Introduction

1In this paper I will discuss different devices provided by INTEX to process semantic information. Semantic descriptions that go beyond the delaf1 categories might be useful for comparative linguistics, style analysis, machine translation or disambiguation. The semantic field that I will account for consists of French verbs denoting anger. The following semantic information will be taken into consideration:

2- number of arguments: in French, there are one-place verbs of anger (rager ‘to be angry’), two-place verbs (enrager de ‘to be angry about’, énerver “to make angry’) and even three-place verbs (irriter quelqu’un contre quelqu’un1 ‘to make somebody angry with somebody’).

3- semantic features of arguments: like all verbs of emotion, French verbs of anger select a Nhum, the human being who experiences an emotion. Two-place verbs select a second argument being the stimulus of the emotion. This argument is “non restricted” in the sense of Gross (1975). It can be realized as a noun phrase or as a sentential clause. Three-place verbs select a third argument denoting a human being (Nhum) who is the person the emotion is directed to.

4- causativity: Most, but not all French verbs of anger that appear in the construction N0 V N1hum (table 4 of Gross 1975) are causative. In French,2 it is common that the stimulus and the Nhum argument of these verbs can be inverted. The resulting pseudo-reflexive construction N0hum se V Prep N1 (s’irriter de lit.: ‘to irritate oneself about’) is non-causative. Two-place verbs of the N0hum V Prep N1 type (enrager de ‘to be angry about’) are non-causative, too. This is proved by the fact that enrager appears in a causativized construction faire enrager (‘to make angry’).3 French three-place verbs of anger are always causative.

5- agentivity of stimulus: Most, but not all French verbs of anger allow an agentive N0hum stimulus argument (‘to make angry on purpose’): e.g. énerver, scandalizer (‘to scandalize’). Another group does not allow an agentive stimulus: horripiler (‘to exasperate’) offusquer (‘to shock’) révolter (‘to revolt’), ulcérer (‘to embitter’).4

6- stage: according to Lakoff (1987: 397-400) the scenario of anger is divided into several stages. I will distinguish the stage of an offending event, i. e. the situation that produces anger, the stage of manifestation of anger, and the stage of performing an act of retribution. There are verbs that literally denote the offending event and by metonymy the feeling of anger itself. Some of the verbs listed by Mathieu (2000) as verbs of offense (e.g. froisser lit. ‘to crumple’) or verbs of disturbance (e.g. déranger, gêner ‘to disturb’) belong to this group.5 The act of retribution can be denoted by two-place verbs listed by Mathieu (2000) as verbs of anger or verbs of revolt: if the stimulus argument refers to a human being, this argument is also the person the anger is directed to (I call this semantic role « point of reference »). But the existence of three-place verbs shows that stimulus and point of reference can be realized by two different arguments: N0 irrite N1hum contre N2hum (‘N0 makes N1hum angry with N2hum’).

7- intensity: According to Mathieu (2000: 111), five levels of intensity should be distinguished: (déranger => énerver => exaspérer => scandaliser => révolter).

8- Source of metaphor or metonymy: For verbs of anger that have synchronically a figurative and a non figurative reading, the source of the figurative reading should be indicated in order to translate them, if possible, by a verb of the target language with a corresponding non figurative reading.

9- register: I will account for dia-systematic connotations distinguishing a formal register (more formal than standard language) and an informal register (colloquial, less formal than standard language).

10The utility of the semantic features illustrated in Table 1 has to be proved by further investigation. It might be doubtful whether it is necessary to indicate the source domain of a figurative reading. One may argue that different readings of polysemic words should be treated as logically independent. But according to Hilty (1995), the different readings are represented as a whole at the cognitive level. When one reading is actualized in a specific context, the others are not just suspended. They are co-present in form of connotations so they should be rendered, if possible, in an adequate translation. Nonetheless, a comparison of even closely related languages as French and Italian shows that the use of metaphorical and metonymical processes differs considerably (cf. Kailuweit 2001). The two languages do not draw anger readings from the same sources. For example, only French possesses verbs that can be tagged ‘dfigA4’ (in relation to an offending event of violating the integrity of hairdo) or ‘dfigA5’ (in relation to an offending event of violating the integrity of clothes): décoiffer (lit. ‘to dishevel’) or froisser (lit. ‘to crumble’). These verbs have to be translated with a non figurative Italian verb of offense (e.g. offendere).

1 – Processing with INTEX

11The following description of some INTEX devices to process the semantic features of table 1 is directed to inexperienced users. Processes will be described in detail. They can be easily reconstructed and may inspire linguists not yet familiar with INTEX to use the program’s different devices as a tool for their own research.

