Argumentation schemes: From genetics to international relations to environmental science policy to AI ethics

2.1.GenIE (Genetics Information Expression) assistant

The GenIE Assistant [32] was a prototype system for generation of the first draft of genetic counseling letters. It was implemented as a testbed for research on natural language generation of transparent biomedical argumentation. The system architecture included (1) a knowledge base (KB), a causal network representation of domain knowledge about genetic disorders and of information about a patient’s case; (2) a discourse grammar with genre-specific rules for creating an abstract representation (DPlan) of a letter to a client; (3) an argument generator which generated arguments in propositional form for claims in the DPlan; (4) an argument presenter which made changes in the DPlan for the sake of text coherence and transparency; and finally, (5) a linguistic realization component to render the DPlan as English text. The DPlan contained propositions about the patient’s case from the KB and was structured by text coherence relations of Rhetorical Structure Theory (RST) [43].

GenIE’s argument generator created arguments for claims in the DPlan using its argument schemes and information in the KB. The argument schemes and the format of the KB were informed by a corpus study of genetic counseling letters.33 The corpus study [14] revealed a conceptual model of genetics used in communication with clients. The conceptual model consisted of general concepts, such as history, symptom, genotype, and test result, and causal relations between concepts, which we described in terms of qualitative constraints of qualitative probabilistic44 networks [10]. The KB could be instantiated for different genetic disorders and for different patients’ cases.

Rather than refer to domain content, the argument schemes were formulated at a higher level of abstraction in terms of qualitative network node variables and relations. For example, to paraphrase one of GenIE’s Effect to Cause schemes shown in Fig. 1, the claim that a node A in the KB was at or above a certain threshold is supported by the Data (premise) that a node B in the KB is at or above a certain threshold, and by the Warrant (premise) that there is a positive influence relation from A to B. Such a scheme could be used to generate an argument that a certain patient has a genotype of two mutated alleles of a GJB2 gene based on the data that he has hearing loss, and the warrant that this genotype can result in hearing loss. Note that, unlike the schemes in the Walton et al. catalogue [60], the GenIE schemes distinguished premises as Data or Warrant [57]. The Data premise represented patient-case-specific information (stored in nodes of the KB), while the Warrant represented general biomedical principles (represented as qualitative constraints in the KB). The Data/Warrant distinction played a role in the organization of the text (DPlan) and in decisions of the argument presenter. In the DPlan, the Warrant was represented as the satellite of an RST Background relation, whose nucleus was an RST Evidence relation; the satellite of the Evidence relation was the Data, and the nucleus was the Claim [17]. The argument presenter applied heuristics to the DPlan to make a letter more concise. One heuristic was if two adjacent arguments in the DPlan included the same warrant, the warrant in the second argument was replaced by an adverb such as ‘similarly’.

Fig. 1.

Simple effect to cause scheme in GenIE assistant (example in italics).

Fig. 2.

Universal condition affective argument scheme.

GenIE’s schemes also included “applicability constraints”, which were used to determine the applicability of a scheme during the argument generation process. In Fig. 1, the Effect to Cause scheme included an applicability constraint that can be paraphrased as “there is no other node in the KB which has positive influence on B.” The applicability constraints were used as critical questions in an interactive version of GenIE.

In subsequent work we identified several affective argument schemes in the corpus for mitigating the client’s possible negative reaction to information presented in the letters [33]. For example, see Fig. 2. If the partially generated letter contained the information in Data, then the Warrant would be inserted into the letter. The phrasing of the sentence expressing the Warrant was derived from the corpus. The Conclusion of the scheme was not explicitly stated in the generated letter, consistent with use of this scheme in the corpus.

2.2.GAIL (Genetics Argument Inquiry Learning) system

The argument generator and schemes developed for GenIE were repurposed as the core of an educational argument modeling system for biology students, GAIL [22, 34]. GAIL’s user interface presented the student with information to use in constructing graphical arguments: the Problem (to give an argument for a certain claim), the Data (information from a fictitious patient’s medical record), several possible Hypotheses, and Connections (possibly relevant principles of genetics, i.e. warrants). See a screenshot in Fig. 3. For example, a Problem was “Give two arguments for the hypothesis that the patient, J. B., has cystic fibrosis (has two mutated copies of the CFTR gene).” The Data about the patient included items such as “History of respiratory problems. During her second year, J.B. developed a chronic cough and has frequent upper respiratory infections”, and “J.B.’s mother does not have a history of respiratory infections”. The Connections included items such as “When both copies of CFTR are mutated, the body produces abnormal CFTR protein”, “Abnormality of the CFTR protein may affect the pancreas”, “People with CFTR protein often have viscous secretions in the lungs”, etc. To construct an argument, the learner could select appropriate items from the Data and Connections lists, drag them into the argument diagramming area, and connect them into a Toulmin-style (Claim-Data-Warrant) diagram of an argument (or chain of arguments as shown in Fig. 3).

