A Contribution to the Study of the Semantics of French Psychological Verbs

Summary
In this document, we investigate the semantics of psychological verbs and verbs expressing feelings in French. After analysis, the lexical conceptual structure seems to be well-adapted because it allows us to represent the semantics of these verbs.
The study of sense variations of these verbs allows us to elaborate generalizations and to define related generative rules.
Introduction
The aim of this work is to propose a description of the semantics of psychological verbs and verbs that express feelings. By identifying sense components, by specifying sense elements which differentiate closely related verbs, and by studying sense variations, we found new criteria permitting the constitution of verbs semantic classes so as to better structure the lexicon and to better organize the semantic descriptions.
We are going first, to represent the verb semantics by means of the Lexical Conceptual Structure (LCS), adding if necessary new primitives and new semantic fields. We then study sense variations with the aim of making semantic descriptions more accurate and more comprehensive.
I. Verb semantic classes
Different approaches to semantic classifications have been proposed, each of them contributing to a different form of classification, whose usefulness and ease of constitution answer to different needs. There are different approaches to semantically classify verbs. Beth Levin [8]has proposed a classification based on syntactic criteria, which are alternations. But these alternations are specific to the English language. Another approach, more oriented towards NLP applications and for the French language, is proposed in [11]. It uses at the same time syntactic criteria and thematic roles.
Approaches based on ontological criteria assume that most, if not all lexical semantic relations, operate within quite precise semantic domains. In [2], verbs are first divided into stative verbs and action verbs, each of these group is then further divided into subclasses, according to some dimensions, properties or domains. In each family of verbs, a generic verb can be specified : act, move, become, make, be, ... One can consider that a generic verb corresponds to an LCS primitive
Jackendoff [3] worked out semantic classes by putting together verbs whose arguments accept the same thematic roles. This method, though based on semantic criteria, has the disadvantage of grouping verbs without semantic resemblances.
Different studies of the semantics of psychological verbs have been made [9], [2], [Sanfilippo 92]. In [9], verbs are classified according to the nature of the feeling. If a verb has more than one meaning, it belongs to as many corresponding classes. This approach is based on the identification of semantic regularities. We propose here a study of the semantics based on semantic criteria, by means of the Lexical Conceptual Structure (LCS).
II. The Lexical Conceptual Structure
The LCS is an elaborated form of semantic representation, with a strong cognitive dimension [4,5]. The LCS has some similarities with approaches used in Artificial Intelligence, such as semantic nets or conceptual graphs. The LCS is basically designed to represent the meaning of predicative elements and the semantics of propositions, it is therefore substantially different from frames and scripts. The LCS should be considered both as a semantic model providing a representational framework and a language of primitives on the one hand, and as a methodology on the other hand.
We use LCS to represent senses of psychological verbs and of verbs denoting feelings.
II.1) LCS representation
Conceptual categories and primitives of the LCS have been suggested for spatial localization verbs, other cognitive / semantic fields may be derived by analogy. We can also, if necessary, introduce new primitives. For example, « aimer » (to love) can be represented by the following pattern :
[event GO+psy ([thing FEELING==« amour »],
[path FROM+psy([thing X : LIVING_OBJECT]), TO+psy([thing/event ]) ])]
This description is nevertheless not satisfactory because there is not an actual transfer of feeling : someone who feels love is not devoid of love, although the goal perceives (or receives metaphorically) his feeling. This is why we introduce a new primitive FEEL which represents the fact that someone feels a feeling towards something or somebody else. The description becomes :
[event FEEL+psy ([thing X : LIVING_OBJECT ], [thing FEELING==« amour »],[path ]) ]
The representations of psychological verbs show that the set of primitives already known, enlarged with FEEL and SEEM ( e.g. to represent the verb « paraître » when it means seem) seems to be sufficiently expressive to describe those verbs.

