Cogex: A semantically and contextually enriched logic prover for question answering

This paper presents the architecture and functionality of a logic prover designed for question answering. The approach transforms questions and answer passages into logic representations based on syntactic, semantic and contextual information. World knowledge supplements the linguistic, ontological, and temporal axioms supplied to the prover which renders a deep understanding of the relationship between the question and answer text. The trace of the proofs provides a basis for generating human comprehensible answer justifications. The results show that the prover boosts the performance of the Question Answering system on TREC 2004 questions by 12%.     

TSUNAMI: Simultaneous Understanding,Answering, and Memory Interaction for Questions

Question processing involves parsing, memory retrieval, question categorization, initiation of appropriate answer-retrieval heuristics, answer formulation, and output. Computational and psychological models have traditionally treated these processes as separate, sequential, independent, and in pursuit of a single answer type at a time. Here this view is challenged and the implications of a theory in which question processes operate simultaneously on multiple question interpretations are explored. A highly interactive model is described in which an expectation-driven parser generates multiple question candidates, including partially-specified candidates. Question candidates act as constraints for a matcher which activates memory items. An answer retrieval process examines question candidates and the active portions of memory in an attempt to generate answer candidates. Answer candidates are examined by an output process that derives the final answer. These processes run simultaneously and interact. Three experiments on human question answering are also described which provide evidence that working memory load during question reading is affected by processes related to answer retrieval.     

Unpriming: the deactivation of thoughts through expression

Unpriming is a decrease in the influence of primed knowledge following a behavior expressing that knowledge. The authors investigated strategies for unpriming the knowledge of an answer that is activated when people are asked to consider a simple question. Experiment 1 found that prior correct answering eliminated the bias people normally show toward correct responding when asked to answer yes-no questions randomly. Experiment 2 revealed that prior answering intended to be random did not unprime knowledge on subsequent attempts to answer randomly. Experiment 3 found that exposure to the correct answer did not influence the knowledge bias but that exposure to the incorrect answer increased bias. Experiment 4 revealed that merely expressing the answer for oneself was sufficient to unprime knowledge. Experiment 5 found that each item of activated knowledge needs to be unprimed specifically, in that correctly answering 1 question does not reduce the knowledge bias in randomly answering another.     

Data-driven type checking in open domain question answering

Many open domain question answering systems answer questions by first harvesting a large number of candidate answers, and then picking the most promising one from the list. One criterion for this answer selection is type checking: deciding whether the candidate answer is of the semantic type expected by the question. We define a general strategy for building redundancy-based type checkers, built around the notions of comparison set and scoring method, where the former provide a set of potential answer types and the latter are meant to capture the relation between a candidate answer and an answer type. Our focus is on scoring methods. We discuss nine such methods, provide a detailed experimental comparison and analysis of these methods, and find that the best performing scoring method performs at the same level as knowledge-intensive methods, although our experiments do not reveal a clear-cut answer on the question whether any of the scoring methods we consider should be preferred over the others.     

Influence of Question Format and Text Availability on the Assessment of Expository Text Comprehension

This study examines how passage availability and reading comprehension question format (open-ended vs. multiple-choice) influence question answering. In two experiments, college undergraduates read an expository passage and answered open-ended and multiple-choice versions of text-based, local, and global bridging inference questions. Half the participants were allowed to refer to the passage when answering the questions and half were not. Participants' prior domain knowledge relating to the text contents was assessed using multiple-choice and open-ended questions. Correlation-based analyses in the two experiments indicated: (a) a decline in the relationship between prior domain knowledge and comprehension when the passage was available during question answering; and (b) a high correlation between multiple-choice and open-ended question answering performance when the passage was not available for reference. Overall the results indicate that the nature of the reading comprehension assessment is influenced by the specific task with which comprehension is assessed.     

