Interpretation of Conditionals in the Suppression Task. Andrea Lechler

Transcription

1 Interpretation of Conditionals in the Suppression Task Andrea Lechler Master of Science Artificial Intelligence School of Informatics University of Edinburgh 2004

2 Abstract If people are presented with a conditional sentence and a corresponding categorical premise, they easily make modus ponens inferences. But several studies have shown that this valid inference of classical logic can be suppressed by adding another conditional sentence containing an additional necessary condition. Although most authors accounting for this suppression of inferences have stressed the importance of interpretative processes, it has not been investigated sufficiently how subjects really interpret such conditional sentences in practice. Thus, this dissertation examines people s interpretations in interviews, utilising statements similar to those used in previous studies. The results indicate that there is a wider range of possible interpretations than assumed in accounts of the suppression effect in the literature, and that subjects do not always adopt the interpretations expected by experimenters. Moreover, they show that there is often no clear-cut distinction between interpretation and reasoning. This study investigates the influence of other factors like participants understanding of the task, character or the use of world knowledge. It also examines the role of background or abnormality conditions and counterexamples for the meaning of conditionals. The findings are used to assess other ways of accounting for the suppression effect and conditional reasoning in general, like the theory of mental models or probabilistic approaches. i

3 Acknowledgements First, I would like to thank my supervisor Keith Stenning for his helpful advice and many fruitful discussions, and Levente Barczy for the good cooperation in doing interviews for our projects. I want to express my appreciation to Wiebke Herding and Daniel Dennis for their useful comments on my thesis and all other forms of support they gave me during the project. I am also grateful to the DAAD (German Academic Exchange Service) for funding my MSc course. ii

4 Declaration I declare that this thesis was composed by myself, that the work contained herein is my own except where explicitly stated otherwise in the text, and that this work has not been submitted for any other degree or professional qualification except as specified. (Andrea Lechler) iii

5 Table of Contents 1 Introduction The suppression effect in conditional reasoning Objectives Terminology Method Motivation Design Procedure Evaluation Different kinds of interpretations Interpretations considered in previous studies A new categorisation Independent statements Alternative conditions True antecedents Correction Summary Subjects understanding of the suppression task The role of instructions Interpretation Reasoning iv

6 5 Involvement of world knowledge 55 6 Conditionals and counterexamples 62 7 Uncertainty in conditional reasoning Uncertainty of the antecedent True antecedents? Different types of assertion How can they be so sure? Individual patterns of interpretation and reasoning Effects of different experimental settings Individual differences Mental abilities Personality Biconditionals Conclusion The suppression effect Why did we observe less suppression? Why was there such a wide range of different conclusions? Is there really suppression? General findings Future work Bibliography 111 v

7 Chapter 1 Introduction 1.1 The suppression effect in conditional reasoning Everybody who has at least some knowledge of formal logic knows inference rules like modus ponens: A B A B From the fact that A implies B and A holds, we can conclude that B holds. Many psychologists have claimed that modus ponens is not only a rule of classical logic, but is also employed in human reasoning, i.e. even people who have never heard of formal logic have a corresponding mental rule. This poses the question as to which representations this mental rule would operate on. In classical logic, modus ponens is formulated for material implication, which is often translated into natural language by means of if...then sentences. Indeed, modus ponens is often defined as an argument of the form: if A then B, and A, therefore B. Does this mean that mental inference rules like modus ponens operate on conditional sentences, or more precisely on their mental representation? And what exactly is the meaning or the mental representation of conditionals? Byrne (1989) questioned these views by showing that the valid inferences modus ponens and modus tollens 1 can be suppressed if more context is provided. 2 1 If A then B, not B, therefore not A 2 The kind of experiment she deploys is hence called the suppression task. Before, 1

