Memory-Based Learning for Article Generation

In: Proceedings Of CoNLL-2000 and LLL-2000, pages 43-48, Lisbon, Portugal, 2000. Memory-Based Learning for Article Generation Guido Minnen* Francis Bond t Cognitive and Computing Sciences MT Research Group University of Sussex NTT Communication Science Labs Palmer BN1 9QH, Brighton, UK 2-4 Hikari-dai, Kyoto 619-0237, JAPAN Guido. Minnen©cogs. susx. ac. uk bond©cslab, kecl. ntt. co. jp Ann Copestake CSLI Stanford University Stanford CA 94305-2150, USA aac©csli, stanford, edu Abstract Article choice can pose difficult problems in ap- plications such as machine translation and auto- mated summarization. In this paper, we investi- gate the use of corpus data to collect statistical generalizations about article use in English in order to be able to generate articles automati- cally to supplement a symbolic generator. We use data from the Penn Treebank as input to a memory-based learner (TiMBL 3.0; Daelemans et al., 2000) which predicts whether to gener- ate an article with respect to an English base noun phrase. We discuss competitive results ob- tained using a variety of lexical, syntactic and semantic features that play an important role in automated article generation. 1 Introduction Article choice can pose difficult problems in nat- ural language applications. Machine transla- tion (MT) is an example of such an applica- tion. When translating from a source language that lacks articles, such as Japanese or Rus- sian, to one that requires them, such as English or German, the system must somehow generate the source language articles (Bond and Ogura, 1998). Similarly in automated summarization: when sentences or fragments are combined or reduced, it is possible that the form of a noun phrase (NP) is changed such that a change of the article associated with the NP's head be- comes necessary. For example, consider the sen- tences A talk will be given on Friday about NLP; The talk will last .for one hour which might get summarized as Friday's NLP talk will last one * Visiting CSLI, Stanford University (2000). t Visiting CSLI, Stanford University (1999-2000). hour. However, given the input sentences, it is not clear how to decide not to generate an arti- cle for the subject NP in the output sentence. Another important application is in the field known as augmentative and alternative com- munication (AAC). In particular, people who have lost the ability to speak sometimes use a text-to-speech generator as a prosthetic de- vice. But most disabilities which affect speech, such as stroke or amyotrophic lateral sclerosis (ALS or Lou Gehrig's disease), also cause some more general motor impairment, which means that prosthesis users cannot achieve a text in- put rate comparable to normal typing speeds even if they are able to use a keyboard. Many have to rely on a slower physical interface (head- stick, head-pointer, eye-tracker etc). We are at- tempting to use a range of NLP technology to improve text input speed for such users. Article choice is particularly important for this applica- tion: many AAC users drop articles and resort to a sort of telegraphese, but this causes degra- dation in …

Content truncated to 500 words