Fig. 3.

GAIL screenshot showing two arguments. (Main claim at top of each chain; warrants at right-angles to chain.)

An authoring tool enabled instructors to create a knowledge base (KB) in the same format as used in GenIE to represent domain knowledge and information about a patient’s case. The authoring tool also created a mapping from the English text to be presented to the student on the screen to the underlying KB. After a student constructed graphical arguments for a claim, GAIL mapped the student’s arguments to their internal representations. Then GAIL used GenIE’s argument generator to create expert arguments for the problem. By comparing the student’s arguments to GenIE’s arguments, GAIL was able to generate feedback on the structure and content of the student’s arguments. Previous educational argument modeling systems required manual authoring of expert arguments in order to enable generation of feedback on content. Also, although not implemented in GAIL, it was noted that use of the critical questions of the argument schemes could support generation of additional feedback, such as “Can you make an argument for a diagnosis other than cystic fibrosis that explains the patient’s malnutrition as well as his frequent respiratory infections?”.

2.3.Analysis of genetics testing company promotional brochure

In this project we analyzed written arguments and persuasive visual devices in a five-page patient brochure published by a genetic testing company [18]. The goal of the brochure was to motivate patients to seek genetic testing from that company. We identified variants of causal and practical reasoning arguments and a specialization of fear appeal arguments based upon Protection Motivation Theory [49].55 We noted that a patient might lack quantitative skills necessary for understanding probability statements used in arguments. Furthermore, we showed that while the brochure addressed critical questions that would support its arguments, analysis of two other sources revealed answers to critical questions that challenged the brochure’s arguments. In addition, the brochure’s arguments could be challenged by reframing, qualifying, or disputing elements of its arguments with information from the other sources.

2.4.Towards argument mining genetics research articles

After considering arguments addressed to the lay reader in the GenIE project, we investigated how argumentation is used in genetics research articles. Such articles are written by and for other scientists. Whereas the genetic counseling letters conveyed a simplified, accepted model of genetics to warrant the claims of the letters, the goal of scientific research is to discover new knowledge by rejecting or refining current models or proposing a new model. Thus, the argument schemes used in GenIE were not sufficient to model scientific research.

As a step towards creating an argument-annotated corpus of genetics research articles [19], we analyzed some of the arguments in a representative article [55]. The analysis distinguished premises as Data and Warrant. It found that some missing (implicit) Warrants were necessary for argument acceptability.66 Also, conclusions of arguments were sometimes implicit; and some conclusions (explicit or implicit) functioned as implicit premises of subsequent arguments. In follow-up work [21], we defined ten causal argument schemes based on analysis of that article and three other genetics articles [6, 9, 44], e.g., Effect to Cause, Failed to Observe Effect of Hypothesized Cause, Consistent with Predicted Effect, Hypothesize Candidates, Eliminate Candidates, and Joint Method of Agreement and Difference. The argument scheme definitions were stated in general terms rather than in terms of genetics concepts.77 Note that Effect to Cause, Failed to Observe Effect of Hypothesized Cause, and Eliminate Candidates are similar to the abductive argumentation scheme, argument from falsification, and argument from alternatives, respectively, in the Walton et al. catalogue [60]. Joint Method of Agreement and Difference was based on Mill’s Method of Agreement and Difference [37].

A pilot study [21] on using our proposed scheme definitions to analyze arguments in genetics research articles found that undergraduate biology students had considerable difficulty applying the schemes correctly.88 However, a smaller follow-up study with faculty as annotators found that they could apply the schemes successfully. Later, other researchers [47] independently found the schemes applicable to the analysis of five biochemistry research articles and speculated that the schemes may be applicable to the experimental biomedical literature in general.

