Beyond adding complexity to the model, a disadvantage of this approach is that it could happen that a speaker that first uttered a Hungarian utterance speaks Quechua in the next, so to speak. Let D be the set of possible data that learners may be exposed to. As k increases, learners have more data to base their inference on and so tend to recover the true types that generated a given sequence with higher probability.
The mutation matrix Q of the replicator mutator dynamic in 1 can then be defined as follows: is the probability that a learner acquires type i when learning from an agent of type j. For a given observation d , the probability of acquiring type i is , so that:. Communicative success and learnability is central to the cultural evolution of language.
These components can be modelled, respectively, as replication based on a measure of fitness in terms of communicative efficiency and iterated Bayesian learning. The grey dotted square in Fig. Populations at the corners of this space are monomorphic; all agents in them are of a single type. Points away from the corners represent mixed populations of various types. But it could equally well be a population that has all four types in equal proportions. This level of granularity will nevertheless be sufficient for our purposes, as we are only interested in sketching rough trajectories after replicator or mutator steps.
When it comes to communicative success, types with have a functional advantage over ; literal has a slight advantage over pragmatic ; all have a sizable advantage over literal. The red arrows in Fig. If instead the prior favors over , we expect repeated mutator steps to drive the population toward the upper half of the space. This is suggested by the solid purple arrows in Fig. Despite its simplicity, the diagram in Fig. Pragmatic language use will not evolve under strong and fully expressive semantics bottom part of Fig.
A weak semantic convention will not evolve without disposition toward pragmatic enrichments left part of Fig. But together a weak semantics and a pragmatic disposition for enrichment can coevolve. The replicator steps shown in the top row of Fig. As for the iterated learning only, the diagrams in the bottom row of Fig. There are three main things to note. It remains to investigate how exactly these forces shape a population over time when they apply together. For this reason, we turn to a somewhat more realistic but still manageable setup with a larger set of types next.
Consider a state space with three states and think of it as a partition of possible worlds into cells where none, some, or all of the A s are B s, for some arbitrary fixed predicates A and B. With three messages there are possible lexica. Some assign the same representations to more than one message and others lexicalize the same representations but associate them with different messages. Out of these possible lexica, three kinds are of particular relevance. First, lexica that assign the same lexical representations to more than one message.
Such lexica lack in expressivity but may be favored by particular learning biases nonetheless see below. Finally, lexica that do not lexicalize upper bounds but would allow for perfect communication under additional pragmatic strengthening. There are six lexica of the second kind and six of the third. The following three lexica exemplify each kind:. Recall that types are a combination of a lexicon and a manner of language use. Accordingly, there is a total of types.
We refer to these as target types because they represent lexica and language use that conform to the majority view of scalar implicatures. Twelve types are either literal or pragmatic users of lexica of the kind. Note that while different types may lexicalize the same representations, they may nevertheless map different states to different overt messages. Consequently, different types of the same kind will fail to understand each other completely.
There is a growing effort to develop empirically testable representational languages that allow for the measure of semantic complexity. At its core, a LOT defines a set of operations and composition rules from which lexical representations can be derived. Applications of generative rules have a cost attached to them. Here, we simply assume that the formation of Boolean combinations of representations incurs two cost units, while all other rule applications incur only one cost unit. The prior probability of a type is just the prior probability of its lexicon.
The prior of a lexicon is a function of the complexity of the lexical representations in its image set. Lexica with simpler representations accordingly have a higher prior. Lexica of type are less likely, but more likely than. The menu of inductive biases argued to shape language acquisition is steadily being refined. Even when these biases are considered in isolation, there are many ways in which they can be translated into priors over types. In view of these numerous possibilities, we should stress that these details—from the generative grammar to its complexity measure—are to be regarded as one convenient operationalization of one general approach to explicating learning biases; this is not an implicit commitment to the claim that this particular instrumentalization is the single most plausible.
We merely want to have a maximally concrete working example in which the priors over types are systematically related to the complexity of lexical representations. Selection based on fitness alone may, therefore, not lead to prevalence of target types in the population. On the other hand, lexica of type are simpler than those of type by the postulated measure from above.
This may make them more likely to be adopted by learners, especially when l is high. Still, lexica of the kind are in turn even more likely a priori than lexica of the kind.
Lexical meaning in context : a web of words
Simulation results will shed light on the question whether target types can emerge, and for which parameter constellations this is likely. As before, we first look at the behavior of the replicator and mutator step in isolation, and then in combination. All reported results are the outcome of 50 update steps. These outcomes correspond to developmental plateaus in which change is, if not absent, then at least very slow. In other words, even if the resulting states do not correspond to an eventual attracting state, they characterize population states in which the system remains for a long time.
