Underspecified Semantic Representations Research Articles

A Notion of Semantic Coherence for Underspecified Semantic Representation

The general problem of finding satisfying solutions to constraint-based underspecified representations of quantifier scope is NP-complete. Existing frameworks, including Dominance Graphs, Minimal Recursion Semantics, and Hole Semantics, have struggled to balance expressivity and tractability in order to cover real natural language sentences with efficient algorithms. We address this trade-off with a general principle of coherence, which requires that every variable introduced in the domain of discourse must contribute to the overall semantics of the sentence. We show that every underspecified representation meeting this criterion can be efficiently processed, and that our set of representations subsumes all previously identified tractable sets.

Opposing effects of semantic diversity in lexical and semantic relatedness decisions.

Semantic ambiguity has often been divided into 2 forms: homonymy, referring to words with 2 unrelated interpretations (e.g., bark), and polysemy, referring to words associated with a number of varying but semantically linked uses (e.g., twist). Typically, polysemous words are thought of as having a fixed number of discrete definitions, or “senses,” with each use of the word corresponding to one of its senses. In this study, we investigated an alternative conception of polysemy, based on the idea that polysemous variation in meaning is a continuous, graded phenomenon that occurs as a function of contextual variation in word usage. We quantified this contextual variation using semantic diversity (SemD), a corpus-based measure of the degree to which a particular word is used in a diverse set of linguistic contexts. In line with other approaches to polysemy, we found a reaction time (RT) advantage for high SemD words in lexical decision, which occurred for words of both high and low imageability. When participants made semantic relatedness decisions to word pairs, however, responses were slower to high SemD pairs, irrespective of whether these were related or unrelated. Again, this result emerged irrespective of the imageability of the word. The latter result diverges from previous findings using homonyms, in which ambiguity effects have only been found for related word pairs. We argue that participants were slower to respond to high SemD words because their high contextual variability resulted in noisy, underspecified semantic representations that were more difficult to compare with one another. We demonstrated this principle in a connectionist computational model that was trained to activate distributed semantic representations from orthographic inputs. Greater variability in the orthography-to-semantic mappings of high SemD words resulted in a lower degree of similarity for related pairs of this type. At the same time, the representations of high SemD unrelated pairs were less distinct from one another. In addition, the model demonstrated more rapid semantic activation for high SemD words, thought to underpin the processing advantage in lexical decision. These results support the view that polysemous variation in word meaning can be conceptualized in terms of graded variation in distributed semantic representations.

Semantic Underspecification

AbstractSemantic underspecification is a technique to capture several readings of an ambiguous expression in one single representation by deliberately omitting the differences between the readings in the representation. First, underspecification formalisms will be presented to introduce underspecification in general and to outline important properties of these formalisms that allow their classification into subgroups. After expounding the kinds of ambiguity to which underspecification can be applied, the article then presents various motivations for the use of underspecification, and shows how underspecified semantic representations can be further processed.

Scope and Situation Binding in LTAG Using Semantic Unification

This paper sets up a framework for Lexicalized Tree Adjoining Grammar (LTAG) semantics that brings together ideas from different recent approaches addressing some shortcomings of LTAG semantics based on the derivation tree. The approach assigns underspecified semantic representations and semantic feature structure descriptions to elementary trees. Semantic computation is guided by the derivation tree and consists of adding feature value equations to the descriptions. A rigorous formal definition of the framework is given. Then, within this framework, an analysis is proposed that accounts for the different scopal properties of quantificational NPs (including nested NPs), adverbs, raising verbs and attitude verbs. Furthermore, by integrating situation variables in the semantics, different situation binding possibilities are derived for different types of quantificational elements.

The effect of context on the processing of type-shifting verbs

Recent research in lexical semantics has suggested that verbs such as begin and enjoy semantically select for a complement that denotes an activity or an event. When no such activity or event is specified in the form of a progressive or infinitival complement, as in John began ( to read/reading) the book, the verb is said to “coerce” the NP direct object to shift its role to encompass the activity that begin requires as complement (e.g., writing, reading). Empirical support for this view has been provided by McElree, Traxler, Pickering, Seely, & Jackendoff (2001). In the present study, however, in two self-paced reading experiments, type-shifting effects (taken to be longer reading times engendered by the computation of the coercion process) were not obtained with sentences in isolation (Experiment 1) or with sentences embedded in contexts that specified the nature of the activity performed over the complement NP (Experiment 2). It is argued that type-shifting verbs are similar to non-preferred verbs for given contexts and that type-shifting operations are pragmatic inferences computed over underspecified semantic representations.

Computational Semantics in Discourse: Underspecification, Resolution, and Inference

In this paper I introduce a formalism for natural language understanding based on a computational implementation of Discourse Representation Theory. The formalism covers a wide variety of semantic phenomena (including scope and lexical ambiguities, anaphora and presupposition), is computationally attractive, and has a genuine inference component. It combines a well-established linguistic formalism (DRT) with advanced techniques to deal with ambiguity (underspecification), and is innovative in the use of first-order theorem proving techniques. The architecture of the formalism for natural language understanding that I advocate consists of three levels of processing: underspecification, resolution, and inference. Each of these levels has a distinct function and therefore employs a different kind of semantic representation. The mappings between these different representations define the interfaces between the levels. I show how underspecified semantic representations can be built in a compositional way (for a fragment of English Grammar) using standard techniques borrowed from the λ-calculus, how inferences can be carried out on discourse representations using a translation to first-order logic, and how existing research prototypes (discourse processing and spoken-dialogue systems) implement the formalism.

Processing underspecified semantic representations in the constraint language for lambda structures

The Constraint Language for Lambda Structures (CLLS)is an expressive language of tree descriptions which combinesdominance constraints with powerful parallelism and bindingconstraints. CLLS was introduced as a uniform framework for definingunderspecified semantics representations of natural languagesentences, covering scope, ellipsis, and anaphora. This articlepresents saturation-based algorithms for processing the completelanguage of CLLS. It also gives an overview of previous results on questions of processing and complexity.

Relaxing Underspecified Semantic Representations for Reinterpretation

Type and sort conflicts in semantics are usually resolved by a process of reinterpretation, which introduces an operator into the semantic representation. We elaborate on the foundations of a recent approach to reinterpretation within a framework for semantic underspecification. In this approach, relaxed underspecified semantic representations are inferred from the syntactic structure, leaving space for subsequent addition of reinterpretation operators. Unfortunately, a structural danger of overgeneration is inherent to the relaxation of underspecified semantic representations. We identify the problem and distinguish structural properties that avoid it. We furthermore develop techniques for proving these properties and apply them to prove the safety of relaxation in a prototypical syntax/semantics interface. In doing so, we present some novel properties of tree descriptions in the constraint language over lambda structures (CLLS).

DISCOURSE ABOUT BELIEFS

Discourse Representation Theory provides an adequate framework for representing sentences that express reports on beliefs. However, all the possible readings of a belief report which discriminate between different ways in which the individual is known to the believer are treated equally. In this paper I demonstrate that some of these readings are more salient than others to the hearer and the systematic gradation of this saliency is parallel to the degrees and types of intentions involved in uttering a belief report. I demonstrate how an account which makes use of referential, informative and communicative intentions supplements DRT and disposes of the need to postulate either ambiguities of reading of belief constructions or the underspecified semantic representation