8 Chapter 1. Introduction 2 others had claimed that fallacies, such as denial of the antecedent 3 and affirmation of the consequent 4, are not mental rules, but only implicatures or invited inferences (Rumain et al. (1983); Markovits (1984); Markovits (1985)). They had showed that these inferences are often not made if a second conditional is added to the statement. Byrne replicated these results by presenting a second conditional containing an alternative condition for the same consequent 5, as in the following statement: If she has an essay to write then she will study late in the library. If she has some textbooks to read then she will study late in the library. She does not have an essay to write. Only a few subjects concluded that she will not study late in the library. Byrne extended the experiment by introducing a second conditional with an additional requirement that must also hold (p. 67): If she has an essay to write then she will study late in the library. If the library stays open then she will study late in the library. She has an essay to write. Now only 38% of subjects made modus ponens inferences, compared to 96% when a simple conditional or two conditionals with alternative conditions were presented. Byrne concludes that either there are no mental rules for the valid inferences, or that suppression by itself tells us nothing about the existence or non-existence of rules of inference in the mind (Byrne, 1989, p. 76) She admits that the results can still be explained in terms of mental rules if one assumes that the joint representation of both sentences makes the application of inference rules impossible. Her proposal is that Formal theories, therefore, need to be supplemented with a detailed account of the process of interpretation, because premises of the same apparent logical form are represented in different ways depending on their meaning. (Byrne, 1989, p. 77) 3 If A then B, not A, therefore not B 4 If A then B, B, therefore A 5 Although some scholars reject the terms antecedent and consequent, I use them to refer to an if-clause and the corresponding main clause. They make talking about conditional sentences easier, and do not indicate preferences for any particular approach.

9 Chapter 1. Introduction 3 Yet, she rejects mental rules in favour of the mental model theory (see e.g. Johnson- Laird and Byrne (1991)). Here, the idea is that people represent all possible states of affairs which are consistent with their world knowledge and the content of the sentences. A conclusion is only endorsed if there is no model which contradicts it. Byrne explains the suppression effect by assuming that introducing alternative or additional conditions results in subjects constructing different sets of models, which provide different kinds of counterexamples. In subsequent papers (e.g. Byrne (1991); Byrne et al. (1999); Johnson-Laird and Byrne (2002)) Byrne refined her approach, but still stuck to these basic ideas. This suppression effect described by Byrne has become one of the phenomena every psychological theory of conditional reasoning has to account for. 6 As it has not been considered in philosophical and linguistic approaches to conditionals, the following review discusses only psychological studies. However, it is worth noting that the use of language, in this case conditionals, in an artificial situation like Byrne s experiments can be interesting for linguists, too. Moreover, linguists tend to neglect the close relationship between conditionals and reasoning. My account of the suppression effect will try to bridge the gap between psychological and linguistic approaches. Let us now examine what has been said about the suppression effect in cognitive psychology. Many scholars accepted the mental model theory and made only slight modifications to it (e.g. Dieussaert et al. (1999); Dieussaert et al. (2000)). There were, however, many others who criticised Byrne s approach and offered new explanations for the suppression effect. Politzer and Braine (1991) defend their syntactic theory of mental logic, and claim that, in contrast to the fallacies, the valid inferences are suppressed due to inconsistencies within the premise set. Subjects cannot draw a conclusion because the second conditional casts doubt on the truth of the first. They also show how pragmatic processes yield the combined representation proposed by Byrne (1989). Although proponents of both mental logic (Politzer and Braine (1991)) and the mental model theory (Byrne (1989)) emphasise the importance of interpretation, Chan and Chua (1994) claim that both theories have not offered a principled account of 6 Others are syllogistic reasoning or Wason s selection task.

10 Chapter 1. Introduction 4 the interpretive processes critical in reasoning (p. 233) and do not consider content and context effects sufficiently. According to them, people s conclusions depend on the relative salience or degree of necessity they assign to an additional requirement by means of their world knowledge (or more precisely by so called reasoning schemas ), and on the order in which the sentences are presented. They observe an increase in suppression rate with increase in relative salience (p. 231), and suggest a cognitive model based on production rules which only allow a conclusion if the relative salience of the second condition is low enough. Other scholars point out that uncertainty of beliefs is an inherent feature of human reasoning and must be accounted for by a theory of reasoning. Stevenson and Over (1995) claim that adding an additional requirement to a conditional can make it uncertain by decreasing the conditional probability of the consequent given the antecedent, which leads to uncertain conclusions. They propose to extend the mental model theory by a notion of conditional probability (see also Johnson-Laird (1994); Oberauer and Wilhelm (2003)). George (1997) confirms the importance of uncertainty in human reasoning, but he shows that none of the existing approaches captures all the data in an adequate way. Oaksford and Chater (2001) reject logic as a base of human reasoning and suggest a model solely based on probabilities. They also consider the probability of exceptions, and explain the suppression of valid inferences (and fallacies) by high values of the exceptions parameter (p. 352). While this model mainly focuses on the reasoning process, Politzer and Bourmaud (2002) try to combine the uncertainty approach with a more detailed pragmatic analysis. They claim that suppression arises because the epistemic implicature that the second condition might not be fulfilled reduces the believability of the first conditional. In their approach, the conditional probability of the consequent given the antecedent indicates the assertability of the conditional. Stenning and van Lambalgen (2003b) come back to the idea of logic as a base of human reasoning by stating that different logical models are appropriate for modelling different tasks. They account for Byrne s suppression data by representing conditionals in a default logic that can be implemented by spreading activation networks. Like others before them, they emphasise the importance of the interpretation process, which