The plot also indicates the proportion of the majority type : the type with the highest proportion in the population. The result is a very long transition with near stagnancy in a rather homogeneous population with many types. As suggested by Fig. In contrast, in cases the majority types had with close to an even share between literal and pragmatic types , corresponding to a mean proportion of about. Its effects when learners exhibit a stronger tendency toward posterior maximization are illustrated in the other two plots of Fig.
The prior shows that while users of are not the most favored by the inductive bias compared, for example, to , they are nevertheless more advantaged than others, such as , in virtue of the relatively simple semantics they conventionalize. As a consequence, a stronger propensity to maximize the posterior increases their proportion in the population. Each is passed on to the next generation with the same faithfulness and, differently from a pressure for communicative success, they do not stand in competition with each other. This makes differences in the likelihood of particular types having generated a learning input less pronounced.
This favors functionally deficient but a priori preferred types such as those that use. A pressure for learnability alone may consequently lead to a spread of communicatively suboptimal types that lexicalize simpler semantics. In sum, when pressured for learnability, pragmatic is promoted over functionally similar but semantically more complex alternatives such as. However, learnability alone does not foment the propagation of a single target type across the population, because it does not differentiate between different ways of mapping the same semantic representations onto different overt signals.
Pressure for communicative success and learnability is not sufficient on their own to have a single target type dominate the population. But the combination of both pressures can lead to the selection of a single target type see Fig. This joint influence is summarized in Fig. This is due to the fact that the likelihood that a small sequence was produced by any type is relatively uniform.
In contrast, larger values increasingly allow learners to differentiate types with different signaling behaviors. In sum, when pressured for communicative success only, target types are outperformed by competitor types. When pressured for learnability only, populations are polymorphic due to lack of competition.
When we combine both pressures, the slight functional disadvantage of targets is counterbalanced by an advantage in learnability. This leads to large proportions of targets and, due to the competition among types applied by each replicator steps, to more monomorphic populations. To conclude, target types can come to dominate the population if three assumptions are met: i language is pressured toward both communicative success and learnability; ii pragmatic language use is an option; iii learners prefer simpler over more complex lexical representations and exhibit a tendency toward the acquisition of the type that best explains the learning data.
Main contributions of the model are i its modular separation of communicative success and learnability on evolutionary trajectories, ii the characterization of language learning as a joint inference over linguistic behavior and lexical meaning, and iii the possibility to trace the coevolution of conventional semantics and pragmatic use. A main difference is that here we considered communicative pressure for mutual understanding.
We see three main reasons for considering communicative success rather than just expressivity, and for looking at communication and learning rather than just learning. Second and more important, types may be equally expressive but their performance as a means of information transfer depends not only on themselves but on the population they find themselves in compare the competition of target types in Fig.
Taking communication into consideration allows the model to be responsive to the task for which language is learned. Lastly, chains of iterated learning alone do not put types in direct competition. Accordingly, learning alone leads to polymorphous populations in which multiple types of a kind coexist Nowak, The main result of our case study is that types that correspond to the majority view of scalar implicatures—scalar readings are nonlexicalized pragmatic enrichments—can come to dominate a population. This can happen if pragmatic language use is recruited indirectly by a preference for simpler lexical representations relative to more complex ones that lead to comparable overt linguistic behavior without pragmatic language use.
Under this view, semantics and pragmatics play a synergic role and can coevolve: Pragmatic use allows maintenance of simpler representations; pressure toward representational simplicity indirectly promotes pragmatic over literal language use. While the results of this case study are interesting, they also raise a number of critical issues. This may just be a technical quirk of the mutator step; but there is a related issue of empirical importance. Independently, semantics is also a well-defined field in its own right, often with synthetic properties.
Further related fields include philology , communication , and semiotics. The formal study of semantics can therefore be manifold and complex.
- Search form.
- Towards a Type-Theoretical Account of Lexical Semantics.
- Aurora: CV-01 (The Frontiers Saga, Book 1).
Semantics contrasts with syntax , the study of the combinatorics of units of a language without reference to their meaning , and pragmatics , the study of the relationships between the symbols of a language, their meaning, and the users of the language. Each of these is related to the general philosophical study of reality and the representation of meaning. In s psychosemantic studies became popular after Osgood 's massive cross-cultural studies using his semantic differential SD method that used thousands of nouns and adjective bipolar scales.
A specific form of the SD, Projective Semantics method  uses only most common and neutral nouns that correspond to the 7 groups factors of adjective-scales most consistently found in cross-cultural studies Evaluation, Potency, Activity as found by Osgood, and Reality, Organization, Complexity, Limitation as found in other studies.