1.1 – Lexical Transducers

12The French verbs of anger irriter and énerver seem to be the less marked verbs of the semantic field. Of course, both have physical reading, too, but these readings have a low text frequency and are considered secondary by native speakers in comparison to the anger reading. Therefore, these verbs will be tagged +sent and +col, having all other features by default. The tagging of irriter is effected by a compiled lexical finite state transducer (FST) stored in the INTEX folder of dictionaries of simple words DELAF.

13The FST corresponding to Graph 1 searches the root irrit-combined with any type of ending. If we apply the FST to Stendhal’s novel le Rouge et le Noir, we get the following output:

irrita, irriter. V+sent+col
irritable, irriter. V+sent+col
irritait, irriter. V+sent+col
irritation, irriter. V+sent+col
irrite, irriter. V+sent+col
irrité, irriter. V+sent+col
irriter, irriter. V+sent+col
irritèrent, irriter. V+sent+col

14The FST has produced some noise tagging the adjective irritable and the noun irritation V+sent+col (verbs of anger). One possibility of avoiding such mistakes consists in using a filter dictionary FiltreCol-. Among the lexical resources of the DELAF folder, resources marked by ‘-‘ have a higher priority. They are applied prior to the other resources. Our filter dictionary FiltreCol- contains the standard (delafm) entries for irritable and irritation:

irritable, irritable. A+z1: ms: fs
irritation, irritation. N+z1: fs

15If we apply the FST corresponding to Graph 1 and the dictionary FiltreCol- at the same time, the dictionary filters the nouns, adjectives and adverbs derived from the same root irrit-. The output is now linguistically correct:

irrita, irriter. V+sent+col
irritable, irritable. A+z1: ms: fs
irritait, irriter. V+sent+col
irritation, irritation. N+z1: fs
irrite, irriter. V+sent+col
irrité, irriter. V+sent+col
irriter, irriter. V+sent+col
irritèrent, irriter. V+sent+col

16This simple tagging procedure has advantages, but also disadvantages. It is not the usual way of effecting lexical descriptions in INTEX (cf. Silberstein 1999-2000: chap.13). The process of creating an INTEX dictionary normally starts with constructing a dictionary of the DELAS type. A DELAS entry consists of a lemma, a part of speech category, a number referring to an inflexion type and a list of syntactic and semantic features. Inflection is effected automatically by a transducer corresponding to the inflection type number of the entry. The advantage of this method is evident: the resulting DELAF type dictionary is completely inflected and does not contain any noise. So there is no need to create an additional filter dictionary. On the other hand, the construction of graphs (or rational expressions) to inflect a verb lemma is intricate and takes up a lot of time. In addition, if one’s aim is semantic description, it is not absolutely necessary to unify semantic features as illustrated in Table 1 with syntactic features of the delafm type. Therefore, a lexical FST is a simple and efficient alternative to a DELAS dictionary.

17The graphs 2 and 3 correspond to FSTs that can be used to tag the occurrences of froisser and exaspérer:

18Froisser denotes the first stage of the anger scenario, the offending event, by referring to a violation of integrity of clothes. Therefore, it takes the tags +ph1 and +dfigA5. Exaspérer differs from irriter in having a higher degree of intensity. Hence, it takes the tag +i1.

19The three graphs are united in patron graph stored in the DELAF folder. Thus, they can be easily applied to a text at the same time. The resulting tagging can be used to realize different searching operations. Used in the locate pattern module, the regular expressions <V+ph1>, <V+col-ph1> or <V+i1> for example will find respectively all occurrences of verbs tagged as related to stage 1 of the anger scenario, occurrences of verbs of anger not related to stage 1 of the scenario, or occurrences of verbs tagged as having the degree of intensity 1.

20The three graphs are united in patron graph stored in the DELAF folder. Thus, they can be easily applied to a text at the same time. The resulting tagging can be used to realize different searching operations. Used in the locate pattern module, the regular expressions <V+ph1>, <V+col-ph1> or <V+i1> for example will find respectively all occurrences of verbs tagged as related to stage 1 of the anger scenario, occurrences of verbs of anger not related to stage 1 of the scenario, or occurrences of verbs tagged as having the degree of intensity 1.

1.2 – The problem of non causative constructions

21Following Croft (1993) and Pesetsky (1995), I consider the pseudo-reflexive construction of Romance verbs of emotion non-causative. Since French irriter is causative by default, we need a device that is able to recognize the pseudo-reflexive constructions of irriter and to tag these constructions as +ncau (non causative). I went through the FRANTEXT corpus (1950-2000) to determine the relevance of this problem. The corpus contains 542 occurrences of irriter.

