GAIL: A Genetics Argumentation Inquiry Learning System Nancy L. Green, Mark Hinshaw, Carl Martensen, Meghana Narasimhan and Tshering Tobgay Department of Computer Science University of North Carolina Greensboro Greensboro, NC 27402 USA nlgreen@uncg.edu Abstract Each GAIL lesson requires learners to construct This paper discusses ongoing work to build an arguments for and against certain hypotheses about a argumentation inquiry learning system, GAIL. The purpose genetics case, e.g., about an infant who may have an of GAIL is to support students in constructing scientific inherited metabolic disorder or someone who inherited a arguments in an undergraduate genetics course in order to genetic variant that is associated with increased risk of facilitate deeper learning and improve argumentation skill. colon cancer. A prototype user interface is shown in Students can construct argument diagrams using a drag- Figure 1 (see last page). Information relevant to the lesson and-drop graphical user interface. The system constructs arguments on-the-fly to use as a knowledge source for is provided by GAIL on the left-hand side of the screen: evaluating the learners’ arguments and providing the Problem (to give a certain argument); Hypotheses intelligent feedback. (which can be used in the argument, but note that some are incorrect); Data from medical records about the patient and the patient’s biological family; and Introduction Connections, a list of facts or principles of genetics. The center of the screen shows two arguments constructed by a Argumentation plays an important role in science. There learner. To construct the arguments, the learner searched has been significant interest within the field of science for text components on the left-hand side of the screen, education in argumentation. However, students’ dragged them into the workspace in the center of the arguments have been shown to be deficient in a number of screen, and connected the components. Arrows point from ways, e.g., lacking support for claims (Bell and Linn support to conclusion. The connection between support 2000; Jiménez-Aleixandre, Rodríguez and Duschl 2000), and conclusion – known as the warrant in argumentation failing to provide alternative explanations (Lawson 2003; theory (Toulmin 1998) – is linked by a line to the arrow. Schwarz et al. 2003), and using inaccurate or irrelevant In Figure 1, the problem is to give two arguments for support (Zohar and Nemet 2002). Computer-supported the hypothesis that the patient (referred to as J.B.) has cooperative learning systems for argumentation have been cystic fibrosis, i.e., has two variant alleles of the CFTR developed (Kirschner et al. 2003; Scheuer et al. 2010; gene. The leftmost “chain” of arguments begins with data Pinkwart and McLaren 2012) but they do not provide (at the bottom of the argument diagram) about J.B.’s human-level expertise in evaluating student respiratory problems. The learner used that data to support argumentation. Furthermore, in larger-enrollment classes an intermediate hypothesis that J.B. has thickened mucus human teachers may not have sufficient time to evaluate in the lungs, which is used to support an intermediate learners’ arguments nor to provide one-on-one feedback hypothesis that J.B. has abnormal CFTR protein, which is as the learner works on an argumentation lesson. used to support the main hypothesis/conclusion that J.B. To address this problem we are implementing a has cystic fibrosis. Branching from the right hand side of prototype genetics argumentation inquiry learning system, the diagram, connections (warrants) provide justification GAIL. GAIL will support learning to argue about cases in for each step of the argument. The second argument for human genetics. This is a field that applies findings from the same hypothesis begins with data about J.B.’s lab test genetics research to biomedical reasoning. GAIL is result. designed for use in an introductory genetics course for GAIL’s innovation is that the system can generate undergraduates that many biology majors find the most arguments for evaluating the correctness of the learner’s challenging course in the biology core curriculum. We arguments, rather than requiring the arguments to be hope that use of GAIL will improve argumentation skill, constructed by a teacher. Use of the generated arguments facilitate deeper learning of genetics, and increase interest enables GAIL to provide intelligent