In this method, seven groups of bipolar adjective scales corresponded to seven types of nouns so the method was thought to have the object-scale symmetry OSS between the scales and nouns for evaluation using these scales. However, deviations in this symmetric and very basic matrix might show underlying biases of two types: scales-related bias and objects-related bias. This OSS design meant to increase the sensitivity of the SD method to any semantic biases in responses of people within the same culture and educational background.
In linguistics , semantics is the subfield that is devoted to the study of meaning, as inherent at the levels of words, phrases, sentences, and larger units of discourse termed texts , or narratives. The study of semantics is also closely linked to the subjects of representation, reference and denotation. The basic study of semantics is oriented to the examination of the meaning of signs , and the study of relations between different linguistic units and compounds : homonymy , synonymy , antonymy , hypernymy , hyponymy , meronymy , metonymy , holonymy , paronyms.
A key concern is how meaning attaches to larger chunks of text, possibly as a result of the composition from smaller units of meaning. Traditionally, semantics has included the study of sense and denotative reference , truth conditions , argument structure, thematic roles , discourse analysis , and the linkage of all of these to syntax. In the late s, Richard Montague proposed a system for defining semantic entries in the lexicon in terms of the lambda calculus.
In these terms, the syntactic parse of the sentence John ate every bagel would consist of a subject John and a predicate ate every bagel ; Montague demonstrated that the meaning of the sentence altogether could be decomposed into the meanings of its parts and in relatively few rules of combination. The logical predicate thus obtained would be elaborated further, e. The notion of such meaning atoms or primitives is basic to the language of thought hypothesis from the s. Despite its elegance, Montague grammar was limited by the context-dependent variability in word sense, and led to several attempts at incorporating context, such as:.
In Chomskyan linguistics there was no mechanism for the learning of semantic relations, and the nativist view considered all semantic notions as inborn. Thus, even novel concepts were proposed to have been dormant in some sense. This view was also thought unable to address many issues such as metaphor or associative meanings, and semantic change , where meanings within a linguistic community change over time, and qualia or subjective experience. Another issue not addressed by the nativist model was how perceptual cues are combined in thought, e.
This view of semantics, as an innate finite meaning inherent in a lexical unit that can be composed to generate meanings for larger chunks of discourse, is now being fiercely debated in the emerging domain of cognitive linguistics  and also in the non- Fodorian camp in philosophy of language. To take an example of one word, red , its meaning in a phrase such as red book is similar to many other usages, and can be viewed as compositional. Indeed, these colours by themselves would not be called red by native speakers.
These instances are contrastive, so red wine is so called only in comparison with the other kind of wine which also is not white for the same reasons. This view goes back to de Saussure :. Each of a set of synonyms like redouter 'to dread' , craindre 'to fear' , avoir peur 'to be afraid' has its particular value only because they stand in contrast with one another.
No word has a value that can be identified independently of what else is in its vicinity. An attempt to defend a system based on propositional meaning for semantic underspecification can be found in the generative lexicon model of James Pustejovsky , who extends contextual operations based on type shifting into the lexicon. Thus meanings are generated "on the fly" as you go , based on finite context. Another set of concepts related to fuzziness in semantics is based on prototypes. The work of Eleanor Rosch in the s led to a view that natural categories are not characterizable in terms of necessary and sufficient conditions, but are graded fuzzy at their boundaries and inconsistent as to the status of their constituent members.
One may compare it with Jung 's archetype , though the concept of archetype sticks to static concept. Some post-structuralists are against the fixed or static meaning of the words.
Coevolution of Lexical Meaning and Pragmatic Use
Derrida , following Nietzsche , talked about slippages in fixed meanings. Systems of categories are not objectively out there in the world but are rooted in people's experience. This leads to another debate see the Sapir—Whorf hypothesis or Eskimo words for snow. Originates from Montague's work see above. A highly formalized theory of natural language semantics in which expressions are assigned denotations meanings such as individuals, truth values, or functions from one of these to another.
The truth of a sentence, and its logical relation to other sentences, is then evaluated relative to a model. Pioneered by the philosopher Donald Davidson , another formalized theory, which aims to associate each natural language sentence with a meta-language description of the conditions under which it is true, for example: 'Snow is white' is true if and only if snow is white. The challenge is to arrive at the truth conditions for any sentences from fixed meanings assigned to the individual words and fixed rules for how to combine them.
- Mountain Hazards and Disaster Risk Reduction.
- A Web of Words: Lexical Meaning in Context.
- Number Theory, Analysis and Geometry: In Memory of Serge Lang.
- Lose Weight Now: The Easy Way!
- Lexical Meaning in Context : Nicholas Asher : !
In practice, truth-conditional semantics is similar to model-theoretic semantics; conceptually, however, they differ in that truth-conditional semantics seeks to connect language with statements about the real world in the form of meta-language statements , rather than with abstract models. This theory is an effort to explain properties of argument structure.