Next, in order to create a freely available argument-annotated corpus, we selected a genetics research article [58] from the CRAFT corpus (CRAFT 17590087) to be annotated.99 In [20] we proposed an annotation system that would specify where in the text an argument’s premises, conclusion, and, possibly, answers to critical questions are expressed. The top part of Fig. 4 shows the proposed annotation of spans of text in an excerpt from the article. The bottom part shows how that excerpt was analyzed as a kind of causal argument from analogy. The first premise was shown as coming from spans 1, 2, and 5 (as indicated at the top of the figure), and the second premise was analyzed as implicit. Spans 3 and 4 were analyzed as responses to two critical questions of the scheme.

Fig. 4.

Example of proposed annotation of article in [20].

Fig. 5.

Genetics-specific method of agreement argument scheme.

Fig. 6.

Taxonomy of argument schemes in genetics research articles from annotation guide.

Fig. 7.

Prolog rule for extracting argument based on method of agreement.

Note that the definitions of argument schemes in [21] did not refer to genetics concepts. However, given the results of the pilot study with undergraduates, we wanted to make the task easier for future annotators. Thus we redefined the argument schemes in terms of a small set of domain concepts such as genotype and phenotype, e.g. as shown in Fig. 5. The revised annotation guidelines1010 include a taxonomy of the argument schemes as an aid to annotators (Fig. 6). The causal schemes are divided based on whether they involve one or two groups of individuals. The One Group branch includes argument schemes related to Effect to Cause and Mill’s Method of Agreement. The Two Group branch includes schemes related to Mill’s Method of Difference and a variant of argument from analogy with a causal conclusion. Both branches of the taxonomy include arguments based on consistency.

After revising the scheme definitions in terms of genetics concepts and relations,1111 we envisioned how the argument schemes could be used for argument mining if implemented as logic programming rules in Prolog [23]. We defined seven argument schemes as Prolog rules based on analysis of the first eight paragraphs of the ten paragraph Results/Discussion section of the article selected for annotation [58]. For example, an argument scheme based on Method of Agreement could be implemented for this domain as shown in Fig. 7. The rule says that an argument (whose scheme name is ‘Agreement’) may consist of the premise that a group of individuals G has abnormal phenotype P, the premise that G has abnormal genotype M, and the conclusion that M may be the cause of P, provided that G is a group such that G’s phenotype is P and G’s genotype is M. Note that by implementing the scheme in this way, it can be used to extract1212 an argument whether the conclusion is explicit or implicit in the text.1313 Also, by providing the name of the argument scheme, it is possible to associate critical questions with an argument.

As described in [23], to make use of the rules to extract arguments, first, BioNLP tools [51] would be used to extract relations from a text for the predicates in the argument schemes, such as have_phenotype. After the relations had been extracted, the argument schemes could be applied to the extracted relations to recognize the arguments in the text. To demonstrate this approach, we manually annotated the entities, relations and arguments in the article as described in [24].1414 An excerpt of the annotated article is shown in Fig. 8. Entities and propositions are listed immediately following the discourse segment in which they are expressed. The arguments are listed immediately following the entities and propositions that are used in the arguments. (Unlike the proposed annotation system in [20], we did not annotate which span(s) expressed a premise or conclusion.) In Fig. 8, three entities have been identified in the preceding discourse segment (not shown in the figure): group1, pheno1, and geno1. Identifiers were assigned by the annotator; paraphrases are optional documentation provided by the annotator. As shown at the top of the figure, two propositions have also been identified in the preceding segment. Following that, the annotation of an argument provides the premises and conclusion of an argument that has been identified in the segment. The argument schemes implemented as rules were tested to confirm that they could extract the annotated arguments.

Fig. 8.

Annotation of some entities and relations, and an argument.

In [23] we suggested that in the future argument schemes might be acquired by inductive logic programming [48] from articles whose entities, relations, and arguments had been similarly annotated. We are currently experimenting with use of inductive logic programming to derive the argument schemes from the annotated article [27].

2.5.Argument Visualization and Evaluation (AVIZE) in international relations

In our next project, we approached a very different domain from the biology-related domains of our previous work. We found that international relations arguments involve the beliefs, goals, appraisals, actions and plans of social actors such as countries, governments, and politicians. The goal of this project was to develop a prototype argument modeling tool, AVIZE (Argument Visualization and Evaluation), to assist in the construction and (human) evaluation of arguments in this domain, based upon evidence of varying plausibility collected from sources of varying reliability. In addition to diagramming tools, AVIZE came with a set of argument schemes that we identified by analysis of several articles written by international relations experts.1515 For example, the Plan Deception scheme is shown in Fig. 9.

