Influences on Attribute Selection in Redescriptions: A Corpus Study
Influences on Attribute Selection in Redescriptions: A Corpus Study
复制标题
重新描述中属性选择的影响：语料库研究
DOI：
--
复制
发表时间：
2000
期刊：
影响因子：
--
通讯作者：
Pamela W. Jordan
中科院分区：
文献类型：
--
作者：
Pamela W. Jordan
对重新描述中属性选择的影响：语料库研究 Pamela W. Jordan (jordan@isp.pitt.edu) 智能系统项目 匹兹堡大学 Pittsburgh PA 15260 摘要 我们报告了语料库分析第一阶段的结果，该分析测试了关于领域和话语目标以及重复功能如何影响对话中重新描述内容的五个假设。我们发现重新描述中表达的属性与预测其中三种目标的上下文之间存在正相关性。这些结果为我们将在下一阶段的分析中测试的选择策略提供了指导。引言 在扩展的话语中，说话者经常重新描述之前介绍过的对象，以便更多地谈论该对象或其参与的事件。重新描述实体时的主要目标是重新唤起适当的话语实体。然而，句子生成的目标导向观点表明，说话者可以尝试用每个话语来满足多个目标[Appelt，1985]，并且单一语言形式可以机会主义地有助于满足多个目标[Stone and Webber，1998]。在生成名义表达式的计算工作中，尚未完全解决除识别之外的目标可能影响名义表达式内容的可能性。目标到语言形式的多一映射通常被称为超载意图 [Pollack, 1991]。重载可能涉及跨语言级别的权衡[Di Eugenio and Webber, 1996, Stone and Webber, 1998]。例如，通过在语义层面上使形式复杂化来实现的意图可能允许说话者通过省略重要信息在句法层面上进行简化[Stone and Webber，1998]。尽管我们已经了解到超载是自然的，甚至可能是必要的，但对于什么程度的超载是合理的以及什么形式可以更容易地解决对话中的多个目标，我们没有得到充分支持的解释。如果没有这样的帐户，我们就没有原则性的方法来在自然语言的自动生成中部署重载。如果没有对过载的良好支持约束，我们很容易以不自然的方式产生过载，这实际上会阻碍有效的沟通。为了调查重载是否适用于重新描述，我们检查了在 13 个关于 COCONUT 标识的对话中发现的 166 个非代词重新描述，满足多个名义表达也值得考虑。语料库 [Di Eugenio et al., 2000]。该语料库包含计算机介导的对话，其中两个人合作完成一项简单的设计任务，为房子的两个房间购买家具。参与者的主要目标是谈判采购；最优先的物品是客厅的沙发和餐厅的一张桌子和四把椅子。参与者还有特定的次要目标，这使得解决问题的任务进一步复杂化。参与者被指示尝试尽可能多地实现这些目标，并通过将分数与满意的目标联系起来来激励他们这样做。第二个目标是：（1）匹配房间内的颜色，（2）购买尽可能多的家具，（3）花掉所有的钱。每个参与者都会获得单独的预算和家具库存，并且必须决定要让哪些内容相互了解。库存中的每件家具都由五个属性描述；类型、颜色、价格、所有者和数量。在本文中，我们报告了语料库分析第一阶段的结果，该分析测试了关于领域和话语目标以及重复功能如何影响对话中重新描述内容的五个假设。我们发现重新描述中表达的属性与预测其中三种类型目标的上下文之间存在正相关性。正相关性将指导我们实施和测试生成名义表达式的选择策略。关于对重新描述的影响的假设我们的假设反映了可能影响重新描述属性选择的非识别目标。这些目标源自对话语或命题层面的重复功能的研究[Walker, 1993, Johnstone, 1994]，以及对任务意图和约束变化（例如匹配颜色）的观察，而这些观察并未由对话参与者直接传达。我们的第一个假设是基于在 COCONUT 语料库中的观察，人们经常在没有明确讨论的情况下调整任务约束（38%）？与命题层面的[Walker，1993]类似，我们建议听者从重新描述的冗余中推断出变化。因此，重复属性可以在规划和调度任务以及设计任务中找到2个约束[Di Eugenio et al., 2000, Jordan, 2000]。
Influences on Attribute Selection in Redescriptions: A Corpus Study Pamela W. Jordan (jordan@isp.pitt.edu) Intelligent Systems Program University of Pittsburgh Pittsburgh PA 15260 Abstract We report on the results of the first stage of a corpus analysis that tests five hypotheses about how domain and discourse goals and the functions of repetition can influence the content of redescriptions in dialogue. We found a positive correlation between the attributes ex- pressed in redescriptions and contexts in which three of these types of goals are predicted. These results provide us with guidance on the selection strategies we will test in the next stage of analysis. Introduction In an extended discourse, speakers often redescribe ob- jects that were introduced earlier in order to say some- thing more about the object or the event in which it participates. The main goal when redescribing an entity is to re—evoke the appropriate discourse entity. How- ever, a goal—directed view of sentence generation suggests that speakers can attempt to satisfy multiple goals with each utterance [Appelt, 1985] and that a single linguistic form can opportunistically contribute to the satisfaction of multiple goals [Stone and Webber, 1998]. The pos- sibility that goals besides identification could influence the content of a nominal expressionl have not been fully addressed in computational work on generating nominal expressions. The many—one mapping of goals to linguistic forms is more generally referred to as overloading inten- tions [Pollack, 1991]. Overloading can involve tradeoff