feedback on both the and engagement in science. structure and content of the learner’s argument. condition predisposing one to colon cancer) based upon System Design the data that genetic testing showed a variant MLH1 allele, The author of an argumentation lesson to be used in and the connection (warrant) that having HNPCC GAIL creates an XML-formatted file that contains: (1) typically leads to that test result. (The exception condition strings of natural language text -- the problem, hypotheses, for this argumentation scheme asks whether there is an data, and connections -- to be displayed to learners on the alternative explanation for the data.) An argument left-hand side of the graphical user interface as shown in diagram representing this argument is shown in Figure 3; Figure 1; (2) a specification of an internal causal domain to save space, instead of natural language text from the model; and (3) mapping of the natural language strings in graphical user interface, the diagram uses letters (1) to concepts and relations in the domain model. GAIL’s representing propositions, where A is the conclusion, B is the data, and S+(A,B) is the warrant. Authoring Tool provides (1) to the user interface, uses (2) to build an internal Knowledge Base, and stores (3) to enable GAIL’s Argument Evaluator to semantically A interpret learners’ argument diagrams, to avoid the challenge of interpreting natural language input. Based on our previous work on modeling genetics S+(A,B) (Green 2005), a Knowledge Base (KB) describes (i) instances of small set of concepts in human genetics (e.g. B genotype, protein, phenotype) and (ii) causal relations between these concepts. Causal relations are defined in Figure 3. Simple argument. terms of influence and synergy relations of a qualitative probabilistic network (QPN) (Druzdzel and Henrion Figure 4 shows a more complicated argument. The 1993). Different genetics KBs can be constructed main claim (A=1) is that a patient’s mother has exactly automatically from XML-language descriptions of the one mutated CFTR allele. The left-hand subargument is causal model specified using the Authoring Tool. for the hypothesis that she has one or two mutated CFTR Argumentation schemes are descriptions of acceptable, alleles. That subargument is supported by the hypothesis but often defeasible, patterns of reasoning (Walton, Reed (D=2) that the patient has cystic fibrosis (two mutated and Macagno 2008). Following the same approach as in CFTR alleles), and is warranted by the synergy relation, our previous research on argument generation (Green, X0(, D=2), i.e., that a child who has two Dwight, Navoraphan and Stadler 2011), GAIL’s mutated alleles inherited one from the mother and one Argument Generator creates arguments by instantiating from the father. Note that the claim D=2 would be abstract argumentation schemes with concepts and supported by another subargument (not shown in Figure relations from a QPN. The argumentation schemes are 4). The right-hand subargument is for the hypothesis that formalized in structures including claim/conclusion, data, the mother does not have two mutated CFTR alleles. This and warrant. The propositions used as claim or data is supported by the data (¬C) that she does not have cystic describe states of variables in a QPN. The warrant fibrosis symptoms, and warranted by the positive expresses formal constraints on the nodes of the QPN in influence relation between having two mutated CFTR terms of probabilistic influence and synergy relations. The alleles and symptoms of cystic fibrosis. distinction between premises as data and warrant reflects their difference in function and source of information. A=1 Data premises refer to a particular case, whereas warrants describe biomedical principles and other generalized knowledge. The condition of GAIL’s argumentation schemes is used to represent possible exceptions. & For example, the argumentation scheme for reasoning from effect to cause, shown in Figure 2, can be instantiated from a KB to create an argument that a patient A=1 or A=2 A≠2 has HNPCC (a mutation in the MLH1 gene, a hereditary X0(, Claim: A ≥ a S+(A=2,C) D=2) Data: B ≥ b D=2 Warrant: S+(, ) ¬C Condition: ¬ exists C X-({C,A},): C ≥ c Figure 2. Argumentation scheme. Figure 4. More complicated argument. Currently, GAIL employs seven argumentation argument. After each try, the Feedback Generator selects schemes. Arguments such as that shown in Figure 4 can be