11 Chapter 1. Introduction 5 they regard as reasoning to an interpretation (p. 11). It results in logical forms which a uniform reasoning mechanism can be applied to. According to them, there is no suppression of inference rules or mechanisms, but only a modification of the representations they operate on. They show that this approach can account for Byrne s suppression data. 1.2 Objectives The overall goal of my dissertation is to shed more light on the meaning of conditionals. It was motivated by the work of Stenning and van Lambalgen (2003b). As their approach emphasises the interpretative component, it is essential to know whether the interpretations suggested by Byrne (1989) are really those that have to be modelled. This is where my study tries to give answers. I want to show how people really understand conditional sentences, especially statements that have been used to elicit suppression effects. Although the importance of interpretation in reasoning has been noted by many scholars, their claims about subjects interpretations are mostly speculative and not based on systematic experiments. But even if they admit that interpretation and reasoning cannot be treated separately, they often give an insufficient account of the interpretation process and fail to consult linguistic theories on discourse understanding. The involvement of linguistics rarely goes beyond mentioning Gricean implicatures. Thus, another objective of this dissertation is to show how semantic and pragmatic theories can help to gain a better understanding of the interpretation and reasoning processes involved in the suppression task. Besides showing which interpretations are possible, I want to examine what makes people adopt one interpretation rather than another, and what determines how people reason from conditional statements. These questions are strongly connected with the way the suppression task is designed and with subjects assumptions about what they are supposed to do. The goal of this dissertation is not to give an exhaustive analysis of all relevant factors, but to show what has to be taken into consideration if one wants to understand conditional reasoning. This work contains new ways of analysing conditionals, so I will re-examine existing theories on conditional reasoning in the

12 Chapter 1. Introduction 6 light of my results. Special attention will be paid to the mental model theory and probabilistic approaches to human reasoning. 1.3 Terminology One of my goals is to gain more insight into the meaning or mental representation of conditional sentences. By mental representation I do not mean the actual representation of sentences or discourse in the brain. This term rather stands for the abstract structure representing a conditional. But in order to examine and describe different ways of interpreting conditional sentences, I need a certain notion of what a conditional or its meaning is, which makes it difficult to investigate the meaning without any biases. Thus, I try to introduce some underlying assumptions about the basic properties of conditionals without fixing too many parameters. In the literature, disagreement starts with the word conditional. There have been many different attempts to define this term, using words like possible, hypothetical, or inference. Wierzbicka (1997) claims that these concepts are more complex than the universal concept IF, which she suggests for what we label with the term conditional. 7 According to her, it is more appropriate to define this concept by some sample sentences illustrating its meaning and preventing confusion with other concepts like WHEN. She suggests sentences containing events that can really happen and specifying a particular time like If it rains tomorrow, I will not go. I think she is right in claiming that the simplicity of conditionals is often not recognised and things are made more complicated than they actually are. However, the fact that conditionals exist in all studied languages does not mean that there is only one way of interpreting them and making inferences from them. The literature 8 usually makes the distinction between indicative conditionals like the example above, and subjunctive or counterfactual conditionals like 7 Wierzbicka suggests a set of universal semantic primitives which have equivalents in every language and cannot be explained by other words. 8 For comments on this distinction see e.g. Bennett (2003); Jackson (1987); Edgington (2001)

13 Chapter 1. Introduction 7 If it had rained, I would have stayed at home. Wierzbicka (1997) even suggests the universal concept IF...WOULD for counterfactuals. Only few authors like Dudman (1991) try to define conditionals in a way that captures all occurring phenomena. This dissertation only deals with indicative conditionals of the type introduced above 9, i.e. without considering sentences like If I didn t have to work today, I would stay at home which are sometimes treated as indicative and sometimes as counterfactual conditionals. Thus, I restrict my attention to the type of conditional examined in the majority of studies on the suppression of valid inferences. It is not clear whether my results about the interpretation or meaning of these conditionals also apply to other types. If we speak of a conditional correcting or denying another conditional 10, we implicitly state that the first sentence is considered to be false (or not true). So, if we want to describe in which cases the second sentence can be understood as correcting or denying the first, we need a notion of what it is for a conditional to be true or false. This is a delicate and controversial issue, and it is not my intention to recapitulate the whole discussion about whether conditionals are truth-functional or not. I rather want to refer to the notion of necessary and sufficient conditions, which is a widespread concept in the literature on conditionals. In these terms, a conditional sentence is true if the antecedent is a sufficient condition for the consequent to hold. 11 This notion of sufficiency is generally relative to a certain context or to the assumption that certain other conditions hold together with the antecedent. The relationship between antecedent and consequent (i.e. the sufficiency of the antecedent) can also be defined as enabling certain inferences, especially what is usually called modus ponens. Maybe it is not only an inference rule operating on the representation of conditional sentences, but itself a substantial part of the meaning of conditionals. This is what Stenning and van Lambalgen (2003b) mean by stating that: the conditional is often not so much a truth functional connective, as a license for certain inferences (p. 14) 9 This type of conditional is what I henceforth refer to as conditional. 10 See chapter Here we see again how difficult it is to define conditionals or the concept IF without being circular and using words like if or condition.

