Jean-Pierre Koenig’s current research program (09/01/08)

Speakers of a language know tens of thousands of words. A survey I conducted of college-educated English speakers suggests they know about 4,000 verbs (Koenig et al. 2003). The sheer size of one's lexicon raises several questions:

· What information is associated with all these words?

· How is it organized?

· How is it used when interpreting sentences?

My research focuses on these questions. During my graduate studies and the early part of my career, I focused on the organization of lexical knowledge. In particular, I developed a model that accounts for the way word structure is similar to, but not identical to, the structure of sentences, i.e., words may have the creativity and productivity of syntax, but they can also be idiosyncratic (see Koenig (1999) for a summary). I also showed that this model accounts for the kinds of regularity that languages display between what words mean and the contexts in which they occur (see Davis and Koenig (2000) and Koenig and Davis (2001)).

In more recent years, while I have continued the investigation of general principles that explain how the meaning of words affects the contexts in which they occur, my research has expanded in three directions.

The first concerns the informational boundaries of word meaning (how complex verb meaning can be and what verb meaning can include). The second concerns the extent to which word meaning and syntactic structure mirror each other within and across languages. The third concerns the role semantic information plays in the on-line interpretation of sentences.

1. The cartography of verb meaning

Given the results of the survey I just mentioned and that, on average, each English verb has between three and four meanings (depending on the particular dictionary), speakers know between 12,000 and 16,000 verb meanings. What is the meaning of all these verbs made up of? This is one of the questions my recent research has tried to answer.

How do we distinguish information that is part and is not part of the meaning of verbs? Verbs describe situations; those situations include participants who play certain roles (e.g., in an eating event, there is an entity acting to ingest some food that is then consumed). Linguists have traditionally distinguished information about situation participants that is part of the meaning of verbs/is strongly associated with verbs and information about situation participants that is not/is weakly associated with verbs. Roles that include the former kind of information are called arguments; roles that include the latter kind of information, adjuncts. Koenig et al. (2003) proposed a new model of the difference between these two kinds of semantic information (participant roles). Arguments are participant roles that are (i) required by the situations described by verbs and (ii) are specific to these verbs and a restricted set of verbs (e.g., all situations occur somewhere, but only some situations involve an entity that causes a change of state in another entity). The relevance of the specificity of participant information to its strength of association with situation-types, I claim, is the verbal parallel of the special role that feature distinctiveness has been shown to play in object categorization. A more information-theoretic way of conceptualizing the model is that participant role information is associated with the meaning of verbs to the degree that it co-occurs with the situations the verb described more than expected by chance (so-called (point-wise) mutual information between the verb's denoted situation-type and the participant role). The research I just outlined provides a model of the factors that determine what information is part of, or is strongly associated with, a verb's meaning.

How complex can the structure of verb meaning be? There seems to be a limit to the structural complexity of verb meanings. Intuitively, verb meanings typically describe situations that are no more complex than a cause leading to a change of state. Thus, there are many words like kill (an activity that causes the death (change of state) of an entity), but there appears to be no single word that would mean something like pay (someone) to kill (an activity that causes another entity to act so that a change of state occurs). In Koenig et al. (2008), we showed on the basis of a survey of 1,800 verbs that verb meaning can be more complex than hitherto believed when a verb describes situations that involve tool manipulation, i.e. situations in which an agent uses an instrument to perform an action (a typical hominoid behavior).

What kind of information differentiates verbs which share meaning structure? Coarse-grained semantic structure (whether the described situations are a state, or include a cause, a change of state, or the use of a tool) leads to conceptual classes that include hundreds of verb meanings. Out of the myriads of ways situation-types denoted by verbs in each of these classes could vary, what are the ways in which verbs typically vary? In other words, why do we have the verbs we have in each of these classes?

In contrast to research on the overall structure of verb meaning, little research has been devoted to this question. To delineate the kind of non-structural information verb meanings can include, Koenig et al. (2008) semantically classified the meaning of all the verbs that describe situations that must or may include the use of a tool to induce a change of state (about 1,800 in total). This study showed that these verbs fall into a small set of semantic classes and that the idiosyncratic information distinguishing one verb meaning from the other is not randomly distributed. Some aspects of the meaning of verbs that describe situations that must or may include use of a toll are more finely carved than others in ways that parallel research on goal/result focus in adults' and children's event descriptions.

How are changes of state categorized differently across languages? In several South Asian and South-East Asian languages (Hindi, Tamil, Thai) one can felicitously say the equivalent of English He killed Lisi, but she didn't die. In recent work, my student Liancheng Chief and I (Koenig and Chief (2008) have argued that such apparently odd verb meanings can be reduced to a variation on the semantics of change of state verbs in better known languages. Simply put, these languages conceptualize the described situations as involving a change that is non-null, but less or equal to the maximum (i.e. paraphraseable for Mandarin sha ‘kill’ as ‘hurt in a way that the change in vital signs is less or equal to the maximum, i.e. death’). We further show that only the meaning of pairs of verbs which describe changes of state that can be conceptualized as gradable (changes in health, persuasion, …) vary between these languages and languages like English.

2. Non-uniformity in the syntax/semantics interface

The structure of expressions in logical or artificial languages “mirrors” the structure of the meaning they express (technically, one can define a homomorphism between a syntactic and semantic algebra). This parallel between the syntax and semantics of natural languages is only partially surface true (i.e. true of the apparent structure of sentences). Much of the research on the interface between syntax and semantics over the last quarter century has studied the extent of this “imperfection” in natural languages and proposed explanations for it (see Koenig (2005) for a survey). To advance our understanding of that issue, I have studied one particular semantic field (aspect operators) and how it is realized syntactically. Aspect operators in natural languages are expressions (typically verbs or verb forms) that indicate whether an event is completed or merely stopped, whether an event is on-going, whether an event has consequences for the present (or some other reference interval).