the most general (lowest level) unused message for an generated by chaining and/or conjoining subarguments. error; each time the student makes the same error on a Arguments are represented internally as directed acyclic subsequent try, the next more specific (next higher level) graphs. Unlike previous educational systems in which all message is selected. A positive message is generated when possible arguments to be used by the system had to be an error is corrected on the next try. Currently, the encoded by an author in natural language (e.g. Woolf et Feedback Generator displays only the message for the al. 2005) or in propositional logic (e.g. Yuan et al. 2008), most serious error to the student, but writes all of the GAIL’s arguments are generated by the system on-the-fly. detected errors to a logfile for inspection by the teacher. Since the argument generator and schemes do not encode domain-specific or patient-specific content, they can be used to generate arguments in any domain whose domain Feedback Example knowledge can be represented in a similar format. To illustrate the learner’s interaction with GAIL, After a learner has created an argument diagram, suppose the problem was to give an argument for the GAIL’s Argument Evaluator’s task is to evaluate the hypothesis that J.B.’s brother might have malnutrition and acceptability of the structure and content of the learners’ poor growth. Internally, GAIL generates a chained argument diagram. First, the learner’s diagram is argument beginning with the data that J.B.’s brother has translated into an argument structure containing KB been diagnosed as having cystic fibrosis, which supports concepts and links. The translation process uses the an intermediate hypothesis that his CFTR protein is correspondences between text the learner sees on the abnormal, which supports an intermediate hypothesis that screen and KB concepts and mappings provided via the he might have pancreatic abnormality, which supports the Authoring Tool from (3). The translated structure is in the main hypothesis that he might have malnutrition and poor same representation as arguments produced automatically growth. (The warrants of GAIL’s argument are not by the Argument Generator. Then the internal described here to save space.) However, on the first try representation of the learner’s argument is compared to all the learner’s argument contains the main claim that J.B. possible arguments created for the given problem by the (rather than J.B.’s brother) has cystic fibrosis, which does Argument Generator. not match the problem. Since this type of error has been GAIL’s Feedback Generator can respond to the given the highest severity code, GAIL would tell the following types of errors, where components of the student that the main claim of his argument does not learner’s argument are enclosed in brackets below: match the problem. On the second try, the student fixes •
does not match the claim to be the main claim and constructs a new argument. GAIL argued for in the problem. informs him that the problem noted on the last try has • is unsupported (i.e. no argument is been fixed. However, the student’s argument is missing provided for it). the intermediate hypothesis that J.B.’s brother might have • does not support the given pancreatic abnormality, so GAIL also informs the student . that one or more intermediate hypotheses are missing • does not support the given between J.B.’s brother having abnormal CFTR protein directly; one or more hypotheses and J.B.’s brother having malnutrition and poor growth. are missing between it and the given On the third try, the student adds the missing hypothesis . but provides an irrelevant warrant. GAIL would inform • Additional data or hypothesis must be conjoined the student that he has made progress but that the warrant to the given . he just added is irrelevant to that subargument. • is given as supporting but it should be conjoined to . Conclusion • The warrant is missing between the given This paper discusses ongoing work to build an and . argumentation inquiry learning system, GAIL. The • The given is irrelevant to the given purpose of GAIL is to support students in constructing and . scientific arguments in an undergraduate genetics course Note that, unique to GAIL, most of the above types of in order to facilitate deeper learning and improve errors are semantic in nature. argumentation skill. The system generates arguments on- For each