across linguistic levels [Di Eugenio and Webber, 1996, Stone and Webber, 1998]. For example, an intention which is achieved by complicating a form at the se- mantic level may allow the speaker to simplify at the syntactic level by omitting important information [Stone and Webber, 1998]. Although we have learned that overloading is natural and perhaps even necessary, we have no well supported account of what degree of overloading is reasonable and what forms can more readily address multiple goals in di- alogue. Without such an account, we have no principled way to deploy overloading in the automatic generation of natural language. Without well supported constraints on overloading, we are liable to create overloads in un- natural ways which will actually impede effective com- munication. To investigate whether overloading applies to re- descriptions, we examined 166 non—pronominal re- descriptions found in 13 dialogues of the COCONUT lldentification being satisfied by more than a nominal ex- pression also deserves consideration. corpus [Di Eugenio et al., 2000]. This corpus contains computer—mediated dialogues in which two people col- laborate on a simple design task, buying furniture for two rooms of a house. The participants’ main goal is to negotiate the purchases; the items of highest priority are a sofa for the living room and a table and four chairs for the dining room. The participants also have specific secondary goals which further complicate the problem solving task. Participants are instructed to try to meet as many of these goals as possible, and are motivated to do so by associating points with satisfied goals. The sec- ondary goals are: (1) Match colors within a room, (2) Buy as much furniture as you can, (3) Spend all your money. Each participant is given a separate budget and inventory of furniture and must decide what to make mutually known. Every furniture item in the inventory is described by five attributes; type, color, price, owner, and quantity. In this article, we report on the results of the first stage of a corpus analysis that tests five hypotheses about how domain and discourse goals and the functions of repeti- tion can influence the content of redescriptions in di- alogue. We found a positive correlation between the attributes expressed in redescriptions and contexts in which three of these types of goals are predicted. The positive correlations will guide us in implementing and testing selection strategies for generating nominal ex- pressions. Hypotheses about influences on redescriptions Our hypotheses refiect non—identification goals that could influence the choice of attributes for a redescrip- tion. These goals are derived from work on the func- tions of repetition at the utterance or propositional level [Walker, 1993, Johnstone, 1994] and from observa- tions about task intentions and constraint changes (e.g. matching colors) that were not directly communicated by the dialogue participants. Our first hypothesis is based on the observation that in the COCONUT corpus, people often adjusted task constraints with no explicit discussion (38%)? Similarly to [Walker, 1993] for the propositional level, we suggest that the hearer is inferring changes from the redundan- cies in the redescription. So the repeated property could 2Constraints are found in planning and scheduling tasks as well as design tasks [Di Eugenio et al., 2000, Jordan, 2000].