work for limited samples of metaphor, CHMeta, to a significant extent, isable to recognize metaphoric sentences in general.

Page 25: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 289

Metaphor Role Dependency Schema Matching Theory differs from other computationalapproaches to metaphor by employing psychological characteristics and pattern recognition.It uses two hypotheses for setting up the theory: A separability hypothesis to separatemetaphor recognition from metaphor inference and a patterning hypothesis to schematizemetaphorical units into small pieces of dependency schemas, which are used to extractlinguistic components for measurement calculation to find unconventionalities.

Chinese metaphorical units are classified into 32 class schemas, each of which has aspecial semantic dependency structure of metaphor roles and corresponding rules for mea-surement calculation and tagging. The metaphor recognition system is able to recognize ametaphorical unit and put it into a certain class, which provides more useful informationfor the metaphor inference process. Besides schematizing and matching to locate semanticanomaly, we also define, specify, and quantify semantic anomaly. The Metaphor Role Depen-dency Schema Theory and its implementation system contribute to metaphor computationresearch by its treatment of large-scale metaphorical expressions. Though the article delin-eates the theory from the perspective of Chinese, the idea and principles can be extended toother languages.

A limitation of CHMeta is that it is restricted to the sentence level. Hence, some difficultproblems like pronoun resolution which require context and discourse comprehension strate-gies present ongoing challenges. Our referential unconventionality calculation has engagedmetaphors with pronouns but has not produced a pronoun resolution module. In our futurework on metaphor recognition, we will incorporate these problems and attempt to advanceour metaphor recognition system from the sentence level to discourse level.

Metaphor recognition is important for its connection of natural language processing andmetaphor understanding. However, it is also an obstacle in metaphor computation not onlyfor its deviation from conventional semantic constraints, but also due to its various linguisticexpressions. Our work may be the initial integrated work in the computable classification,recognition, and tagging of Chinese metaphors computation. Overall, however, the workcould be improved to reduce limitations caused by knowledge resources-based computationstrategies.

This article delineates this theory from the perspective of Chinese. However, the ideaand principles can be extended to other languages. There still remains a great deal of workto do in metaphor computation. In our future work, we will concentrate on improvements inCHMeta system and in inference and interpretation work for recognized metaphors.

ACKNOWLEDGMENTS

This work was supported by the National Nature Science Foundation (NNSF) of Chinaunder grant No. 60373080. We would like to thank the HIT IR-Lab for sharing their corpus,resources, and Chinese dependency parser and the founders of HowNet for sharing their free2000 version. We would also like to thank our anonymous reviewers for their constructivesuggestions which helped improve this article and will assist our future research work.

REFERENCES

AGERRI, R., J. A. BARNDEN, M. G. LEE, and A. M. WALLINGTON. 2007a. Default inferences in metaphorinterpretation. In CONTEXT 2007, Lecture Notes on Artificial Intelligence Series (LNAI) 4635. Edited byB. Kokinov et al. Springer-Verlag, Heidelberg, pp. 1–14.

Page 26: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

290 COMPUTATIONAL INTELLIGENCE

AGERRI, R., J. A. BARNDEN, M. G. LEE, and A. M. WALLINGTON. 2007b. On the formalization of invariantmappings for metaphor interpretation. In Companion Proceedings of the Association of ComputationalLinguistics Conference (ACL-07), Prague.

DONG, Z. D., and Q. DONG. 1999. HowNet version2000, http://www.keenage.com/.

DONG, Z. D., and Q. DONG. 2006. HowNet and the Computation of Meaning. World Scientific Publishing Co.Pte.Ltd., Singapore.

FASS, D. 1991. Met∗: A method for discriminating metonymy and metaphor by computer. Computational Lin-guistics, 17(1):49–90.

FELLBAUM C. 1998. WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA.

GOATLY, A. 1997. The Language of Metaphors. Routledge, London.

HIT-IRLab. 2005. TongYiCi CiLin (Extended_Version), http://ir.hit.edu.cn/.

HIT-IRLab. 2006. HIT IRLab Shared Parser, http://ir.hit.edu.cn/.

KRIAHNAKUMARAN, S., and X. ZHU. 2007. Hunting elusive metaphors using lexical resources. In Proceed-ings of the Workshop on Computational Approaches to Figurative Language, Rochester, NY, pp. 13–20.

LAKOFF, G. 1993. The contemporary theory of metaphor. In Metaphor and Thought (2nd ed.). Edited by A.Ortony. Cambridge University Press, Cambridge, UK, pp. 202–251.

LAKOFF, G., and M. JOHNSON. 1980. Metaphors We Live By. University of Chicago Press, Chicago.

LAKOFF, G. et al. 1994. Conceptual Metaphor List, UC Berkeley, http://cogsci.berkeley.edu/lakoff/.

LI, J. F., Y. YANG, and C. L. ZHOU. 2007a. Corpus designing and constructing for metaphor computation. Mindand Computation, 1(1):142–146.

LI, J. F., Y. YANG, and C. L. ZHOU. 2007b. An embedded tree matching algorithm on metaphorical dependencysemantic structure extraction. In Proceedings of 2007 International Conference on Convergence InformationTechnology (ICCIT2007). IEEE Computer Society, pp. 607–611.