14 Chapter 1. Introduction 8 In their implementation of a neural network, they represent the relationship between antecedent and consequent as a link between nodes, thus emphasising the fundamental function of conditionals in representing discourse and reasoning on these representations. The notion of the antecedent being a necessary condition for the consequent to hold is not relevant in defining the meaning of conditional sentences as such. But we will see later that it is important in interpreting the relationship between two conditionals and in making inferences.

15 Chapter 2 Method 2.1 Motivation Previous studies on the suppression effect were based on experiments where subjects were presented with a sequence of sentences and asked what they could conclude. In most cases, participants could only choose conclusions from a given list (Byrne (1989); Byrne et al. (1999), Experiment 2-4; Stevenson and Over (1995); Stevenson and Over (2001)), with the result that they only made inferences fitting into the categories expected by the experimenters. But even when subjects could generate their own answers, the analysis of these remained superficial. In their first experiment, Byrne, Espino, and Santamaria (1999) gave participants the opportunity to draw their own conclusions. But again, they only categorised the results in accordance with the mental model theory and returned to the previous design in the remaining experiments. Dieussaert et al. (2000) present a more detailed analysis of freely produced conclusions. Dieussaert et al. (1999) make subjects even describe the reasoning process in a think aloud study. The problem with their studies, however, is that they focus on the reasoning process itself, especially on the deployment of different reasoning strategies, but do not use the material to examine underlying interpretations. Like many other studies, they accept the stance on interpretation expressed in the mental model theory and build their own analysis upon it. Another shortcoming of most studies on conditional reasoning is that they only 9

16 Chapter 2. Method 10 account for answers given by the majority of subjects, without explaining why there is such a great variety of answers. Thus, they implicitly state that all other subjects did not interpret and reason in a natural and representative way which would be worth accounting for. Of course, there may be subjects who do not understand the task or do not take it seriously, but it is not likely that this is such a significant percentage, especially considering the fact that most subjects make the expected inferences in the standard condition. Again, Dieussaert et al. (1999; 2000) constitute an exception, but as said above their analysis does not attach enough significance to interpretation. 2.2 Design This shows the need for a study which tries to account for different types of answers by giving people the opportunity to express how they interpret conditional sentences. The method I used within my project can be called interview rather than experiment. 1 I did some of the interviews 2 together with Levente Barczy, who was investigating how people s understanding of the suppression task and the ontologies introduced by the instructions determine which logic they use. 3 The material was very similar to that used in previous studies, we only changed it slightly to make the sentences sound more natural, and added our own statements to examine certain factors, like the role of the content or the formulation of the sentences, more systematically. Yet, the way the sentences were presented and people were asked to draw conclusions differed significantly from other experiments. After the presentation of a statement, subjects were engaged in a conversation about how they understood the sentences and what their conclusions were. In contrast to most previous studies, I used a qualitative rather than a quantitative or statistical research method. Some psychologists are sceptical about qualitative research because it is very subjective and generalisations are not easily possible. It is also objected that the results are not natural in the sense that they would not occur without 1 Stenning and van Lambalgen (2003a) call a very similar method used in a study on Wason s selection task socratic tutorial dialogue. This terminology is misleading because the goal of these conversations is not to make the participants learn. 2 Subjects 13 to 26 in the excerpts presented below. 3 I will hence use the pronoun we when talking about the interviews.