22Graph 4 recognizes 93 of the 542 occurrences as pseudo-reflexive constructions. In a Lexicon-Grammar framework, the pseudo-reflexive construction might be considered a transformation. Some French verbs of anger appear in the pseudo-reflexive transformation others do not. Nonetheless, for the purpose of semantic description it is necessary to attribute the same tag +ncau to pseudo-reflexive constructions like s’irriter de and intransitive two-place verbs like enrager or rager. Since in current Italian intransitive two-place verbs of anger as imbestialire, inferocire or infuriare are not used, non causative French verbs like enrager should be translated by a pseudo-reflexive or by a support verbs construction: j’enrage (‘I am/become angry’) => mi arrabbio, mi viene rabbia.

23How can we treat this problem in INTEX? One possibility would be doubling the lexical entries for all occurrences of irriter. Two different transducers of the Graph 1 type could be used to tag every occurrence of irriter +ncau and –ncau at the same time. The FST representation of the sentence Paul s’irrite will be the following:

24In a second step, a compiled FST corresponding to Graph 4 could be stored in the folder DSISAMB to be applied in the disambiguation module.6 If we use the tagging mode ‘replace words with lexical entry’, ambiguities disappear, but the tags corresponding to the semantic features are maintained:

25An alternative way of treating the problem of non causative pseudo-reflexive constructions is the application of the function ‘Merge FST outputs with input text’ of the locate pattern module. Graph 7 inserts the tags +sent+col+ncau after the occurrences of the recognized pseudo-reflexive constructions of irriter.

26If we apply this graph to the sentence Paul s’irrite the result is the following modified text:

Paul s’irrite [V+sent+col+ncau].

27Nonetheless, there is an important disadvantage. Further research operations on the inserted tags can only be brought about after having copied, saved and recharged the modified text. But even then the inserted tags cannot be represented in a text FST.

1.3 – Processing of support verb constructions

28Since Italian support verb constructions like mi viene rabbia correspond to French non causative two-place verbs like enrager, it is necessary to tag theses constructions in the same way as simple verbs. From a general point of view, it is desirable to treat all kind of words that denote an emotion – verbs like irriter, nouns like irritation and adjectives like irritant or irritable – in an uniform and coherent way (cf. Blanco 1999, Gross 1995).

29In the following, I will refer to the French support verb construction être en colère (lit. ‘to be in anger’) to discuss two ways of processing support verb constructions in INTEX. It has to be taken into account that support verb constructions differ from compound words in allowing insertions between the support verb and the supported elements.

30The FST corresponding to Graph 8, a transducer with variables,7 recognizes the occurrences of être en colère with and without an inserted adverb.

31We will apply Graph 8 to the sentences Marie était bien en colère in the locate pattern module. If we use the ‘Replace recognized sequences’ mode for FST outputs, the FST recognizes an inserted adverb, effects a permutation to the right, reorders the constituents of the support verb construction and inserts the appropriate tags. Hence, it is possible to tag pseudo-reflexive constructions and support verb constructions in the same way:

Marie était en colère [V+sent+col+ncau] bien.

32The tags can be used for further searching operation, but only if we copy, save and recharge the modified text.

33An alternative would be applying a compiled modified version of Graph 8 as a lexical transducer for frozen expressions stored in the INTEX folder DELAE. Graph 9, that corresponds to the lexical transducer, differs from Graph 8 in deleting the inserted adverb instead of permuting it. In addition, the transducer does not only insert the appropriate tags, but produces a complete lexical description (lemma + category + semantic features).

34The advantage of processing support verb construction at the lexical level is that the lexical description is taken into account when the text FST (Graph 10) is constructed. However, if we apply a graph in the locate pattern module replacing or merging the text with the FST outputs modifications cannot be represented with a text FST.

35Nonetheless, there is still an important shortcoming of the lexical transducers for frozen expressions. The tags they provide are not taken into account in the locate pattern module. Hence, if we want to search for e.g. the occurrences tagged +ncau we will not find the support verb construction être en colère.

Conclusion

36INTEX provides several devices to realize the semantic description of French verbs of anger sketched in Table 1. The insertion of tags into the text using the merging or replacing mode of the locate pattern module is not really satisfying. The modified text has to be copied, saved and recharged, and even then it is not possible to represent the inserted tags in a text FST. Therefore, it is preferable to effect the semantic tagging at the level of lexical resources. This can be done combining lexical transducers for simple words, a filter dictionary and lexical transducers for frozen expressions. Ambiguities can be treated by local grammars of disambiguation. Lexical processing and disambiguation lead to an appropriate text FST representation. But, as far as research operations are concerned, it is inconvenient that tags inserted by lexical transducers of frozen expressions are not taken into account. This should be amended in a future version of INTEX.