LIU, T., and J. S. MA. 2006. Chinese dependency parsing model based on lexical governing degree. Journal ofSoftware, 17(9):1876–1883.

MARTIN, J. H. 1990. A Computational Model of Metaphor Interpretation. Academic Press, San Diego, CA.

MARTIN, J. H. 2000. Representing UNIX domain metaphors. Artificial Intelligence Review, 14(4–5):377–401.

MASON, Z. 2002. Corpus-based metaphor extraction system, Ph.D.dissertation. Brandeis University, Waltham,MA.

MASON, Z. 2004. CorMet: A computational, corpus-based conventional metaphor extraction system. Computa-tional Linguistics, 30(1):23–44.

MEI, J. J. 1983. TongYiCi CiLin. Shanghai Lexicographical Publishing House, Shanghai.

ROBINSON, J. J. 1970. Dependency structures and transformational rules. Language, 46(2):259–285.

SARDINHA, B. T. 2006. A tagger for metaphors. In Paper presented at the Sixth Researching and ApplyingMetaphors (RAAM) Conference, Leeds University, 10–12 April.

WANG, Z. M. 2006. Chinese noun phrase metaphor recognition, computer software and theory, Ph.D. dissertation.School of Electronics Engineering and Computer Science, Peking University, Beijing.

WILKS, Y. 1975. A preferential pattern-seeking semantics for natural language inference. Artificial Intelligence,6(1):53–74.

XU, Y. 2007. Recognition of the Chinese metaphor phenomena based on the maximum entropy model. ComputerEngineering & Science, 29(4):95–103.

YANG, Y. et al. 2004. Research on machine understanding based classification of Chinese metaphor. Journal ofChinese Information Processing, 18(4):31–36.

YANG, Y., J. F. LI, and C. L. ZHOU. 2008. The identical natural of psychological process of metaphor recognition.Psychology Science, 31(5):1229–1231.

Page 27: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 291

ZHANG, W. 2003. Study on meta-anaphora resolution and metaphor comprehension in discourse understanding.Ph.D. dissertation. College of Computer Science and Technology, Zhejiang University, HangZhou.

ZHOU, C. L. 2003. Introduction to Mind Computation. Tsinghua University Press, Beijing.

ZHOU, C. L., Y. YANG, and X. X. HUANG. 2007. Computational mechanisms for metaphor in languages: A survey.Journal of Computer Science and Technology, 22(2):308–319.

APPENDIX A: THE INTERPRETATION OF TAGS AND THEIR MEANINGS

TABLE A1. Chinese Part-of-Speech Tags

Tag Meaning Tag Meaning Tag Meaning Tag Meaning

a adjective i idiom ni organization q quantifierb degree j short form nl location r pronounc conjunction k suffix ns site name u auxiliary wordd adverb m numeral nt time v verbe exclamation n noun nz proper noun wp punctuationg morpheme nd direction o onomatopoeia ws character stringh prefix nh name p preposition x nonmorphemevx copula vg content verb vb auxiliary verb vz modal verb

TABLE A2. Dependency Parsing Tags

Tag Meaning Tag Meaning Tag Meaning Tag Meaning

ATT Attributive VOB Verb–objective CNJ Conjunctive DI Chinese DI( )structure

QUN Quantity POB Preposition– MT Mood–tense DEI Chinese DEI( )objective structure

COO Coordinate SBV Subject–verb IS Independent BA Chinese BA ( )structure structure

APP Appositive SIM Similarity ADV Adverbial BEI Chinese BEI( )structure

LAD Left adjunct HED Head CMP Complement IC Independent clauseARD Right adjunct VV Verb–verb DE Chinese DE DC Dependent clause

structure

TABLE A3. Metaphor Role Tags

Role name Tag Role name Tag

Topic role Met_t Ground role Met_gVehicle role Met_v Mark role Met_m

Page 28: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

292 COMPUTATIONAL INTELLIGENCE

APPENDIX B: CATEGORIES OF REFERENTIAL METAPHORS

In the dependency figure of the example sentence, English word translations undercorresponding Chinese words are written in accordance with the original Chinese ones toreflect Chinese sentence constructions. Consequently, they are not necessarily correct Englishsentences. Full sentence translations of the Chinese examples are below the example figure.Met_t is the topic role and Met_v is the vehicle role.

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 29: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 293

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 30: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

294 COMPUTATIONAL INTELLIGENCE

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 31: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 295

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 32: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

296 COMPUTATIONAL INTELLIGENCE

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 33: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 297

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 34: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

298 COMPUTATIONAL INTELLIGENCE

APPENDIX B: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

Page 35: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 299

APPENDIX C: THE CATEGORIES OF COLLOCATIONAL METAPHOR

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 36: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

300 COMPUTATIONAL INTELLIGENCE

APPENDIX C: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

(Continued)

Page 37: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through

METAPHOR RECOGNITION 301

APPENDIX C: Continued

Description and MetaphorRole Dependency and

No. Its Class Schema Example

Page 38: METAPHOR RECOGNITION: CHMETA, A PATTERN-BASED …of metaphorical expressions) by providing a strategy to locate potential metaphorical anomalies in a target input sentence through