A metaphysics for practical knowledge

Is Anscombean practical knowledge independent of what the agent actually does on an occasion? Failure to understand Anscombe’s answer to this question is a major obstacle to appreciating the subtlety and plausibility of her view. I argue that Anscombe’s answer is negative, and turns on the nature of mistakes in performance, and reveals a distinctive implicit metaphysics of mind and knowledge, structured by related capacities and exercises of capacities. If my interpretation is correct, then practical knowledge shares features with knowledge-how and knowledge-that, but deserves its own epistemic category.     

Answering yes-no questions about causes: Question acts and question categories

Three experiments examined the cognitive process of answering yes-no questions about causes. Singer’s VAIL model of question answering predicted that readers would take longer to correctly answer “no” than “don’t know” to such questions. In Experiment 1, the antecedent sentences used either the causal conjunction so orbecause. Experiment 2 compared so with an implicit causal link. In all conditions, the main prediction was strongly supported. However, when the questions referred to brief stories in Experiment 3, correct “no” and “don’t know” response latencies did not differ. It was concluded that (1) VAIL identifies the cognitive operations underlying the answering of causal questions; (2) answering yes-no questions about causes resembles answering yes-no questions about case-filling elements; (3) the yes-no versus wh- distinction is orthogonal to the type of relation asked about; and (4) studying question answering about sentences will contribute to the understanding of question answering about text.     

Response selection strategies and realism of confidence judgments

Two studies were conducted to examine how response selection strategy is related to confidence ratings and to performance on general knowledge questions. In both studies subjects were asked to answer 80 general knowledge questions and to rate their confidence in the correctness of the answer selected. A pilot study, in which subjects thought aloud while answering general knowledge questions, was carried out to identify different response selection strategies. In the first study, 40 subjects were asked to indicate which of four strategies (immediate recognition, inference, intuition, or guessing) they used for selecting an answer. In Study 2, think aloud reports from 20 subjects were coded into the same four strategies. The distribution of strategies differed between the studies, but there were very similar relations among strategy, confidence, and correctness of answer in the two studies. Response selection strategy was related to correctness of answer when confidence was partialed out. More specifically, immediate recognition was associated with higher proportion correct than with the other strategies. It was also found that ratings of how difficult the knowledge questions were to fellow students of the subjects were on a much more realistic level than the confidence ratings were. It is concluded that people could improve their confidence judgments by taking into account (a) how difficult a question is to other people, and (b) the response selection strategy used for answering the question.     

Domain-specific learning of grammatical structure in musical and phonological sequences

Artificial grammar learning depends on acquisition of abstract structural representations rather than domain-specific representational constraints, or so many studies tell us. Using an artificial grammar task, we compared learning performance in two stimulus domains in which respondents have differing tacit prior knowledge. We found that despite grammatically identical sequence structures, learning was better for harmonically related chord sequences than for letter name sequences or harmonically unrelated chord sequences. We also found transfer effects within the musical and letter name tasks, but not across the domains. We conclude that knowledge acquired in implicit learning depends not only on abstract features of structured stimuli, but that the learning of regularities is in some respects domain-specific and strongly linked to particular features of the stimulus domain.     

ALTERNATIVE QUESTIONS AND KNOWLEDGE ATTRIBUTIONS

We discuss the 'problem of convergent knowledge', an argument presented by J. Schaffer in favour of contextualism about knowledge attributions, and against the idea that knowledge- wh can be simply reduced to knowledge of the proposition answering the question. Schaffer's argument centrally involves alternative questions of the form 'whether A or B'. We propose an analysis of these on which the problem of convergent knowledge does not arise. While alternative questions can contextually restrict the possibilities relevant for knowledge attributions, what Schaffer's puzzle reveals is a pragmatic ambiguity in what 'knowing the answer' means: in his problematic cases, the subject knows only a partial answer to the question. This partial knowledge can be counted as adequate only on externalist grounds.     

