Applied computational models of discourse, argument, and text

点击翻译按钮获取中文摘要

Our research is in computational linguistics (CL), natural language processing (NLP), and their applications. Themes that run through the work are (i) computationally determining the structure of discourse and argumentation; (ii) computational notions of paraphrase and of the semantic distance between words, larger linguistic expressions, or documents; and (iii) the use of statistical classification methods in text analysis. We propose the following new research: (1) Finding the structure and framing of discourse and arguments: Understanding a speaker's argument entails understanding not only what is presented as evidence and what as conclusion but also how the conclusion follows from the evidence, what the unstated premises (enthymemes) are, and what the implicit framing of the issue is. We will look in particular at opinionated texts in which argumentation structure is fairly explicit on the surface but determining enthymemes and framing is crucial to full understanding. For automatic text analysis, we take quantifiable semantic characteristics of the speaker's presentation of a position as indicators or proxies of the framing, which can then be interpreted qualitatively. In a simple analysis, this could be merely a statistical analysis of the key concepts of the text, as denoted by content words and significant collocations -- something like a topic model. Here, however, we propose a novel, more-sophisticated analysis in which we also look at the actual argumentation structures and discourse relationships of the text and how the concepts adduced by the lower-level linguistic components are used in these structures. This will draw on and extend our recent work on discourse parsing and the identification of argumentation schemes in text. Although these are difficult tasks for which the state-of-the-art is far from perfect, we hypothesize that typical political speech contains a sufficiently well-cued discourse structure that the analyses that we can achieve, although still quite imperfect, will be usefully indicative of issue framing. (2) Finding precedent scientific literature: Researchers often have difficulty searching for past research relevant to, or precedent to, their new or proposed research, and often resort simply to Google keyword searches, which are rarely adequate. We will develop methods for searching scientific literature that use semantic and structural relationships to find publications that are possibly relevant to a new text. We will concentrate in particular on the legacy literature of biodiversity, for which conventional keyword searches are almost invariably insufficient because in this literature, more so than most other fields of science, related concepts are often described or explained in different terms, or in completely different conceptual frameworks, from those of contemporary research. As a result, relevant legacy publications, or even whole literatures, may remain hidden to term-based methods. This goal will not be reached in five years, but it motivates the next stage of our work because it requires bringing together many of the methods of natural language processing developed in our own and other researchers' work of the past decade or more: (a) the recognition of paraphrase and of textual entailment, and measurement of semantic similarity at the sentence level and above; (b) the automatic analysis of the structure and argumentation of scholarly papers and scientific discourse (this is a point of overlap with (1) above, but in scientific texts we expect the micro-structure to be less explicit and the macro-structure more explicit than in opinion texts). (3) We will continue our work on automatic authorship identification; on characterizing aphasic speech; and on the philosophy of CL.