II.2) LCS based classification
Some verbs share the same LCS pattern. For a given pattern, we have put together verbs which accept this model, and thus constitute a new class. Models allow us to classify verbs according to different levels of granularity and different conceptual dimensions.
For example, the following models accept :

[event ]
{ éprouver, ressentir, aimer, adorer, désirer, apprécier, envier, renoncer, hésiter, décider, regretter, se désoler, ...}
[state ]
{ rire, pleurer, déprimer, délirer, supporter, apprécier, ... }
[event CAUSE([ ]), [event INCH( [state BE+psy ([thing X : LIVING_OBJECT], [place ])]) ]]
{ supporter, inquiéter, terroriser, tourmenter, ... }
etc ...
We note that verbs such as inquiéter, tourmenter, terroriser, ... belong to the same semantic class, but the LCS pattern is not sufficient to take into account the level of anxiety which differentiate these verbs. The non-branching proportional series [1] [6]allow us to establish this gradation. For instance, the scale for these verbs is :

This enable us to complement the LCS representations. Terroriser can then be described by both the previous serie and by the following LCS pattern :
[event CAUSE([ ]), [event INCH([state BE+psy ([thing X : LIVING_OBJECT ],
[place AT+psy ([prop INQUIETUDE) ])]) ]]

We notice however that verbs of a given non-branching proportional serie do not necessarily belong to the same semantic class (in the sense that they « accept » the same LCS pattern). We can improve the semantic description by the study of sense variations, and show how generative actions can be represented with the LCS.
III. The Generative framework
To describe the senses of a verb, we can either adopt an enumerative approach where senses are close to each other, or create a system of rules which generates the new senses of a verb. This second approach is interesting since it allows us to represent the dynamics of a language [10].
Our work consists in :
investigating the notion of sense, and developping a generic model for each sense, without falling into an excessive generalization. Our goal is to define the different senses and usages of a verb, and to adjust the different usages around a small number of senses, from which a generative process can produce derived usages.
using a sense theory, such as LCS, and showing how the lexical semantic generative operations can be described with the LCS.
III.1) Hypothesis on sense representation
A study of sense variations of psychological verbs and verbs denoting feelings shows that :
a) LCS representations are different for 2 senses of a verb supporter in the sense of « to offer resistance to an agression » :

Jean supporte le froid
[event STAY+psy,+char,+ident ([thing X: LIVING_OBJECT],
[place AT+psy,+char,+ident ([prop PSYCHOLOGICAL_STATUS(X)==« résistant » AND PSYSICAL_STATUS(X) == «
résistant » ]) ]) ,
DESPITE ([thing/event Y : ELEMENT_AGRESSIF ]) ]
supporter in the sense of « to take upon » : Jean supporte la responsabilité de sa fonction

[state BE+psy ([thing X : LIVING_OBJECT ], [place AT+psy ([prop PSYCHOLOGICAL_STATUS(X)==« acceptation »
]) ]),
[BECAUSE_OF( [event GO+psy ([thing ],
[path FROM([thing ])])]]
supporter in the sense of « to run a software»
[event CAUSE([thing I: ORDINATEUR]), [state BE+char,+ident ([thing LOGICIEL ], [place AT+char,+ident ([prop
STATUS==« exécutable » ]) ]) ,
BY_MEANS_OF ([thing CONFIGURATION(I) ]) ] ]

b) for a given sense, the LCS pattern kernel can be more or less instanciated, depending on arguments.
III.2) Operations related to sense variations
The usage analysis is useful for delimiting senses. Among the most frequent usage relations, we have selection, metonymy, metaphor, co-composition [7, 11]. A study of the behaviour of these generative operations shows that they would tolerate the rules described below.
a) selection : the type of an argument is subsumed by the type expected by the predicate for that argument. For example, exciter may have for its second argument a function, or chemical elements : « exciter la soif », « exciter des ions »
exciter une fonction or exciter des ions :

[event CAUSE([thing/event ]),[event GO+char,+ident ([thing X : LIVING_OBJECT ??X : FUNCTION ? X :
CHEMICAL_ELEMENTS ],
[path FROM+char,+ident ([prop VALUE_OF(ACTIVITY_LEVEL(X) == « low » ]) ,
TO+char,+ident ([prop VALUE_OF(ACTIVITY_LEVEL(X) == « high » ]) ]) ] ]

selection may be considered if we define a constraint on the conceptual category instanciation.
b) metonymy : in the argument, an entity is used to refer to another one, via lexical relations (is-a, part-of, ...). We try first to resolve the metonymy, and then describe the sentence sense, by means of a lexical relation.
In la bourse se calme (literally, the stock exchange quiets down), stock exchange may refer to brokers and/or to currencies.The PART_OF lexical relation denotes the different entities of the stock exchange (brokers are taking into account with +psy, and currencies with +char, +ident).
[event FEEL+psy,+char,+ident ([thing WHOLE_OF(X) ], [thing FEELING==« calme »],
[path FROM+char,+ident,+psy([prop ACTIVITY_LEVEL(X)) ==« high » ]) ,
TO+char,+ident,+psy ([prop ACTIVITY_LEVEL(X)) ==« low » ]) ])]

c) metaphor : it is generally a partial homomorphism between ontologies of different conceptual domains. supporter can be used in the sense of « to hold », but also in the sense of « to encourage ». To encourage somebody can mean to « sustain » his courage in a high position.
supporter in the meaning of « to hold » :

