Department of Psychology Department of Computer Science Institute for Intelligent Systems The University of Memphis The University of Memphis The University of Memphis Memphis, TN 38152, USA Memphis, TN 38152, USA Memphis, TN 38152, USA a-graesser@memphis.edu vrus@memphis.edu zcai@memphis.edu
Abstract Correction of knowledge deficits. This occurs
when there is an obstacle to a goal, a contradiction, One important first step in the Question Genera- an anomalous event, a glitch in an explanation, an tion campaign is to identify question taxonomies obvious gap in knowledge, or a decision required and analytical schemes that are grounded in between equally attractive alternatives. The person theory and empirical research. We present some experiences cognitive disequilibrium so a question of the proposed schemes in artificial intelligence, is asked to obtain information to restore equili- computational linguistics, discourse processes, education, and cognitive science. brium. Most sincere information seeking questions are in this category (van der Meij, 1987). 1 Introduction Monitoring common ground. Questions are asked to gauge, assess, confirm, or ratify what each Researchers in several fields have proposed other knows about a topic. schemes for classifying questions. An important Social coordination of action. These include in- first step in a Question Generation (QG) campaign direct requests, indirect advice, requests for per- is to take stock of the landscape of question cate- mission, and moves in bargaining. gories so that researchers can specify what types of Control of conversation and attention. These questions they have in mind, as well as the educa- include greetings, directives to change the speaker, tional context (Rus, Cai, & Graesser, 2007). rhetorical questions, gripes, and directives to focus Question taxonomies have been proposed by re- on an agent’s actions. searchers who have developed models of question asking and answering in the fields of artificial in- 3 Assumptions behind Questions telligence (Lehnert, 1978; Schank, 1986), compu- tational linguistics (Harabagiu, Maiorano, & Pasca, Van der Meij (1987) identified 11 assumptions that 2002; Voorhees, 2001), discourse processing all need to be true in order for a question to qualify (Graesser & Person, 1994; Graesser, Person, & as a sincere information seeking question. Huber, 1992), education (Beck, McKeown, Hamil- ton, & Kucan, 1997; Mosenthal, 1996) and a num- 1. The questioner does not know the information ber of other fields in the cognitive sciences (for a asked for with the question. reviews, see Graesser, Ozuru, and Sullins, in 2. The question specifies the information sought. press). 3. The questioner believes that the presupposi- This paper identifies some QG mechanisms, as- tions of the question are true. sumptions, categories, and dimensions that should 4. The questioner believes that an answer exists. help launch discussions about the landscape of 5. The questioner wants to know the answer. questions to consider in the QG campaign. 6. The questioner can assess whether a reply constitutes an answer. 2 Question Generation Mechanisms 7. The questioner poses the question only if the benefits exceed the costs. Graesser et al. (1992) identified four classes of 8. The questioner believes that the respondent psychological mechanisms that drive the genera- knows the answer. tion of questions. These are listed below. 9. The questioner believes that the respondent 3. Type of knowledge: Is the knowledge orga- will not give the answer in absence of a ques- nized as a semantic network, plan, causal struc- tion. ture, spatial layout, rule set, list of facts, etc.? 10. The questioner believes that the respondent 4. Cognitive process: Recognition, recall, com- will supply the answer. prehension, inference, application, synthesis, 11. A question solicits a reply. comparison, evaluation, etc.
4 Question Categories References
The following 16 question categories were either Graesser, A.C., Ozuru, Y., Sullins, J. (in press). What is proposed by Lehnert (1978) or by Graesser and a good question. In M. McKeown (Ed.), Festscrift for Isabel Beck. Mahwah, NJ: Erlbaum. Person (1994) in their analysis of tutoring. Cate- gories 1-4 were classified as simple/shallow, 5-8 as Beck, I.L., McKeown, M.G., Hamilton, R.L., & intermediate, and 9-16 as complex/deep questions Kucan, L. (1997). Questioning the Author: An in Graesser and Person’s empirical analyses of approach for enhancing student engagement with questions in educational settings. text. Delaware: International Reading Association. 1. Verification: invites a yes or no answer. Graesser, A. C., & Person, N. K. (1994). Question 2. Disjunctive: Is X, Y, or Z the case? asking during tutoring. American Educational 3. Concept completion: Who? What? When? Research Journal, 31, 104-137. Where? Graesser, A. C., Person, N., & Huber, J. (1992). 4. Example: What is an example of X? Mechanisms that generate questions. In T. Lau- 5. Feature specification: What are the proper- er, E. Peacock, & A. C. Graesser (Eds.), Ques- ties of X? tions and information systems. Hillsdale, NJ: 6. Quantification: How much? How many? Erlbaum. 7. Definition: What does X mean? Harabagiu, S.M. Maiorano, S.J. & Pasca, M.A. 8. Comparison: How is X similar to Y? (2002). Open-domain question answering tech- 9. Interpretation: What does X mean? niques. Natural Language Engineering, 1, 1-38. 10. Causal antecedent: Why/how did X occur? Lehnert, W. G. (1978). The process of question- 11. Causal consequence: What next? What if? answering. Hillsdale, NJ: Erlbaum. 12: Goal orientation: Why did an agent do X? Mosenthal, P. (1996). Understanding the strategies 13: Instrumental/procedural: How did an agent of document literacy and their conditions of use. do X? Journal of Educational Psychology, 88, 314- 14: Enablement: What enabled X to occur? 332. 15: Expectation: Why didn’t X occur? Rus, V., Cai, Z., Graesser, A.C. (2007). Evaluation 16: Judgmental: What do you think of X? in Natural Language Generation: The Question Generation Task. It should be noted that sometimes a question can Schank, R.C. (1986). Explanation patterns: be a hybrid between two categories. Understanding mechanically and creatively. Hillsdale, NJ: Erlbaum. 5 Other Dimensions of Questions Van der Meij, H. (1987). Assumptions if informa- tion-seeking questions. Questioning Exchange, Some other dimensions of questions are frequently 1, 111-118. addressed in classification schemes, as discussed Voorhees, E. (2001). The TREC Question Answer- by Graesser et al. (in press). ing Track. Natural Language Engineering, 7, 361-378. 1. Information sources. Does the answer come from a text, world knowledge, both, elsewhere? 2. Length of answer: Is the answer a single word, a phrase, a sentence, or a paragraph?