A contrastive account of explanation generation

In this article, we propose a contrastive account of explanation generation. Though researchers have long wrestled with the concepts of explanation and understanding, as well as with the procedures by which we might evaluate explanations, less attention has been paid to the initial generation stages of explanation. Before an explainer can answer a question, he or she must come to some understanding of the explanandum—what the question is asking—and of the explanatory form and content called for by the context. Here candidate explanations are constructed to respond to the particular interpretation of the question, which, according to the pragmatic approach to explanation, is constrained by a contrast class—a set of related but nonoccurring alternatives to the topic that emerge from the surrounding context and the explainer’s prior knowledge. In this article, we suggest that generating an explanation involves two operations: one that homes in on an interpretation of the question, and a second one that locates an answer. We review empirical work that supports this account, consider the implications of these contrastive processes, and identify areas for future study.     

Undermining Motivations for Universalism

Universalism (the thesis that for any ys, those ys compose a further object) is an answer to the Special Composition Question. In the literature there are three arguments – what I call the arguments from elegance – that universalists often rely upon, but which are rarely examined in‐depth. I argue that these motivations cannot be had by the perdurantist, for to avoid a commitment to badly behaved superluminal objects perdurantists must answer the ‘Proper Continuant Question’. Any answer to that question necessarily ensures that there is a restricted answer to the Special Composition Question that is just as elegant as universalism. Thus, if you are a perdurantist, the arguments from elegance fail to motivate universalism for there will always be a restricted composition that is just as good.     

Answer comparison in automated question answering

In the context of Question Answering (QA) on free text, we assess the value of answer comparison and information fusion in handling multiple answers. We report improvements in answer re-ranking using fusion on a set of location questions and show the advantages of considering candidates as allies rather than competitors. We conclude with some observations about answer modeling and evaluation methodology, arising from a more recent experiment with a larger set of questions and a greater diversity of question types and candidates.     

Six-year-olds' difficulties handling intensional contexts

Previous research shows that children have difficulties handling intensional contexts even when they can pass a test of false belief (e.g. Cognition 67 (1998) 287; Cognition 25 (1987) 289). Some authors (Perner, J. (1991). Understanding the representational mind. Cambridge, MA: MIT Press; Cognition 25 (1987) 289) place these difficulties in the linguistic and not the mental representational domain. The experiments reported here examined whether 6-year-old children could answer questions in an intensional context that did not require the explicit verbal characterization of a belief. We replicated previous findings and found that children answered according to their own knowledge in an intensional context. This occurred even though they responded by choosing a picture to insert into a protagonist's thought bubble rather than report the belief verbally. Children could correctly answer questions about the knowledge state of the protagonist and pass a test of false belief. Further experiments ruled out methodological explanations. Experiment 2 showed that the difference in answering according to own knowledge between the false belief and intensional stories is not accounted for by procedural factors in the two types of test. Experiment 3 revealed that children did not answer according to their own knowledge by default. Experiment 4 suggested that answering according to own knowledge was not a result of pictorial salience. Results are discussed in relation to the simulation-theory debate.     

Domain-specific conflict adaptation without feature repetitions

An influential account of how cognitive control deals with conflicting sources of information holds that conflict is monitored by a module that automatically recruits attention to resolve the conflict. This leads to reduced effects of conflict on the subsequent trial, a phenomenon termed conflict adaptation. A prominent question is whether control processes are domain specific—that is, recruited only by the particular type of conflict they resolve. Previous studies that have examined this question used two-choice tasks in which feature repetition effects could be responsible for domain-specific adaptation effects. We report two experiments using four-choice (Experiment 1) and five-choice (Experiment 2) tasks that contain two types of irrelevant sources of potentially conflicting information: stimulus location (Simon conflict) and distractors (flanker conflict). In both experiments, we found within-type conflict adaptation for both types of conflict after eliminating trials on which stimulus features were repeated from one trial to the next. Across-type conflict adaptation, however, was not significant. Thus, conflict adaptation was due to domain-specific recruitment of cognitive control. Our results add converging evidence to the idea that multiple independent control processes are involved in reactive cognitive control, although whether control is always local remains to be determined.     