Fig. 9.

Plan deception scheme in AVIZE.

We identified the following other schemes used to argue about an Actor’s intentions: Plan Distraction, Inferred Plan, Coercion, Increasing Boldness, Behavior Pattern, Inferred Positive Appraisal, Inferred Negative Appraisal. Like the Plan Deception scheme, these schemes are domain-specific but are related to AI plan recognition heuristics. In addition, we identified the following schemes for arguing in favor of a planned action of the protagonist: Practical Reasoning, Resist Coercion, and Avoid Negative Consequences.1616 These latter schemes are related to more abstract schemes such as practical reasoning, argument from threat, and argument from negative consequences, respectively. (In subsequent work [25] in this domain, we proposed an Intentional Cause to Effect argument scheme and its critical questions for reasoning about future events resulting from an agent’s intentional acts.)

Fig. 10.

AVIZE screenshot of inferred plan argument with evidence for each premise.

Fig. 11.

AVIZE screenshot of arguments supporting and attacking a claim.

Challenges to arguments were represented in several ways in AVIZE. First, the user could attach both supporting and opposing pieces of data to the premise of an argument and decide which to credit. In this application, it was possible for sources to conflict and to have varying credibility. As shown in Fig. 10, for example, the premise that Martians have landed on Earth was supported by a news item that hikers had found a Martian space ship in the Nevada desert but challenged by two other items attached to that premise. Second, AVIZE enabled the user to attach counterarguments and provide answers to critical questions challenging a claim (Fig. 11). Also, Fig. 11 shows that users could color-code claims and premises to show their confidence in a premise or conclusion.

2.6.Analysis of argument and rhetoric in environmental science policy arguments

In our next project, we analyzed uses of rhetoric in a genre combining scientific and policy arguments in two journal articles in the domain of environmental science [38, 39]. The long-term goal was to build a rhetorically-annotated digital corpus for research on rhetorical persuasion in argumentation. The short-term goal was to inform the design of tools to help students analyze use of rhetoric in science policy arguments. The overarching arguments in the articles were instances of value-based practical reasoning (VBPR), including a variant (VBPR-Knowledge Precondition) which argues for the need to acquire information as a precondition for action. In addition to the main VBPR arguments, in one article we noticed use of enthymematic VBPR arguments in the introduction to signal the authors’ values and position on climate change. The articles also used two “rhetorical” argument strategies [45]: protocatalepsis (“the rebuttal or refutation of anticipated arguments” (p. 241) and prolepsis as presage (warning of environmental catastrophe as motivation for action).

The analysis of rhetorical figures in the two articles was based on descriptions of figures in [11, 35, 36]. We found instances of a variety of types in each class of rhetorical figure as defined in [35]: schemes1717 (phonetic, lexical or syntactic pattern), tropes (figures involving semantics such as metaphor), chroma (figures involving pragmatics such as rhetorical questions), and moves (discourse patterns such as protocatalepsis). While our study did not reveal an association between rhetorical devices and argumentation schemes, it suggested that certain devices might play a role in automatic detection of rebuttals, e.g., rhetorical questions, sarcasm, satire, attacking the holder of an opposing view, concession, and conciliato (identifying with certain concerns). Currently, we are studying the argumentative role of the rhetorical figure of antithesis in five environmental science articles [28].

2.7.AI ethics debate (AIED)

In an AI Ethics1818 course for computer science students, we focused on the design of explicit ethical agents [46]. An explicit ethical agent is an artificial agent that reasons about the ethical acceptability of its action using an explicit representation of ethical principles. In contrast, an implicit ethical agent is programmed so that its actions are consistent with human ethical judgments but it has no explicit representation of ethics. According to some AI researchers [1, 54], it is preferable to develop explicit ethical agents since an autonomous agent may encounter situations requiring ethical decision making that were not anticipated by the agent’s creators. Moreover, it is possible to examine the ethical justification for an explicit ethical agent’s actions. (It is assumed that moral responsibility lies with the humans involved in the agent’s creation or use: programmer, designer, purchaser, user, etc.). As a pedagogical tool, we developed several argument schemes for modeling the ethical acceptability of an explicit ethical agent’s action in military and healthcare applications, incorporating issues that ethicists have raised for these domains.

