Francis Jeffry Pelletier
Most linguists have heard of semantic compositionality. Some will have heard that it is the fundamental truth of semantics. Others will have been told that it is so thoroughly and completely wrong that it is astonishing that it is still being taught. The present article attempts to explain all this. Much of the discussion of semantic compositionality takes place in three arenas that are rather insulated from one another: (a) philosophy of mind and language, (b) formal semantics, and (c) cognitive linguistics and cognitive psychology. A truly comprehensive overview of the writings in all these areas is not possible here. However, this article does discuss some of the work that occurs in each of these areas. A bibliography of general works, and some Internet resources, will help guide the reader to some further, undiscussed works (including further material in all three categories).
This is an advance summary of a forthcoming article in the Oxford Research Encyclopedia of Linguistics. Please check back later for the full article.
Computational semantics performs automatic meaning analysis of natural language. Research in computational semantics designs meaning representations and develops mechanisms for automatically assigning those representations and reasoning over them. Computational semantics is not a single monolithic task but consists of many subtasks, including word sense disambiguation, multi-word expression analysis, semantic role labeling, the construction of sentence semantic structure, coreference resolution, and the automatic induction of semantic information from data.
The development of manually constructed resources has been vastly important in driving the field forward. Examples include WordNet, PropBank, FrameNet, VerbNet, and TimeBank. These resources specify the linguistic structures to be targeted in automatic analysis, and they provide high quality human-generated data that can be used to train machine learning systems. Supervised machine learning based on manually constructed resources is a widely used technique.
A second core strand has been the induction of lexical knowledge from text data. For example, words can be represented through the contexts in which they appear (called distributional vectors or embeddings), such that semantically similar words have similar representations. Or semantic relations between words can be inferred from patterns of words that link them. Wide-coverage semantic analysis always needs more data, both lexical knowledge and world knowledge, and automatic induction at least alleviates the problem.
Compositionality is a third core theme: the systematic construction of structural meaning representations of larger expressions from the meaning representations of their parts. The representations typically use logics of varying expressivity, which makes them well suited to performing automatic inferences with theorem provers.
Manual specification and automatic acquisition of knowledge are closely intertwined. Manually created resources are automatically extended or merged. The automatic induction of semantic information is guided and constrained by manually specified information, which is much more reliable. And for restricted domains, the construction of logical representations is learned from data.
It is at the intersection of manual specification and machine learning that some of the current larger questions of computational semantics are located. For instance, should we build general-purpose semantic representations, or is lexical knowledge simply too domain-specific, and would we be better off learning task-specific representations every time? When performing inference, is it more beneficial to have the solid ground of a human-generated ontology, or is it better to reason directly with text snippets for more fine-grained and gradual inference? Do we obtain a better and deeper semantic analysis as we use better and deeper manually specified linguistic knowledge, or is the future in powerful learning paradigms that learn to carry out an entire task from natural language input and output alone, without pre-specified linguistic knowledge?
Hearers and readers make inferences on the basis of what they hear or read. These inferences are partly determined by the linguistic form that the writer or speaker chooses to give to her utterance. The inferences can be about the state of the world that the speaker or writer wants the hearer or reader to conclude are pertinent, or they can be about the attitude of the speaker or writer vis-à-vis this state of affairs. The attention here goes to the inferences of the first type. Research in semantics and pragmatics has isolated a number of linguistic phenomena that make specific contributions to the process of inference. Broadly, entailments of asserted material, presuppositions (e.g., factive constructions), and invited inferences (especially scalar implicatures) can be distinguished.
While we make these inferences all the time, they have been studied piecemeal only in theoretical linguistics. When attempts are made to build natural language understanding systems, the need for a more systematic and wholesale approach to the problem is felt. Some of the approaches developed in Natural Language Processing are based on linguistic insights, whereas others use methods that do not require (full) semantic analysis.
In this article, I give an overview of the main linguistic issues and of a variety of computational approaches, especially those stimulated by the RTE challenges first proposed in 2004.
The central goal of the Lexical Semantic Framework (LSF) is to characterize the meaning of simple lexemes and affixes and to show how these meanings can be integrated in the creation of complex words. LSF offers a systematic treatment of issues that figure prominently in the study of word formation, such as the polysemy question, the multiple-affix question, the zero-derivation question, and the form and meaning mismatches question.
LSF has its source in a confluence of research approaches that follow a decompositional approach to meaning and, thus, defines simple lexemes and affixes by way of a systematic representation that is achieved via a constrained formal language that enforces consistency of annotation. Lexical-semantic representations in LSF consist of two parts: the Semantic/Grammatical Skeleton and the Semantic/Pragmatic Body (henceforth ‘skeleton’ and ‘body’ respectively). The skeleton is comprised of features that are of relevance to the syntax. These features act as functions and may take arguments. Functions and arguments of a skeleton are hierarchically arranged. The body encodes all those aspects of meaning that are perceptual, cultural, and encyclopedic.