Priming and communication: Social determinants of information use in judgments of life satisfaction

Two experiments examined the effects of answering a question about a specific component of life satisfaction on respondents' assessment of their overall satisfaction with life. The results suggest that the use of primed information in forming subsequent judgments is determined by Grice's conversational norms. In general, answering the specific question increases the accessibility of information relevant to that question. However, the effect that this has on the general judgment depends on the way in which the two questions are presented. When the two questions are merely placed in sequence without a conversational context, the answer to the subsequent general question is based in part on the primed specific information. As a result, the answer to the general question becomes similar to that for the specific question (i.e. assimilation). However, this does not occur when the two questions are placed in a communication context. Conversational rules dictate that communicators should be informative and should avoid redundancy in their answers. Therefore, when a specific and a general question are perceived as belonging to the same conversational context, the information on which the answer to the specific question was based is disregarded when answering the general one. This attenuates the assimilation effect. The conditions under which these different processes occur are identified and experimentally manipulated, and the implications of these findings for models of information use in judgment are discussed.     

A comprehensive semantic framework for data integration systems

A data integration system provides the user with a unified view, called global schema, of the data residing at different sources. Users issue their queries against the global schema, and the system computes answers to queries by suitably accessing the sources, through the mapping, i.e., the specification of the relationship between the global schema and the sources. Since sources are in general autonomous subsystems, the information provided by the data at the sources and the mapping is likely not to be consistent with the knowledge expressed by the global schema. Therefore, the question arises of how to interpret user queries in such a situation, i.e., in the presence of data contradicting the global schema and the mapping. In this paper, we provide an in-depth analysis of the problem of dealing with inconsistencies in data integration systems. In this respect, we highlight the central role played by the mapping, and propose a general “mapping-centered” semantics that allows for computing significant answers to user queries even in the presence of inconsistent information. Based on such a semantic analysis, we define a general formal framework for data integration. Then, we argue that our semantic approach formalizes a very reasonable way of handling inconsistency in such systems, since practically all the existing proposals in the literature can be reconstructed in our framework. This allows for comparing and evaluating the different existing proposals.     

Defining answer classes using resolution refutation

Resolution theorem proving provides a useful paradigm for the exploration of question answering. A partition of the clauses generated during resolution refutation based on their syntactic structure is presented. The three classes comprising this partition correspond to semantically intuitive types of answers. This work encompasses and expands upon previous work on question answering in a theorem proving paradigm, which began with the association of answers with proofs. A complete, formal definition of what is meant by answer in the context of resolution theorem proving is presented. In this context, clauses that are relevant are all identified as answers, where relevance is determined with respect to a question and knowledge base: any clause descended from the clause form of a negated question is deemed relevant. This definition of relevance is not in and of itself novel; rather, it is the way in which the set of relevant clauses is partitioned that provides the key to interpreting clauses as answers. The three answer classes identified are: specific, generic, and hypothetical. These classes are formally distinguished by the way in which literals in a clause share variables, with class membership based on a property termed the closure of variable sharing of a literal. The results presented provide a foundation for further work by establishing a context-independent logical pragmatics of question answering.     

The generation effect extended: Memory enhancement for generation cues

The generation effect is the greater memorability of a response that is actively produced (e.g., in answering a question from memory) than one that is more passively produced (as in reading the answer). The present three experiments addressed a question that is critical to the theoretical interpretation of the generation effect: Is memory enhanced for the cues that are used to elicit generated responses? Using incidental learning procedures, Experiments 1 and 2 gave an affirmative answer (although the effect was substantially weaker than the generation effect for responses). Enhancement of memory for generation cues was observed both in a within-subject/within-list design (reading and generation items within the same trial blocks; Experiment 1) and in a between-subjects design (reading and generation tasks for different groups of subjects; Experiment 2). In Experiment 3, memory enhancement for generation cues was used to produce a previously unobtained result—a generation effect for nonsense responses under incidental learning conditions. These findings provide critical evidence required by theories that interpret the generation effect in terms of enhanced processing of the cue-response item.