Bunt H., Bos J., Pulman S. (eds.) Computing Meaning. Volume 4

Springer, 2014. — 258 p.While computers are very good at computing in general, they are not very good at computing meaning. There are at least three reasons why this may be so: (R1) the very notion of meaning, as expressed in natural language, is something extremely complex, and therefore difficult to compute; (R2) the process of computing meanings is extremely complex, because it requires the effective use of a variety of extremely rich information sources (linguistic knowledge, general knowledge of the world, specific knowledge of the domain of discourse, knowledge of interactive settings...); and (R3) the very notion of meaning is not well enough understood to effectively program and/or teach computers what it is and how it can be computed for a given natural language expression, occurring in a given context.
Most of the work in formal as well as in computational semantics tacitly assumes, different from (R3), that we do have a clear understanding of what we mean by meaning, and different from (R1), that natural language meanings are simple enough to be represented by very simple structures, such as formulas in first-order logic (or, equivalently, Discourse Representation Structures). Assuming that such structures are adequate representations of meaning, computing the meaning of a given natural language expression comes down to syntactic parsing of it and composing the semantic representations of the parts to form the meaning representation, which itself has a semantics defined by the representation formalism.
Since computational semantics started to develop, in the last two decades of the twentieth century (see Blackburn and Bos 2005), it has become clear that the dream of computing meaning representations by syntactic/semantic (de-)composition, made popular especially through the work of Richard Montague (see Thomason 1974), cannot become reality, simply because natural language expressions much of the time do not contain sufficiently much information to construct such a representation. Other information sources are indispensable. This insight has inspired the introduction of the notion of an underspecified meaning representation, which represents the semantic information that is present in the sentence without disambiguating those aspects for which the sentence does not contain sufficient information. It also became very clear that relying solely on linguistic information for computing meanings would lead to impossibly complex interpretation processes, due to the astronomical number of readings that ordinary sentences have when considered in isolation (see Bunt and Muskens 1999). Again, underspecified meaning representations offer solace here, as they obviate the need to fully disambiguate. Several of the chapters in this book, in particular in Part I, witness the ongoing search for appropriate forms of meaning representation and for methods of exploiting linguistic as well as other information in their computation.
A problematic aspect of the use of underspecified semantic representations is that they do not allow straightforward application of logic-based inference methods, since different resolutions of underspecifications may result in interpretations that allow different inferences (see e.g. van Deemter 1996; Blackburn et al. 2001). This is indeed problematic on the traditional view of meaning representations as unambiguously supporting a specific set of inferences, thereby explaining differences in meaning and relations between different meanings. One way to deal with this problem is to move away from strictly deductive approaches to inferencing, and instead turn to abductive methods (Hobbs et al. 1993) or to textual entailment, where inferencing is performed directly on natural language expressions, rather than on their interpretations, and logical proof is replaced by psychological plausibility (see e.g. Dagan et al. 2008 and Bos 2013). One way or another, the use of inference processes involving natural language expressions and/or their interpretations is needed, since nonlinguistic information must be exploited in order to arrive at intended and contextually appropriate interpretations; methods for combining pieces of information therefore have to be applied in order to arrive at a appropriate interpretations. The chapters in Part II of this book are all concerned with forms of inferencing (or combining pieces of information) in the computation of meanings.
Related to the limitations of effectively following strictly logic- and rule-based methods in the computation of meaning is the exploration of statistical and machine learning techniques that have been successfully applied in other areas of computational linguistics. These techniques presuppose the availability of large corpora, and can benefit in particular from semantically annotated resources. The development of such corpora (e.g. Basile et al. 2012), and of well-founded semantic annotation methodologies (see Bunt 2013), have supported the use of these new methods in computational semantics research (see e.g. Clark and Pulman 2007), as reflected in several of the chapters in this book, both in Part I and in Part III.
The chapters in this book are organized into three parts. A first cluster of four chapters is focused on aspects of the representation of meaning and the computation of these representations. A second group of four chapters is concerned with issues of inferencing and its role in language understanding. The chapters in the third and final cluster of four deal with resources for meaning computation and their use.Computing Meaning: Annotation, Representation, and Inference
Semantic Representation and Compositionality
Deterministic Statistical Mapping of Sentences to Underspecified Semantics
A Formal Approach to Linking Logical Form and Vector-Space Lexical Semantics
Annotations that Effectively Contribute to Semantic Interpretation
Concrete Sentence Spaces for Compositional Distributional Models of Meaning
Inference and Understanding
Recognizing Textual Entailment and Computational Semantics
Abductive Reasoning with a Large Knowledge Base for Discourse Processing
Natural Logic and Natural Language Inference
Designing Efficient Controlled Languages for Ontologies
Semantic Resources and Annotation
A Context-Change Semantics for Dialogue Acts
VerbNet Class Assignment as aWSD Task
Annotation of Compositional Operations with GLML
Incremental Recognition and Prediction of Dialogue Acts