Features in LSF are used in (a) a cross-categorial, (b) an equipollent, and (c) a privative way. This means that they are used to account for the distinction between the major ontological categories, may have a binary (i.e., positive or negative) value, and may or may not form part of the skeleton of a given lexeme. In order to account for the fact that several distinct parts integrate into a single referential unit that projects its arguments to the syntax, LSF makes use of the Principle of Co-indexation. Co-indexation is a device needed in order to tie together the arguments that come with different parts of a complex word to yield only those arguments that are syntactically active.
LSF has an important impact on the study of the morphology-lexical semantics interface and provides a unitary theory of meaning in word formation.
Natural language ontology is a branch of both metaphysics and linguistic semantics. Its aim is to uncover the ontological categories, notions, and structures that are implicit in the use of natural language, that is, the ontology that a speaker accepts when using a language. Natural language ontology is part of “descriptive metaphysics,” to use Strawson’s term, or “naive metaphysics,” to use Fine’s term, that is, the metaphysics of appearances as opposed to foundational metaphysics, whose interest is in what there really is.
What sorts of entities natural language involves is closely linked to compositional semantics, namely what the contribution of occurrences of expressions in a sentence is taken to be. Most importantly, entities play a role as semantic values of referential terms, but also as implicit arguments of predicates and as parameters of evaluation.
Natural language appears to involve a particularly rich ontology of abstract, minor, derivative, and merely intentional objects, an ontology many philosophers are not willing to accept. At the same time, a serious investigation of the linguistic facts often reveals that natural language does not in fact involve the sort of ontology that philosophers had assumed it does.
Natural language ontology is concerned not only with the categories of entities that natural language commits itself to, but also with various metaphysical notions, for example the relation of part-whole, causation, material constitution, notions of existence, plurality and unity, and the mass-count distinction.
An important question regarding natural language ontology is what linguistic data it should take into account. Looking at the sorts of data that researchers who practice natural language ontology have in fact taken into account makes clear that it is only presuppositions, not assertions, that reflect the ontology implicit in natural language.
The ontology of language may be distinctive in that it may in part be driven specifically by language or the use of it in a discourse. Examples are pleonastic entities, discourse referents conceived of as entities of a sort, and an information-based notion of part structure involved in the semantics of plurals and mass nouns. Finally, there is the question of the universality of the ontology of natural language. Certainly, the same sort of reasoning should apply to consider it universal, in a suitable sense, as has been applied for the case of (generative) syntax.
The term coordination refers to the juxtaposition of two or more conjuncts often linked by a conjunction such as and or or. The conjuncts (e.g., our friend and your teacher in Our friend and your teacher sent greetings) may be words or phrases of any type. They are a defining property of coordination, while the presence or absence of a conjunction depends on the specifics of the particular language. As a general phenomenon, coordination differs from subordination in that the conjuncts are typically symmetric in many ways: they often belong to like syntactic categories, and if nominal, each carries the same case. Additionally, if there is extraction, this must typically be out of all conjuncts in parallel, a phenomenon known as Across-the-Board extraction. Extraction of a single conjunct, or out of a single conjunct, is prohibited by the Coordinate Structure Constraint. Despite this overall symmetry, coordination does sometimes behave in an asymmetric fashion. Under certain circumstances, the conjuncts may be of unlike categories or extraction may occur out of one conjunct, but not another, thus yielding apparent violations of the Coordinate Structure Constraint. In addition, case and agreement show a wide range of complex and sometimes asymmetric behavior cross-linguistically. This tension between the symmetric and asymmetric properties of coordination is one of the reasons that coordination has remained an interesting analytical puzzle for many decades.
Within the general area of coordination, a number of specific sentence types have generated much interest. One is Gapping, in which two sentences are conjoined, but material (often the verb) is missing from the middle of the second conjunct, as in Mary ate beans and John _ potatoes. Another is Right Node Raising, in which shared material from the right edge of sentential conjuncts is placed in the right periphery of the entire sentence, as in The chefs prepared __ and the customers ate __ [a very elaborately constructed dessert]. Finally, some languages have a phenomenon known as comitative coordination, in which a verb has two arguments, one morphologically plural and the other comitative (e.g., with the preposition with), but the plural argument may be understood as singular. English does not have this phenomenon, but if it did, a sentence like We went to the movies with John could be understood as John and I went to the movies.