17 Chapter 2. Method 11 the intervention of an experimenter. In our case, this could mean that the questions in the interviews trigger interpretations or conclusions people would not come up with naturally. But I think it is unlikely that people have a totally new way of understanding natural language in the interviews, it is rather a question of which of the available interpretations they adopt under such artificial conditions. Another issue is whether the assertions people make can really be taken as evidence of mental processes. But as I investigated possible outcomes of the interpretation and reasoning process and not the mental process itself, this does not seem to be a big problem in my case. Moreover, qualitative research seemed to be more adequate because I was not interested in the percentage of subjects making a certain conclusion, I rather wanted to find out why people come to their conclusions. There were no variables that could be evaluated statistically because my goal was to determine the range of possible interpretations and factors influencing which interpretations and conclusions people choose. The interviews also gave the possibility to develop new hypotheses. Quantitative research is often a testing of hypotheses in terms of predefined categories, but my goal was to question the categorisation introduced by previous studies and show that other interpretations are also possible. Of course, it is difficult to say why exactly people choose a particular interpretation, but such interviews can at least show the existence of certain interpretations. As with qualitative research in general, it is difficult to present a comprehensive analysis of all the data because there is no numerical data to explain. Another drawback is the danger of picking single quotations as evidence for one s hypothesis whilst withholding contradicting data. But sometimes qualitative research may yield more honest data than quantitative methods that suggest a uniform pattern of behaviour which can be described with numbers. That this is not always the case is shown by previous studies on the suppression effect where analysing the percentages of inferences led to wrong conclusions about the nature of reasoning. To sum, the qualitative method used in this dissertation is a useful complement to quantitative studies. I try to combine a psychological with a linguistic analysis of the interpretation and reasoning process. So, it is important to compare the method used in this dissertation with the methods standardly employed in linguistic research. Some linguists rely only

18 Chapter 2. Method 12 on their own intuitions or those of their colleagues, but their intuitions may not be correct. Others base their theories on the analysis of text and speech corpora, which are likely to contain more natural and representative data. However, it is often difficult to find enough significant data in corpora to investigate a particular phenomenon. I would place the method I deployed somewhere between these two extremes. The interviews are not perfect situations of natural language usage. Subjects are aware that they are asked about their understanding of language and thus likely to reflect upon their interpretation of the material. But asking others about their interpretations in an indirect way goes beyond just consulting one s own intuitions. Thus, this method offers a way of gathering data on a specific phenomenon systematically, i.e. without depending on rare appearances in corpora, and is at the same time a step away from a mere armchair analysis of natural language. My hypothesis was that people do not always interpret the presented conditional statements in the way expected by the experimenters, and that this fact has to be considered when interpreting the reasoning data. So, the first objective of the interviews was to investigate which interpretations people adopt in the suppression task. An obvious starting point was to use the material of previous studies to make my results comparable and clarify their data. But the form of the statements - two conditionals with the same consequent followed by one of the antecedents (the categorical premise) - does not seem very natural and does not occur very often in real written and spoken language. Thus, the question is whether this is an appropriate way to examine how people usually interpret conditionals. Should we not use more natural material, i.e. conditionals embedded in a broader discourse with the categorical premise presented in a less marked way? Firstly, it is important to note that experiments always constitute somewhat artificial situations, so even more natural material could not make them really natural. The deciding criterion of a good experiment is not naturalness, but that it allows conclusions about the real world. In this study, using rather abstract material is adequate because its interpretation is presumably based on the same principles or processes as the interpretation of more realistic material. People try to make sense of these sentences by using the mechanisms they use for interpreting

19 Chapter 2. Method 13 language in daily life. So, using more abstract, unnatural material might even be better because the processes which are involved are easier to investigate. Factors which might influence the interpretation process can be controlled more easily. For instance, the degree to which people use their own world knowledge can be examined systematically by varying the amount of context given to them. Conditionals presented without any context might help to find out if there is a default interpretation for conditionals. Adding further context may show which factors make people choose another interpretation instead of or in addition to the default interpretation, and what makes them doubt the truth of a conditional. Presenting two conditional sentences together may also reveal different ways of interpreting natural language conditionals. Of special interest is the case where the second conditional triggers a revision of the first. This provides insight into the interpretation process itself, and allows conclusions about the mental representations of conditionals. Another source of information are the inferences people make from the presented statements. They shed light on the role of natural language conditionals in human reasoning and the relationship between the antecedent and the consequent, e.g. whether the antecedent is a necessary or sufficient condition for the consequent to be true. But the invitation to make inferences is also one of the drawbacks of this kind of material. If people are presented with a conditional and a corresponding categorical premise and asked what follows, their interpretation of the conditional will be such that the truth of the antecedent implies the truth of the consequent. This does not permit us to find out whether there are other interpretations of the conditional not allowing such inferences. For this reason, I presented conditional sentences without a categorical premise at the beginning of the conversation, and only added the categorical premise after having interviewed subjects about their understanding of the conditional. Participants were told that the sentences were part of a conversation they overheard, or text messages on a mobile phone they had found. To find out what the default interpretation of a conditional is, we asked them what they thought the discussion was about, or why the messages had been sent. After the interviewees had been given the categorical premise, we asked them what followed from the sentences and why it followed. Sometimes we presented only one conditional and the corresponding