Arkin [4] has done extensive research on design and implementation of explicit ethical agents for autonomous machines capable of lethal force, e.g., autonomous tanks and robot soldiers. In this highly codified domain, it is possible to implement military principles such as the Laws of War and Rules of Engagement, which are based on Just War Theory [61], rather than more abstract ethical approaches such as utilitarianism or deontological theories. These principles are encoded in Arkin’s system as constraints in propositional logic. After Arkin’s agent has proposed an action, the action is evaluated in accordance with these constraints by an “ethical governor”. Note that the action of an autonomous agent in Arkin’s system is limited to the action of firing (or not firing) on a target, with the only variation being in terms of choice of armament. Arkin states that the problem of accurate target identification, which involves subsymbolic reasoning, is outside of the scope of his research.1919

We defined an argument scheme (Fig. 12) which is more abstract than the specific military rules encoded in Arkin’s agents, yet which summarizes many of the key principles of international law and norms on just warfare [50]. To illustrate an application of this argument scheme to a fictitious case study, consider use of lethal force by Homelandia’s AI-controlled drone against a missile base in the country of Malevolentia that has been firing missiles across the border into Homelandia. The missiles have caused damage to several apartment buildings in Homelandia and killed several occupants. The Just cause premise is satisfied since the purpose of the action is defense of Homelandia. Suppose that all reasonable alternatives, such as proposing a cease fire to negotiate a truce, etc., have been exhausted, so the Last resort premise is satisfied as well. According to international norms, drone-fired missiles are not considered inhumane weapons, and Homelandia has no covert reason for attacking the Malevolentian missile base, respectively satisfying the Humane weapon and Right intention premises. However, suppose that it is not possible for Homelandia’s counterattack against the Malevolentian missile base to succeed without inflicting a very high number of civilian casualties since the missile base is located on the grounds of a hospital. In this case, the Proportionality and Legitimate target premises are not satisfied. Thus, there is conflicting support and opposition from the premises. In the Ethical justification, one may provide additional explanation as to why the action is ethically acceptable to some degree (or not).

Fig. 12.

Just war argument scheme.

To give a related example, Arkin’s agent is forbidden from doing an action if that would violate certain constraints relating to Proportionality and Legitimate target. (Deployment of Arkin’s agent assumes that the others of the first six premises have been satisfied.) However, in Arkin’s system a human operator can override the agent’s prohibition after providing a justification for purposes of later oversight of the operator’s override. Similarly, the Ethical justification of the Just War scheme can be used to explain a justifiable exception to the requirement to satisfy a certain premise. (See also the related discussion of the Ethical justification premise of the Healthcare Argument scheme.)

The field of biomedical ethics has provided ethical principles for the design of healthcare applications. Anderson and Anderson [2, 3] have done extensive research on design of explicit ethical agents in this domain. Adapting Ross’ prima facie duty approach to ethics [53], they implemented several duties of biomedical ethics [7]: beneficence (e.g. promoting a patient’s welfare), nonmaleficence (e.g. intentionally avoiding causing harm), justice (e.g. healthcare equity), and respect for the patient’s autonomy (e.g. freedom from interference by others). In addition to the above principles, the healthcare literature also cites the need to respect the patient’s privacy. Under the virtue of fidelity, Beauchamp and Childers [7] discuss the need for the professional to give priority to the patient’s interests, e.g. that the agent has no covert goal such as to endorse a particular medical device or prescription drug. The following Healthcare Argument scheme encodes these principles as premises (Fig. 13).

Fig. 13.

AI ethics healthcare argument schemes.

However, as Ross noted, prima facie duties may conflict. To handle such situations, Anderson and Anderson created a process for training their system using inductive logic programming to derive rules that generalize the decisions of medical ethicists on training cases. The ethicists first assigned numerical ratings to represent how strongly each duty was satisfied or violated in a particular training case. An example rule that was induced was “A healthcare worker should challenge a patient’s decision [e.g. to reject the healthcare professional’s recommended treatment option] if it isn’t fully autonomous and there’s either any violation of nonmaleficence or a severe violation of beneficence” [2, p. 71]. Such a rule, or a student’s justification for favoring certain principles in S, could be used to provide the Ethical justification.