Instead, we find nouns denoting relevant attributes of humans, such as victory , success and health :. However, there are many activities e. While with toast in 2 the focus is on the relevant attributes of the entity that is affected by the verb, rather than on the entity itself, in 3 the verb attend modulates the meaning of the nouns that fill the object position and causes words denoting [[Location]]s such as school , church and clinic to be reinterpreted as [[Activity]]s contextually.
Thus, each canonical member of a lexical set is recorded with statistical contextual information, like in 4 :. Coercion is a principled mechanism for accounting for the variety of interpretations that words exhibit in different contexts. In particular, coercion occurs when the meaning that a word exhibits in context is not inherent to the word itself i. In a broad definition, coercion covers conventionalized sense modulations as those discussed in Pustejovsky and Copestake and Briscoe , but also non-conventional i. Type coercion is an operation of type adjustment that occurs when none of the selectional preferences of a predicator match the type of a noun that it combines with in a particular text.
In this case, type coercion is invoked to explain how a mismatching verb-argument combination can be interpreted. This occurs for instance when a verb that normally selects an [[Activity]] as direct object e. In this case, the verb induces a re-interpretation of the noun from [[Physical Object]] to [[Activity]], so that it can successfully fill the object argument slot 8 :. In other words, the reconstructed event is an event in which the object is typically involved 9. Thus, there is subset of nouns with semantic type [[Location]] that occur as direct object of the verb ring , and a second subset that do not.
On the other hand, if a particular word or group of words is a unique outlier, coercion is the best option. While some cases are clear-cut, others are not. For example, the verb visit selects both for [[Location]] and [[Human]] as its direct objects. Unlike ring , these two uses of visit seem to map quite neatly onto two distinct senses and patterns e. It is perfectly possible to visit Paris for business, social or political purposes, but this is not the normal phraseology.
Thus, intuitively, storms abating and riots abating seem to be very different kinds of events and could be entered in the dictionary as different senses of the verb abate , corresponding roughly to a more literal and a more metaphorical sense. In the case of attend, for example, schools , clinic s, and churches are [[Location]]s where particular [[Activity]]s take place a class, a treatment, a mass and so on. It is to these [[Activity]]s that we refer to when we say that we attend such-and-such a location. So, as with ring , it makes sense to coerce all these nouns to the type [[Activity]] in the context of attend.
These are:. Alternations of semantic type are regular choices of types within an overall pattern in relation to a target lexical item.
Lexical Meaning in Context: A Web of Words
Two common alternations are for instance [[Human Institution]] very frequent in the subject of a large number of verbs that denote a cognitive action and [[Human Body Part]]:. The distinction between a semantic type and a semantic role can be defined as follows. The semantic type is an intrinsic attribute of a noun, while a semantic role has the attribute thrust upon it by the context.
Thus, for instance, the verb arrest is found in the following pattern:. This is a role assigned by context — specifically, by the selectional preferences of the verb arrest , which expects a subject with such characteristics. For example in 15 the role of [[Driver]] is assigned contextually to the type [[Human]] by the verb accelerate. An exploitation is a deliberately creative or unusual use of a norm — an established pattern of word use.
While alternations are regular choices of elements within an overall pattern, exploitations are dynamic and creative choices. Newly created metaphors are exploitations, but there are many other kinds of exploitation, too, some of them quite unremarkable, as in But what people normally cultivate is a subset of the set of all plants — flowers and vegetables, mainly. Weeds are outliers, and must be classified here as an exploitation. This preserves the homogeneity of the set of cultivated plants. Guardian Weekly , 13 November cited by Copestake and Briscoe The lexical sets populating a node in the ontology i.
Different verbs select different prototypical members of a semantic type even if the rest of the set remains the same.
Towards a Type-Theoretical Account of Lexical Semantics
For example, two verbs, wash and amputate , both typically select [[Body Part]] as their direct object. But while pretty well the entire set of [[Garment]]s occurs as direct object of put on and wear , in the case of hitch up this is not true. What do people hitch up? Typically: trousers, pants, skirt.
Maybe socks and stockings, nightdress, pyjamas. Outliers are: brassiere, bosoms. A newspaper is typically read, while a message is typically sent, a report is typically published, and so on see Jezek and Lenci for a full discussion of the distributional behaviour of words denoting [[Documents]]. Also, we typically take a taxi or a train to a destination, but it would be very odd to talk of someone taking an ambulance to hospital.
Take is something you do with vehicles that run a public service. Moreover, for cars, we typically predicate some activity involving causing damage damage , smash , crash , wreck , hit , destroy , whereas the typical co-occurrence of verbs like require , need , provide with ambulance shows how its purpose of use is central to its meaning. Verb selectional behaviour is closely connected with the structure of our conceptual system, but selectional constraints do not always map neatly onto semantic types.