20 Chapter 2. Method 14 categorical premise at the beginning and asked subjects what they could conclude. Then we gave them the second conditional to see whether the new information would make them change their conclusion, and how they would interpret this sentence in the context provided by the first two. As already mentioned above, we usually presented one of the antecedents as a categorical premise, which means that we focused our attention on modus ponens inferences. Since I was mainly interested in the interpretation of conditionals, and not so much in the reasoning process, it seemed legitimate to use only this kind of inference, which I assumed to be most obvious. This permitted me to analyse the interpretation process without worrying about the type of inference. In order to investigate what role probabilities and uncertainty play in interpreting conditionals, we also asked people how sure they were about their conclusion, or what information they would need to be sure. Another issue this dissertation explores is the relationship between the two conditionals, especially between the two consequents. Subjects answers to questions like How many people are saying these sentences? or Why is the second person saying this in response to the first person?, indicated what they thought the function of the second conditional was. 4 Making people rephrase the sentences was another way of finding out how they interpreted the relation between the two sentences. We asked them whether the sentences sounded natural, whether these were sentences they would say, or, if not, how they would convey the same information using different 4 However, these questions did not always have the intended effect, as the following excerpt shows: Subject 18 A: Imagine you are sitting in a bus, and there are some people at the back of the bus having a discussion, and you want to find out what they are talking about. And you hear these sentences: If Rosa meets her friend Liz, she goes to the cinema. If she has money for a cinema-ticket, she goes to the cinema. She has money for a cinema-ticket. (...) A: What do you think, how many different people are saying these sentences? S: When they are talking about that? A: Yeah. S: Maybe just two people. A: Why two? S: I don t know, it s not a big conversation, really, is it, or a discussion point. A: Who would say which sentence? S: I mean if it was like secondary school children, there might be more, there might be five or six, but with adults it would just be a couple, I think.

21 Chapter 2. Method 15 words. These questions also elicited statements about subjects understanding of the task and the naturalness of our material. As the controversy about the role of exceptions and counterexamples in interpreting conditionals is an essential part of the literature on conditional reasoning 5, I wanted to test the plausibility of different hypotheses by asking questions like What could prevent her from being in the library? or Can you think of situations where she has an essay to write, but isn t in the library?. This was also one way of gaining insight into the involvement of world knowledge in interpreting the statements. Generally, it is important to note that not all of our questions were intended to provide information about one particular aspect of the interpretation and reasoning process. They often mainly served the purpose of making people talk about the sentences, which we hoped would reveal how they interpreted them. 2.3 Procedure The design of the interviews as described in the previous section gives only a rough overview of the presentation and questioning methods we used. This is due to the fact that interviewing people requires flexibility and adapting the questions to the interviewee. Moreover, our methods and materials continuously changed and improved as we did the interviews. I started by interviewing friends to get a feeling for how people reacted to my questions and how I could evoke information about their understanding of the sentences. Most of the subsequent interviews were done at a cafeteria at university, which means that most of our subjects were students. Yet, the ages of our 30 interviewees ranged from 13 to 65. They had different occupations and subjects of studies, which gave us the opportunity to examine the influence of someone s education and world knowledge on the interpretation and reasoning process. Except for one, all subjects were native speakers of English and most of them did not have any prior knowledge of linguistics or formal logic. We told the subjects that we were interested in how they reasoned and how they interpreted sentences. As people tended to be confused and uncertain about what they were expected to say, it was 5 See chapter 6.