[event CAUSE([thing X ] ), [event STAY+loc,+cont ([thing ]j, [place ON+loc,+cont ([thing X ] ) ]) ] ]

supporter in the meaning of « to encourage » (psychological holding) :

[event CAUSE([thing X] ), [event STAY+psy ([thing ] , [place AT+psy ([thing X ] ) ]) ] ]
Metaphors may be characterized by a switch in the semantic fields, w.r.t. this verbs class.
d) co-composition : deals with the emergence of unexpected new senses, often with an important sense variation. This phenomenom usually may be at the origin of semi-fixed or fixed forms. We can say fatiguer la terre for remuer la terre, whereas fatiguer le sable has no meaning.

fatiguer la terre in the sense of « to move earth »
[event CAUSE([thing/event ]) , INCH ([state BE+char,+ident ([thing EARTH ],
[place AT+char,+ident ([prop PHYSICAL_STATUS == « remué » ] ) ]) ]) ]

co-composition may correspond to closed LCS structures : all the conceptual categories are instanciated.
e) sense restrictions may be represented by selectional restrictions : paraître in the sense of « to appear before a court of justice » :

[event GO+loc ([thing X : LIVING_OBJECT
[path AT+loc ([place IN_FRONT_OF+loc ([thing TRIBUNAL ] ]) ]) ]) ]

paraître in the sense of « to appear on the scene » :

[event GO+loc ([thing X : LIVING_OBJECT], [path AT+loc ([place ON+loc ([thing SCENE ]) ]) ]) ]
It is now interesting to show some generalizations over sets of verbs w.r.t. sense variations, which can be expressed by means of rewriting rules. .
An example is the semantics of a verb which denotes some physical characteristics. It can be extended to make the semantics of this verb denotes psychological characteristics. This is the case of verbs like to:
{ inquiéter, fatiguer, supporter }
which accept the same metaphor, consequently the same ontological domain changes from the physical characteristics to psychological ones(+char,+ident --> +psy)
{ supporter, inspirer }
also accept a metaphor, from a physical motion to a psychological one (+loc --> +psy)
This work tends to show that sense variation phenomenoms can be described according to LCS alterations from the original sense of the verbs. This changes from the analysis of sense variation based on a type concordance and the operation of type coercion.
Concerning the verbs semantic classes, we can add orthogonally the classification criteria mentionned above in section II, to the criteria of sense variations.
Conclusion
In this work, we propose a representation for psychological verbs, by means of the LCS. We have also shown that some generative operations can be applied with some regularities on this verb family. The definition of these representations makes it possible to build new semantic classes.
References
1. A. Cruse Lexical Semantics C.U.P. 1986
2. C. Fellbaum « English verbs as semantic net » Journal of lexicography - 1993
3. R. S. Jackendoff « Semantic Interpretation in generative grammar » The MIT Press - 1972
4. R. S. Jackendoff « Semantics and Cognition » Cambridge - Mass - The MIT Press - 1983
5. R.S. Jackendoff « Semantic Structures » Cambridge - Mass - The MIT Press - 1990
6. M. Kamel, P. Saint-Dizier « La relation d'opposition selon des séries proportionnelles sans branchement »Terminologie et Intelligence Artificielle - Toulouse 1997
7. G. Lakoff, M. Johnson « Metaphors we live by » University of Chicago Press - 1980
8. Beth Levin « English verb classes and alternations : a preliminary investigation » Chicago - The University ofChicago Press - 1993
9. Y.Mathieu "Verbes Psychologiques et Interprétation Sémantique" Langue française - ED. Larousse Feb. 95 -n°105
10. J. Pustejovsky « The Generative Lexicon » MIT Press - 1995
11. P. Saint-Dizier « Verb semantic classes in French » Proc. Coling'96 - Copenhaguen