License to Coll: How to Bind Bound Words and Readings to their Contexts

In this contribution we propose a new module for handling idioms and distributional idiosyncrasies. Based on the concept by Richter/Sailer (1999) the new feature COLL (context of lexical licensing) plays the central role in our approach. We provide a way to handle decomposable and nondecomposable idioms and idioms containing bound words. Our module guarantees the co-occurrence of all idiom parts and of bound word and licensing context, respectively. A prerequisite for our analysis is a means to select for particular elements in the lexicon. We introduce another feature, LISTEME, which gives each lexical item its unique identifier and makes it possible to select for a particular lexical word or phrase. Finally, we compare our proposal with alternative approaches and give some ideas regarding further applications beyond idiomaticity. 1 Motivation Idioms are omnipresent in everyday language, enriching our communication with metaphoric imagery and fulfilling various communicative goals. Nonetheless, they have been widely neglected by linguists developing grammar fragments. And even where an account for idioms has been given, most approaches have their shortcomings (cf. Riehemann, 2001, ch. 4). In this contribution we want to focus on decomposable and non-decomposable idioms1 and idioms containing bound words. We concentrate on technical aspects of the analysis and refrain from presenting detailed linguistic corpus data due to space limitations. By “idiom” we mean idiomatic expressions that do not form complete sentences as would be the case for e. g. His bark is worse than his bite. (1) make waves (“cause trouble”) (2) spill the beans (“divulge a secret”) The expressions in (1) and (2) are instances of decomposable idioms, i. e. their meaning can be derived from the idiom parts. Note that idiom parts are not necessarily to be understood literally. In (1), e. g., we can attribute the meaning “cause” to make and “trouble” to waves. The idiomatic meaning of the whole idiom consists of the idiomatic meanings of its parts. Where this is not the case, an idiom is non-decomposable: the meaning of the whole phrase has nothing to do with the meaning of the words the idiom consists of. Consider (3) and (4): The research to the paper was funded by the Deutsche Forschungsgemeinschaft. I am grateful to Stefan Müller, Christine Römer, Manfred Sailer, Adrian Simpson, the reviewers and the audience of HPSG’04 for comments and Michelle Wibraham for help with English. Cf. Gibbs et al. (1989) or Nunberg et al. (1994) for this distinction. 262 (3) saw logs (“snore”) (4) shoot the breeze (“chat”) It is not clear how to assign the meaning “snore” to the words saw and logs, the same holds for “chat”. Finally, we want to draw the attention to idioms comprising bound words or “cranberry words” (Aronoff, 1976). These are expressions which are highly collocationally restricted. Dobrovol’skij (1988) compiled quite a lot of examples for German, Dutch and English. (5) to learn/do sth. by rote (automatically, by heart) (6) to cock a snook (to thumb the nose) The underlined words are restricted to the given contexts. Sometimes there is some variation, as in to lie/go/lay doggo (Brit. slang; “to hide oneself”), but a free distribution is not possible. Such idioms can be either decomposable or nondecomposable. 2 Lexemes and Listemes Before we present our analysis, we point out a way that enables us to select a specific word. This forms a prerequisite of our approach. Idioms often consist of particular words which cannot be substituted by semantically equivalent terms. It seems in general that each word has a unique “identity” with an idiosyncratic behavior. The possibility to select a particular word would, thus, be a useful feature. Up to now, there has been a discussion about the necessity of having such kind of selection. One could argue that any data in question are to be handled as Constructions or collocations. But why impose such a “heavy thing” on an expression like to furrow one’s brow? Would it not be plausible that the verb furrow simply selects a word of the form brow? For perfect tense in German a main verb has to be combined with the right auxiliary (haben/sein; in HPSG with the attribute AUXF, cf. Heinz and Matiasek, 1994, p. 222). Here one does nothing other than to select a particular lexeme. A mechanism for selection of lexical elements has to meet three requirements: 1. The information has to be locally available (below SYNSEM). 