22 Chapter 2. Method 16 important to emphasise that there were no right or wrong answers and that the interview was nothing like an intelligence or personality test. Then each person was presented with three to five different statements in an interview of about 10 to 20 minutes. 2.4 Evaluation It could be objected that people are biased by the way they interpret the first set of sentences and the questions they are asked about it, and thus do not process the following statements in a natural way. 6 The presentation of several statements allows us, however, to investigate this influence of the conversational context, and to examine if there are individual preferences for certain kinds of interpretation. Of course, we cannot even guarantee for the first statement that people interpret it in a natural and unbiased way. It is not clear how strongly people are influenced by the experimental character of the interview, and how conscious they are of the process which makes them choose their interpretations. In some cases it was obvious that the subjects did not interpret the sentences in a natural, unconscious way. It was notable that participants with a linguistic or logical background often gave several possible interpretations. Yet, the environment and character of the interviews were probably more similar to a real life situation than the experiments in previous studies. The interviewees could stay in familiar surroundings and were engaged in a conversation rather than asked to solve an abstract task. But, as mentioned above, they were often suspicious and thought we were interested in something different from what we told them. This effect was more noticeable with two interviewers. Subjects were more relaxed when we stopped asking them too much about the reasoning process and why they had changed their conclusions. 7 The results show that interviewing people about their understanding of conditional statements helps to evaluate the results of experiments on conditional reasoning, and to compare different approaches. This method gives the opportunity to ask further questions if someone s interpretation of the sentences is not clear from his conclusions. 6 Indeed such effects could be observed. See chapter 8 for details. 7 As the interviews are numbered in chronological order, the excerpts presented in this dissertation will allow the reader to follow the change of interviewing techniques.

23 Chapter 2. Method 17 A great advantage over the methods used in previous studies is that the types of answers are not so restricted because subjects are encouraged to express themselves freely. This helped us to get new ideas and not only to think in terms of previous hypotheses and categories. Indeed, the interviews suggested types of interpretation that have not been thought of before. 8 8 See section 3.2 for details.

24 Chapter 3 Different kinds of interpretations 3.1 Interpretations considered in previous studies Byrne (1989) introduces two different ways of interpreting two consecutive conditionals: the condition stated in the second conditional can be either an alternative possibility or an additional requirement for the consequent to be true. She puts each statement used in her experiments in one of these categories, and analyses the data about inferences made according to this categorisation, thus assuming that subjects always adopt one of these two interpretations. Subsequent studies on the suppression effect kept this distinction and hardly thought about other possible interpretations. In the few cases where other interpretations, e.g. the second conditional as a denial of the first (Stevenson and Over (2001)), were taken into consideration, the authors did not try to give a systematic account of possible interpretations. Yet, they are right in claiming that Byrne s two categories are not the only ones which have to be considered. I would even say that her treatment of the second conditional as an additional requirement is not precise enough. In order to account for all essential differences in people s conclusions, this category has to be subdivided again. Below, I will try to give a more comprehensive description of the different ways subjects can interpret the second conditional in the context provided by the first. The second major flaw in how many scholars dealt with the issue of interpretation is that they assumed subjects would adopt exactly the interpretation they deemed 18

25 Chapter 3. Different kinds of interpretations 19 most plausible themselves. Consequently, they analysed the reasoning data only with regard to this interpretation. The inappropriateness of this assumption becomes clear in Dieussaert et al. s think aloud study (Dieussaert et al. (1999)) which uses similar material as previous studies on the suppression effect. They take the following statements as evidence for their amendment strategy, but do not analyse the relationship between the two conditionals. 1 So she takes a domestic help, but this doesn t mean that she earns enough money, but one of the two conditions are fulfilled, so she takes a domestic help. 2 John goes out fishing, so there will be fish on the menu, whether he catches one or not. (Dieussaert et al., 1999, p. 229) Obviously, subjects have not adopted the intended interpretation of the second conditional as an additional requirement, and this might have influenced their conclusions to a greater extent than a particular reasoning strategy. The authors thought of the possibility that subjects would not adopt the intended interpretation, so they made other subjects score how alternative or additional the second antecedent is, and chose their material based on this scoring. But they fail to use this fact to interpret the results of their think aloud experiment: in the presentation of their results they keep the original categorisation. This means that they do not consider the actual interpretation of the sentences in their analysis. The hypothesis that subjects often do not have the reading that was intended by the authors of all these studies was confirmed in our interviews. In many cases where the second condition was meant to be an additional requirement (in Byrne s sense), subjects interpreted it as an alternative, as can be seen in the following examples. 3 1 Based on the statement: If she goes out working, then the woman next door takes a housekeeper. If she earns enough money, then the woman next door takes a housekeeper. 2 Based on the statement: If Jan goes out fishing, then there is fish on the menu tonight. If Jan catches a fish, then there is fish on the menu tonight. 3 The following abbreviations or markers are used in the transcriptions of the interviews: S = subject L = Levente A = Andrea

