Reformulating Discourse Connectives for Non-Expert Readers

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 1002–1010, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Reformulating Discourse Connectives for Non-Expert Readers Advaith Siddharthan Napoleon Katsos Department of Computing Science Research Centre for English and Applied Linguistics University of Aberdeen University of Cambridge advaith@abdn.ac.uk nk248@cam.ac.uk Abstract In this paper we report a behavioural ex- periment documenting that different lexico- syntactic formulations of the discourse rela- tion of causation are deemed more or less ac- ceptable by different categories of readers. We further report promising results for automati- cally selecting the formulation that is most ap- propriate for a given category of reader using supervised learning. This investigation is em- bedded within a longer term research agenda aimed at summarising scientific writing for lay readers using appropriate paraphrasing. 1 Introduction There are many reasons why a speaker/writer would want to choose one formulation of a discourse rela- tion over another; for example, maintaining thread of discourse, avoiding shifts in focus and issues of salience and end weight. There are also reasons to use different formulations for different audiences; for example, to account for differences in reading skills and domain knowledge. In this paper, we present a psycholinguistic experiment designed to il- luminate the factors that determine the appropriate- ness of particular realisations of discourse relations for different audiences. The second part of this pa- per focuses on training a natural language generation system to predict which realisation choices are more felicitous than others for a given audience. Our para- phrases include eight different constructions. Con- sider 1a.–d. below: (1) a. Tom ate because he was hungry. b. Tom ate because of his hunger. c. Tom’s hunger caused him to eat. d. The cause of Tom’s eating was his hunger. These differ in terms of the lexico-syntactic prop- erties of the discourse marker (shown in bold font). Indeed the discourse markers here are conjunctions, prepositions, verbs and nouns. As a consequence the propositional content is expressed either as a clause or a noun phrase (“he was hungry” vs “his hunger”, etc.). Additionally, the order of presenta- tion of propositional content can be varied to give four more lexico-syntactic paraphrases: (1) e. Because Tom was hungry, he ate. f. Because of his hunger, Tom ate. g. Tom’s eating was caused by his hunger. h. Tom’s hunger was the cause of his eating. It is clear that some formulations of this propo- sitional content are more felicitous than others; for example, 1a. seems preferable to 1d., but for a different propositional content, other formulations might be more felicitous (for instance, example 4, section 3.1, where the passive seems in fact prefer- able). While discourse level choices based on infor- mation ordering play a role in choosing a formula- tion, it is of particular interest to us that some de- contextualised information orderings within a sen- tence are deemed unacceptable. Any summarisation task that considers discourse coherence should not introduce sentence-level unacceptability. We now …

Content truncated to 500 words