Eve V. Clark
The words and word-parts children acquire at different stages offer insights into how the mental lexicon might be organized. Children first identify ‘words,’ recurring sequences of sounds, in the speech stream, attach some meaning to them, and, later, analyze such words further into parts, namely stems and affixes. These are the elements they store in memory in order to recognize them on subsequent occasions. They also serve as target models when children try to produce those words themselves. When they coin words, they make use of bare stems, combine certain stems with each other, and sometimes add affixes as well. The options they choose depend on how much they need to add to coin a new word, which familiar elements they can draw on, and how productive that option is in the language. Children’s uses of stems and affixes in coining new words also reveal that they must be relying on one representation in comprehension and a different representation in production. For comprehension, they need to store information about the acoustic properties of a word, taking into account different occasions, different speakers, and different dialects, not to mention second-language speakers. For production, they need to work out which articulatory plan to follow in order to reproduce the target word. And they take time to get their production of a word aligned with the representation they have stored for comprehension. In fact, there is a general asymmetry here, with comprehension being ahead of production for children, and also being far more extensive than production, for both children and adults. Finally, as children add more words to their repertoires, they organize and reorganize their vocabulary into semantic domains. In doing this, they make use of pragmatic directions from adults that help them link related words through a variety of semantic relations.
Lexical semantics is the study of word meaning. Descriptively speaking, the main topics studied within lexical semantics involve either the internal semantic structure of words, or the semantic relations that occur within the vocabulary. Within the first set, major phenomena include polysemy (in contrast with vagueness), metonymy, metaphor, and prototypicality. Within the second set, dominant topics include lexical fields, lexical relations, conceptual metaphor and metonymy, and frames. Theoretically speaking, the main theoretical approaches that have succeeded each other in the history of lexical semantics are prestructuralist historical semantics, structuralist semantics, and cognitive semantics. These theoretical frameworks differ as to whether they take a system-oriented rather than a usage-oriented approach to word-meaning research but, at the same time, in the historical development of the discipline, they have each contributed significantly to the descriptive and conceptual apparatus of lexical semantics.
Željko Bošković and Troy Messick
Economy considerations have always played an important role in the generative theory of grammar. They are particularly prominent in the most recent instantiation of this approach, the Minimalist Program, which explores the possibility that Universal Grammar is an optimal way of satisfying requirements that are imposed on the language faculty by the external systems that interface with the language faculty which is also characterized by optimal, computationally efficient design. In this respect, the operations of the computational system that produce linguistic expressions must be optimal in that they must satisfy general considerations of simplicity and efficient design. Simply put, the guiding principles here are (a) do something only if you need to and (b) if you do need to, do it in the most economical/efficient way. These considerations ban superfluous steps in derivations and superfluous symbols in representations. Under economy guidelines, movement takes place only when there is a need for it (with both syntactic and semantic considerations playing a role here), and when it does take place, it takes place in the most economical way: it is as short as possible and carries as little material as possible. Furthermore, economy is evaluated locally, on the basis of immediately available structure. The locality of syntactic dependencies is also enforced by minimal search and by limiting the number of syntactic objects and the amount of structure accessible in the derivation. This is achieved by transferring parts of syntactic structure to the interfaces during the derivation, the transferred parts not being accessible for further syntactic operations.
Haihua Pan and Yuli Feng
Cross-linguistic data can add new insights to the development of semantic theories or even induce the shift of the research paradigm. The major topics in semantic studies such as bare noun denotation, quantification, degree semantics, polarity items, donkey anaphora and binding principles, long-distance reflexives, negation, tense and aspects, eventuality are all discussed by semanticists working on the Chinese language. The issues which are of particular interest include and are not limited to: (i) the denotation of Chinese bare nouns; (ii) categorization and quantificational mapping strategies of Chinese quantifier expressions (i.e., whether the behaviors of Chinese quantifier expressions fit into the dichotomy of A-Quantification and D-quantification); (iii) multiple uses of quantifier expressions (e.g., dou) and their implication on the inter-relation of semantic concepts like distributivity, scalarity, exclusiveness, exhaustivity, maximality, etc.; (iv) the interaction among universal adverbials and that between universal adverbials and various types of noun phrases, which may pose a challenge to the Principle of Compositionality; (v) the semantics of degree expressions in Chinese; (vi) the non-interrogative uses of wh-phrases in Chinese and their influence on the theories of polarity items, free choice items, and epistemic indefinites; (vii) how the concepts of E-type pronouns and D-type pronouns are manifested in the Chinese language and whether such pronoun interpretations correspond to specific sentence types; (viii) what devices Chinese adopts to locate time (i.e., does tense interpretation correspond to certain syntactic projections or it is solely determined by semantic information and pragmatic reasoning); (ix) how the interpretation of Chinese aspect markers can be captured by event structures, possible world semantics, and quantification; (x) how the long-distance binding of Chinese ziji ‘self’ and the blocking effect by first and second person pronouns can be accounted for by the existing theories of beliefs, attitude reports, and logophoricity; (xi) the distribution of various negation markers and their correspondence to the semantic properties of predicates with which they are combined; and (xii) whether Chinese topic-comment structures are constrained by both semantic and pragmatic factors or syntactic factors only.