26 Chapter 3. Different kinds of interpretations 20 This is probably one of the reasons why we did not observe many suppression effects with modus ponens inferences and why others studies have not obtained uniform data. Subject 17 A: If Rosa meets her friend Liz, she goes to the cinema. If she has money for a cinema ticket, she goes to the cinema. And you know: She has money for a cinema ticket. S: Um, then she would go to the cinema, presumably, Um, again it s just the first sentence isn t really relevant because she s gonna have, if she has the money, and if she has the money, she goes, then she goes, whether she goes with Liz or not is unclear. L: So, basically, she doesn t need Liz around, as long as she has the money? S: Yeah, if, the first two statements aren t, are like two alternates and only one of them needs to be confirmed, and one of them is, so, she will go to the cinema. But we don t know whether she will be going with Liz or not. Subject 1 S: (reading) 4 If Maria studies hard then she gets high grades in the exams. If Maria is intelligent then she gets high grades in the exams. She studies hard. A: What follows? S: Hm, that she will get high grades in the exams. But I guess it doesn t say whether she is intelligent or not. So she might get high grades anyway, even if she doesn t study. Um... A: Do you think she will get high grades? S: Yes. A: Why can you say that? S: Because you said that if she does study hard then she will get good grades. And also if she is intelligent, she will get... There s kind of two options, you said she does study hard... A: Does she have to be intelligent to get high grades? S: No, she can also just study hard. Subject 20 A: If Tom studies hard, he ll pass. If he s intelligent he ll pass. And Tom # = speech not intelligible enclosed by * = not sure capital letters = pitch accent (I did not make a detailed prosodic analysis, pitch accents are only marked in cases where they seemed relevant for my purposes.) 4 In some interviews, I made subjects read the sentences because I did not want my intonation to influence their interpretation.

27 Chapter 3. Different kinds of interpretations 21 is studying hard. What follows? S: Think it follows that he ll pass, then. If he studies hard he ll pass, and he s studyin hard so he s gonna pass. It sounds like... it sounds like if he s intelligent he doesn t need to study hard to pass. But if he s intelligent and he s studyin hard then he ll pass. You know, the first one means he ll pass anyway, so he s gonna pass no matter what, really. 3.2 A new categorisation Interpretations of the conditional statements we are dealing with can be classified along three different dimensions. One of them is the relation between antecedent and consequent. In chapter one, I used the terms necessary and sufficient condition to describe this relationship. They will also be used here to describe the different ways of interpreting two successive conditional sentences. Another variable is people s stance on the coherence of the statement. Discourse that people are confronted with in real life (as opposed to experiments or interviews) is usually meant to be coherent, i.e. sentences refer to each other and are interpreted in the context provided by the previous ones. But in the suppression task, people find themselves in an unnatural situation and have to interpret sentences out of context. So, it is possible that they do not always regard the sentences as part of a coherent text or discourse, but interpret them as independent of each other. I will examine in the next chapter how likely it is that people actually adopt this stance; but the fact that we found examples of it in our interviews shows that it has to be considered in a description of possible interpretations. If people assume that a statement is coherent, they try to find an interpretation of it that supports this assumption, so the interpretations of both sentences are more strongly influenced by each other. I want to distinguish two types of coherent statements. In a consistent statement, the function of adding the second conditional is to accept the truth of the first conditional and add new information to the existing context. In an inconsistent statement, the second sentence states that the first sentence (i.e. the current context) is false and corrects it. The following analysis will be based on the common classification of rhetorical or discourse relations where this is possible, mainly referring to Asher and Lascarides

28 Chapter 3. Different kinds of interpretations 22 (2003). To explain the connection between two conditional sentences in terms of these relations, we also need the notion of focus (See e.g. Kadmon (2001); Rooth (1996)), which is used to describe the correlation between prosody and certain semantic or pragmatic phenomena. The term focus usually denotes the syntactic constituent of a sentence with the greatest prosodic prominence. The focused part of a sentence answers the salient question in the current context and is chosen out of a set of alternatives. The rest of the sentence, corresponding to this open question, is often called background Independent statements The first category comprises cases where the sentences are not treated as coherent. As one is not assumed to be said in response to the other, this kind of interpretation is always possible. It means that both antecedents are independent sufficient conditions for the consequent to be true. In this case, subjects come up with an interpretation for Byrne s library example which is totally different from the one she expected, namely that the student is in the library whenever it is open. Subject 20 L: Then the third message says, if the library is open, she ll be in the library. What do you think that means? S: Okay so it means that she s gonna be working full out during the day, so she s gonna be working like all day, so she prefers to work in the library, so she ll be in the library if it s open, but if it s not open obviously she ll be working elsewhere, maybe. Subject 23 L: Hm, hm. Then you read the next message, which says: If the library is open, she ll be in the library. S: Um, yeah, then I think it was getting increasingly likely that she was there. L: What do you think that person meant? S: Um, just exactly what she, exactly what they said, really. That she can be found most of the time in the library, um, she s hard-working.