Developing Language Independent Event Representations that are Inferrable from Linguistic Expressions in Large Text Corpora
Abstract
The project aims to develop a representation of the structure of events in order to identify particular types of events expressed linguistically in large text corpora. A major challenge in this task is that the same event can be described by different verbs and different grammatical constructions, for example "I opened the safe with a blowtorch", "I used a blowtorch to cut open the safe", and "I took a blowtorch and cut the safe open". The goal is to provide a theoretical framework for event representation of linguistic expressions that can account for this variation in the linguistic expression of events. Identifying an event no matter how it is expressed requires a more language-independent event representation than is usually found in linguistic semantic analyses. Most current linguistic semantic analyses use an event representation similar to formal logic, such as predicate calculus, but these event representations usually capture just one way to express the event. This fact can make it difficult to recognize other possible ways to express the same event, in English or in other languages. The current project expands on an event representation presented in a recent publication by the principal investigator that is language-independent in two important respects. First, the event representation models events in three dimensions: time, causal interactions among participants, and qualitative states of the participants, rather than a predicate-calculus type of representation. These three dimensions represent major properties of events in the real world that are largely independent of linguistic expression in English or another language. Second, the event representation recognizes the fact that the structure of an event is not an inherent property of the meaning of a verb. Instead, the structure of an event is represented by a more abstract semantic structure called an image schema. For example, the verb "tap" is normally assumed to describe contact between two objects. But in "He tapped the ball into the pocket" (in a game of billiards), "tap" is used to describe the event of causing the ball to move into a location. Caused motion is an example of an event image schema, but one cannot identify the event image schema from the verb "tap" alone. The event image schema is inferred by the combination of the verb, the construction the verb occurs in, and the context of language use. Recognizing that event structure is not solely determined by the verb meaning allows for a more flexible and realistic relationship between linguistic expressions and the events they denote, and should improve accurate event identification in large corpora. The project has four primary tasks. The first is to identify the major event image schemas to allow for coverage of a large range of verbs and their occurrence in different grammatical constructions, and to provide a semantic analysis of the structure of the event image schemas. The second task is to implement a computational model of the event structure representation that captures the semantic analysis of the event image schemas. The third task is to inventory and analyze the actual combinations of verbs and different constructions, and the event image schema each verb-plus-construction expresses, and to construct similarity measures that would help to predict the likelihood of novel verb-plus-construction combinations and their event image schemas. The fourth task is to provide a deeper analysis of the basic meanings of verbs in order to better understand why individual verbs can express the range of event image schemas that they do.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 26, 2016
- Source ID
- HDTRA11510063
Entities
People
- William Croft
Organizations
- Defense Threat Reduction Agency
- University of New Mexico