Automatically summarizing vast amounts of on-line quantitative data with a short natural language paragraph has a wide range of real-world applications. However, this specific task raises a number of difficult issues that are quite distinct from the generic task of language generation: conciseness, complex sentences, floating concepts, historical background, paraphrasing power and implicit content. In this thesis, I address these specific issues by proposing a new generation model in which a first pass builds a draft containing only the essential new facts to report and a second pass incrementally revises this draft to opportunistically add as many background facts as can fit within the space limit. This model requires a new type of linguistic knowledge: revision operations, which specifyies the various ways a draft can...

Recent years have witnessed rapid progress in explanation generation. Despite these advances, the quality of prose produced by explanation generators warrants significant improvement. Revision-based explanation generation offers a promising means for improving explanations at runtime. In contrast to singledraft explanation generation architectures, a revision-based generator could dynamically create, evaluate, and refine multiple drafts of explanations. However, because of the inherent complexity of revision, previous multisentential revision-based approaches have not scaled up. We have developed a scalable revision-based model of explanation generation that dynamically improves multi-sentential explanations. By operating on abstract discourse plans encoded in a minimalist representation, it combats both the conceptual complexities and the efficiency problems posed by revision. This approach has been implemented in REVISOR, a unification-based revision system. Evaluations of REVISOR'S performance in generating a corpus of extended multi-sentential scientific explanations yielded encouraging results. 1

this paper is that neither has been implemented." ([Vaughan and McDonald, 1986], page 95). Although Meteer [1990] gives a detail description of the relationship between text structure and revision it is unclear how the proposed model could contribute to the choice problem of paraphrases (see section 5.2). How- ever, from the approach described above and from the system described in [Meteer, 1990] we can draw the following conclusions. Only the generatoFs input is marked. If the generator encounters alternative realizations the revision component is asked to make the decision. However, to be able to do this it needs detailed knowledge about the grammar. Therefore grammatical knowledge has to be duplicated. The linguistic realization component used in [Meteer, 1990] is MUMBLE-86 [McDonald, 1986]. The text structural representation level must completely specify the infor- mation to be expressed by the utterance. The mapping has to ensure that all the necessary linguistic information is present. Mumblers procedural grammar is used only for generation purposes. Therefore it is without reach for the revision model to take into account relevant sources of ambiguities

Natural language reports generated by existing systems ignore the historical context of the facts and events they relate. In this paper, I argue that going beyond this limitation requires abandoning the pipelined architecture of existing report generators. I propose a new architecture in which a first draft of the report is organized around new information and then incrementally revised to opportunistically add related historical information. This type of information-adding revision allows to elaborate inside a clause or a nominal while taking into account surface structure constraints from any other portion of the report. In addition to providing the additional flexibility required to convey historical information, the proposed architecture constitutes an interesting testbed to investigate a wide range of open questions concerning content planning below the sentence level, generation with revision and generation architecture. 1 Introduction: generating reports in their historical co...

We present anew model of natural language processing in which natural language parsing and generation are strongly interleaved tasks. Interleaving of parsing and generation is important if we assume that natural language understanding and production are not only performed in isolation but also work together to obtain subsentential interactions in text revision or dialog systems. The core of the model is a new uniform agenda-driven tabular algorithm, called UTA. Although uniformly de ned, UTA is able to con gure itself dynamically for either parsing or generation, because it is fully driven by thestructure of the actual input|a string for parsing and a semantic expression for generation. Efficient interleaving of parsing and generation is obtained through item sharing between parsing and generation. This novel processing strategy facilitates the automatic exchange of items (i.e., partial results) computed in one direction to the other direction as well. The advantage of UTA in combination with the item sharing method is that we are able to extend the use of memorization techniques to the case of an interleaved approach. In order to demonstrate UTA's utility for developing high-level performance methods, we present a new algorithm for incremental self-monitoring during natural language production.

We have begun to see the emergence of large-scale knowledge bases that house tens of thousands of facts encoded in expressive representational languages. The richness of these representations o er the promise of significantly improving the quality of natural language generation, but their representational complexity, scale, and task-independence pose great challenges to generators. We have designed, implemented, and empirically evaluated Fare, a functional realization system that exploits message specifications drawn from large-scale knowledge bases to create functional descriptions, which are expressions that encode both functional information (case assignment) and structural information (phrasal constituent embeddings). Given a message specification, Fare exploits lexical and grammatical annotations on knowledge base objects to construct functional descriptions, which are then converted to text by a surface generator. Two empirical studies -- one with an explanation generator and one with a qualitative model builder -- suggest that Fare is robust, efficient, expressive, and appropriate for a broad range of applications.

In this paper, I present a model of the local organization of extended text. I show that texts with weak rhetorical structure and strong domain structure, such as descriptions of houses, digital circuits, and families, are best analyzed in terms of local domain structure, and argue that 1obal structures that may be inferred from a domain are not always appropriate for constructing descriptions in the domain. I present a system I am implementing that uses short-rasge struteies to organize text, and show how part of a description is organized by these strategies. I also briefly discuss a model of incremental text generation that dovetails with the model of local organization presented here.

In this paper I investigate the issue of providing historical background in computer-generated reports. I first observe that ignoring this issue is the most drastic limitation of existing report generation systems. I then present an empirical corpus analysis of basketball summaries aimed at discovering the specific means by which historical information is conveyed in human-generated reports. This analysis resulted in a set of data that forms the basis for the implementation of streak, a system generating basketball report leads with historical information. This data shows that building such a system cannot be based on the single-pass pipelined architecture of previous systems. Instead, I propose an entirely new architecture in which generation proceeds in two passes. The first pass builds a report draft containing only the basic facts. The second pass incrementally revises this draft to include additional facts providing the historical background. Independently of the issue of...

This paper describes the use of a system of semantic rules to generate noun compounds, vague or polysemous words, and cases of metonymy. The rules are bidirectional and are used by the understanding system to interpret the same constructions.