type of error, the author of a GAIL lesson or a the-fly to use as a knowledge source for evaluating the system developer can provide a severity code and three learners’ arguments and providing formative and levels of feedback message templates in an XML- summative intelligent feedback. formatted file. In the current implementation of GAIL, a All of the components described in this paper have student is allowed three tries to construct an acceptable been implemented in Java. Future work includes improvements to the user interface and the Feedback Generator. The Feedback Generator will be made more Kirschner, P.A., Buckingham Shum, S.J., and Carr, C.S. intelligent to address certain types of errors that we have (Eds.) 2003. Visualizing Argumentation. London, UK: observed in our formative evaluation studies. For Springer. example, a learner “flattened” a chained argument into a one-level structure by conjoining together all of the data Lawson, A. 2003. The Nature and Development of and intermediate hypotheses. Note that in this case, the Hypothetico-Predictive Argumentation with Implications learner has selected the correct content but has just not for Science Teaching. International Journal of Science structured the argument properly into subarguments and Education 25(11): 1387-1408. shown how one subargument builds upon another subargument. Because the Feedback Generator has access Pinkwart, N. and McLaren, B.M. (Eds.) 2012. Educa- to arguments constructed by GAIL’s Argument Generator, tional Technologies for Teaching Argumentation Skills. the Feedback Generator will be able to detect this type of Sharjah: Bentham Science Publishers error and provide more meaningful feedback than systems that do not have access to content. After these Scheuer, O., Loll, F., Pinkwart, N., and McLaren, B.M. improvements are made, we plan to evaluate GAIL’s 2010. Computer-Supported Argumentation: A Review of effectiveness in an undergraduate genetics course. the State of the Art. Computer-Supported Collaborative Learning 5(1): 43-102. Acknowledgments Schwarz, B., Neuman, Y., Gil, J., and Ilya, M. 2003. Former graduate students Mark Hinshaw, Carl Martensen, Construction of Collective and Individual Knowledge in Meghana Narasimhan, and Tshering Tobgay contributed Argumentative Activity. Journal of the Learning Sciences to the implementation of GAIL for their MS Projects. 12(2): 219-256. Former graduate students Benjamin Wyatt and Chris Cain also contributed to the implementation of GAIL. Wyatt Toulmin, S.E. 1998. The Uses of Argument, Cambridge, and Martensen received support from a UNCG Regular UK: Cambridge University Press. Faculty grant and Cain received support from the Computer Science Department. Dr. Malcolm Schug of the Walton, D., Reed, C., and Macagno, F. 2008. UNCG Department of Biology has provided helpful Argumentation Schemes. Cambridge, UK: Cambridge feedback on the project. University Press. Woolf, B., Reid, J., Stillings, N., Bruno, M., Murray, D., References Rees, P., Peterfreund, A., and Rath, K. 2002. A General Bell, P., and Linn, M. 2000. Scientific Arguments as Platform for Inquiry Learning. In S.A. Cerri, G. Learning Artifacts: Designing for Learning from the Web Gouardères, and F. Paraguaçu (Eds.), ITS 2002, LNCS with KIE. International Journal of Science Education, 2363, 681-697. London: Springer. 22(8), 797-817. Yuan, T., Moore, D., and Grierson, A. 2008. A Human- Druzdzel, M.J. and Henrion, M. 1993. Efficient Computer Dialogue System for Educational Debate: A Reasoning in Qualitative Probabilistic Networks. In Proc. Computational Dialectics Approach. International of the 11th Nat. Conf. on AI (AAAI-93), pp. 548-53. Journal of AI in Education 18, 3-26. Green, N.L. 2005. A Bayesian Network-Based Coding Zohar, A. and Nemet, F. 2002. Fostering Students’ Scheme for Annotating Biomedical Information Presented Knowledge and Argumentation Skills through Dilemmas to Genetic Counseling Clients. Journal of Biomedical in Human Genetics. Journal of Research in Science Informatics. 38, 130-144. Teaching 39(1): 35-62. Green, N., Dwight, R., Navoraphan, K., and Stadler, B. 2011. Natural Language Generation of Biomedical Argumentation for Lay Audiences. Argument and Computation 2(1), 23-50. Jiminez-Aleixandre, M., Rodriguez, M., and Duschl, R.A. 2000. ‘Doing the Lesson’ or ‘Doing Science’: Argument in High School Genetics. Science Education, 84(6), 757- 792. Figure 1. Screen shot of prototype GAIL user interface