2. The information has to be available along the syntactic projection line. 3. The information must be identical for a pronoun and its antecendent. Krenn and Erbach (1994) made an important contribution to idiom analysis within the HPSG framework. They suggested selecting particular lexemes via their feature LEXEME below CONTENT INDEX. This idea of having lexeme information 263 in the CONTENT is questionable. A lexeme combines phonetic, morphological, syntactic and semantic properties all together, not only semantic information. Besides, their approach had several technical shortcomings (cf. Soehn and Sailer, 2003): the locality principle for selection (Pollard and Sag, 1994, p. 23) was not implemented and there was not means for the LEXEME value to percolate. We therefore propose that the LEXEME approach has to be discarded. A different concept that helps to distinguish between individual words is that of a listeme2. As the concept holds the characteristic of listedness in a lexicon, we use it in our grammar to identify a particular word or phrase. Thus, we insert LISTEME into the feature geometry below CATEGORY, emphasizing the morphosyntactic character of information. More precisely, we put it below HEAD. This has two consequences: firstly, it is available for selection, as a HEAD value is below SYNSEM. Secondly, the LISTEME value of a projection is the same as the one of the head, as all HEAD features “percolate” according to the HEAD-FEATUREPRINCIPLE. For our furrow-example that means that a modified direct object his heavy brow still has the same LISTEME value as brow alone. A third question to address is the handling of pronominalization. It is necessary that pronouns have the same LISTEME value as their antecedent.3 In Krenn and Erbach’s approach this was the major motivation of putting the LEXEME feature in the INDEX. To emulate this quality, we propose a constraint ensuring that each pronoun which is co-indexed with an antecedent takes over its LISTEME value. In the lexical entries of pronouns that value would be left underspecified in that way, that it consists of a disjunction of an identifying value (she, her, etc.) and a wildcard. In case of co-indexation the wildcard is identical to the LISTEME value of the antecedent and – by virtue of the constraint – becomes the actual and concrete LISTEME value of the pronoun. An informal description of such a pronoun constraint is illustrated in (7). (7) PRONOUN-LISTEME-CONSTRAINT: If a pronoun is co-indexed with an antecedent, it takes over the LISTEME value of that antecedent. Otherwise the LISTEME value of this pronoun is that of the other disjunct. The value of LISTEME is an atomic sort as brow, heavy, furrow, take, she etc. In order to identify listemes for the same words having different meanings, we use numeric indices just as in a dictionary. In summary, discarding the LEXEME approach, we propose a more adequate solution for the problem of selecting particular words, at least with respect to terminology, technical feasibility and the feature geometry. We introduce a feature LISTEME which is appropriate for the sort head taking atomic sorts as its value. This term has been introduced by (Di Sciullo and Williams, 1988, p. 1) for a sign that is listed in the lexicon. E. g. in the phrase “He furrowed it.” the pronoun has the same LISTEME value as its antecedent, satisfying the subcategorizational requirement of the verb. 264 3 Licensing Contexts Getting to the analysis, we have to define a second attribute in the feature geometry. We declare objects of sort sign to bear a list-valued feature COLL (Context Of Lexical Licensing), first introduced by Richter and Sailer (1999). The COLL list may contain objects of sort barrier. These barriers are particular nodes in the syntactic configuration, like XPs, complete clauses or utterances (a complete clause with an illocutionary force). The concept of barriers is borrowed from the tradition of generative grammar, where these form boundaries for government and binding principles. We avail ourselves of this concept and use similar barriers to restrict the range of influence of the COLL feature. barrier objects have an attribute LOCAL-LICENSER (LOC-LIC) which has a value of sort local. In the lexical entry of an idiomatic word one can thus specify a barrier on its COLL list with a specific local configuration. Subsorts of barrier are illustrated in figure 1: complete-clause, utterance, np, vp and pp. The subsorts of barrier correspond to nodes in the syntactic tree with particular properties. The relations depicted in figure 2 identify the nodes which relate to the subsorts of barrier.4 complete-clause utterance np vp pp xp [ barrier LOC-LIC local ] Figure 1: Sort hierarchy for barrier The LICENSING-PRINCIPLE (informally in 8) makes sure that if there is a barrier specified on a word’s COLL list, there is an actual barrier in the phrase our word occurs in. This barrier must fulfill the local requirements and it has to be minimal, i. e., there is no other potential barrier of the same kind between the word and the actual barrier. (8) LICENSING-PRINCIPLE (LIP): For each barrier object on the COLL list of a sign x and for each phrase z: the LOCAL value of z is identical with the LOC-LIC value, iff z dominates x, z can be identified as the barrier specified and z dominates no sign y which in turn dominates x and forms an equivalent barrier. In this principle there are three conditions to be satisfied. The first one is simply that there must be a domination relation between the phrase z and the idiomatic element x. The second condition, that a particular barrier can be identified, means Cf. (Richter, 1997, pp. 68f) for the STATUS feature. 265