One dominant view is that the syntax of aspect markers is uniform across languages, that there is a single, universal mapping between aspectual operators and syntactic positions. In Koenig and Muansuwan (2005), original data from Thai challenges this view. Briefly put, the seventeen Thai aspect markers are all verbs, but they fall into two distinct classes. Members of the first class are syntactic heads that take following verb phrases as complements; members of the second class are syntactic modifiers that modify preceding verb phrases. We show that the fact that one and the same language has two (symmetric) ways to map aspect operators onto syntactic structure seriously undermines the uniformity hypothesis. Our research supports claims that the architecture of natural languages allows for a dissociation between semantic operators and syntactic heads. Despite their common semantic operator status, Thai aspect markers may, but need not, be syntactic heads. In my most recent work (presented at conferences, but not yet published), Poornima Shakthi and I have extended this argument on the basis of data from Hindi. Hindi aspect markers combine with lexical verbs to form complex predicates. Most interestingly, Hindi shows the same “dual” mapping between syntax and semantics that Thai does. Some of the aspect operators are syntactic heads that take preceding verbs as complements, and some are syntactic modifiers that modify following verbs. (Hindi is a verb final language, i.e. heads follow their complements.) Data from Hindi support the same conclusion regarding the architecture of the syntax/semantics interface as the Thai data supports.

Some of my recent research has also expanded on the semantics of aspect operators. Koenig and Muansuwan (2001) show how semantic operators that select non-necessarily proper subparts of an event can be used to model the meaning of Thai aspect classes. Nishiyama and Koenig (2008) show that these kinds of operators can be used to provide a new approach to the meaning of the English and Japanese perfect. They argue that the various traditional interpretations of the perfect come from further specification of an underspecified meaning, typically through pragmatic inferences and validate their hypothesis through a corpus study of a ps-random sample of over 600 English and Japanese example discourses.

3. How much of the syntactic context in which words occur is truly semantically determined?

It is well-known that the meaning of words partially determines how participant roles that are included in the meaning of verbs are realized syntactically. In the last few years, expanding on Koenig and Davis (2001), I have re-examined to what degree the meaning of verbs determines the realization of arguments. In Koenig and Davis (2003) and Koenig and Davis (2006), we argue that previous research, including our own, has overestimated the role of semantics by surreptitiously positing unwarranted semantic representations. Koenig and Davis (2003) and Koenig and Davis (2006) also use semantic underspecification to properly bound the effects of verb meaning on the syntactic realization of participant roles.

Most recently, in joint research with Karin Michelson (Koenig and Michelson (To appear)), I have looked at the interplay of nominal reference and argument realization. Our analysis of Oneida kin terms (an Iroquoian language), which are at the same time nominal and verbal stems suggests that (i) the issue of nominal reference (nominal index selection, technically) is in principle distinct from the issue of the realization of the members of kin relations, (ii) that the range of properties relevant to argument realization may include non-event-dependent properties (e.g., the absolute age of members of a kin relation), and (iii) it seems to be indeed a universal that only n-1 of a noun's arguments need be realized by phrases co-occurring with the noun. Oneida kin noun stems only seem to be an exception to this putative universal because they are also verb stems.

4. The influence of lexical meaning on sentence processing

Over the last ten years, my research has combined theoretical linguistics, quantitative linguistics, and experimental methodologies to answer the fundamental questions about speakers' word knowledge I have just described. In most of my experimental work to date, I have tried to determine what kind of participant information is used in the on-line interpretation of sentences and when it is used. In a series of papers from 1999 through 2002, Gail Mauner and I showed that whether a semantic argument is activated or not upon reading a verb form (i.e. whether readers of sentences activate the notion of an agent/cause …) depends on whether or not the semantic argument is syntactically active (is part of the verb's argument structure in Linguistics parlance). Critically, the activation of the semantic argument does not depend on the realization of that argument, since it was unexpressed in all our critical conditions. More importantly, we have experimentally validated our model of participant role activation through a series of experiments (see Koenig et al. (2003), Conklin et al. (2004), Koenig et al. (Under revision)). In all these experiments, extensive corpus work showed that the effects of strength of association between situation-types and participant roles was not due to extraneous syntactic factors, in particular how frequently that role is expressed.

In recent work (experiments conducted, presented at conferences and being written up for publications), we have expanded this research based on the results of the semantic results of the survey of verbs that require or allow instruments detailed in Koenig et al. (2008). Verbs that require an instrument role tend to constrain the range of fillers of that role more than verbs that allow an instrument (compare the range of likely instruments for chop and injure). In a series of experiments conducted by Breton Bienvenue, we showed that this subtle difference in the event information verbs include can affect eye-movements: Participants launch earlier looks for highly constraining verbs and use these semantic constraints to infer which of a set of possible depicted objects are likely instruments of the verb. Gail Mauner and I have also shown that the fact that a syntactically similar preceding sentence contains a verb that belongs to the same narrow semantic class (e.g. membership in the class of cutting cuts), can facilitate first and later inhibit the processing of the verb and post-verb regions of an immediately following sentence. Both results provide support for the semantic classification discussed in Koenig et al. (2008).

Finally, a former student of mine and I have shown that part-of-speech information associated with preceding words can affect the lexical processing of subsequent words above and beyond any effect of semantic similarity between the two words (see Melinger and Koenig (2007)).