To illustrate, this scheme can be used to analyze an argument as to whether it is ethically acceptable for a robot, Halbot, to steal insulin from Carla (a human) to give to Hal (a human) to save Hal’s life. This is a variant of an account given in [8] in which Hal stole the insulin. In both accounts, Hal and Carla are diabetics who need insulin to live. Hal will die unless he gets some insulin right away. In our version, Halbot breaks into Carla’s house and takes her insulin without her knowledge or permission. Presumably, Carla does not need the insulin right away. Applying the above Healthcare Argument scheme, the Beneficence premise is that the action A will prevent Hal’s death. However, the Nonmaleficence premise opposes A since A might cause harm to Carla if she is unable to obtain a replacement dose of insulin in time. The Justice premise is that Hal deserves equal access to insulin. However, Halbot’s action of taking away Carla’s insulin without her knowledge or permission is a violation of Respect for (Carla’s) Autonomy. An Ethical justification premise in the style of [2] could say that the positive contribution to Beneficence and Equity due to A in S outweighs A’s negative contribution to Maleficence and Autonomy in S.

Fig. 14.

Critical questions of AI ethics schemes.

Figure 14 shows a number of critical questions for challenging the acceptability of an action A of an artificial ethical agent shared by the preceding schemes. (To save space, they are listed here rather than with each scheme.) The Data question is especially significant for explicit ethical agents that must rely in part on subsymbolic processing such as facial recognition.2020

The purpose of these argument schemes is to help the student to analyze whether an artificial agent’s action is or is not ethically acceptable to some degree. (As the state-of-the-art advances, it might be possible for an artificial explicit ethical agent to use such a scheme to explain its actions.) Unlike the argumentation schemes described in [60], the premises are not considered to be jointly necessary conditions. Also, the hedge ‘to some degree’ in the conclusion is intended to reflect the intuition that some actions are more ethically acceptable than others, e.g., that telling a lie with the goal of cheering someone up is more acceptable than lying to secure investors in a pyramid scheme. In some cases some of the premises may support the conclusion that the action is ethically acceptable while other premises may oppose that conclusion, weakening its degree of ethical acceptability.

These schemes are related to value-based practical reasoning (VBPR) schemes proposed by argumentation researchers to model an agent’s argument for why the agent should do some action in consideration of the agent’s goals, values, and available means of achieving those goals [60] and in the current context or circumstances [12]. In addition to its use to describe what a rational agent should do, VBPR has been used to model what an ethical agent should do [8]. However, we distinguish ethical principles from values, and ethically acceptable acts from rational acts. In examples of VBPR, the value is a general concept such as ‘freedom’ and ‘safety’ and ethical dilemmas are modeled by specifying value preferences. In our schemes, a group of ethical principles, elucidated by ethicists for particular domains, contribute to ethical acceptability, and an act may be ethically acceptable only to some degree. Furthermore, a rational agent need not behave ethically (when circumstances do not require it), nor an ethical agent rationally. A rational agent whose circumstances require it to behave ethically may plan an action that is both rational and ethically acceptable by providing a VBPR argument whose circumstances require the agent to behave ethically, supported by a subargument that the action is ethically acceptable – using ethical argument schemes such as we have proposed.

AIED (AI Ethics Debate) was designed to support creation and graphical realization of AI ethics arguments using argument schemes such as those described above as templates for constructing arguments. AIED provides the student with drop-down menus for selecting case studies and ethics materials, which when selected appear in windows on the screen. Argument scheme definitions are listed in a panel on the right-hand side of the screen. Course instructors may provide case studies and ethics materials of their choosing. If desired, they can author their own argument schemes too.2121

Fig. 15.

Screenshot of AIED with case study and ethics windows and argument scheme panel minimized. The argument scheme has been dragged into the argument diagram construction workspace, creating a box-and-arrow template for the user’s argument, as shown. Boxes have been marked red (con) or green (pro) by the user. The user left the conclusion uncolored since there are both pro and con issues regarding its ethical acceptability.

The center of the AIED screen is a drag-and-drop style argument diagram construction workspace. When the student selects an argument scheme from the right-hand side panel, a box-and-arrow style template is rendered in the center of the screen (Fig. 15). The student may cut-and-paste text from the case study and ethics windows, and enter their own words into the diagram. Critical questions can be selected from a menu and are rendered as text boxes attached to the argument. Premise and critical question boxes can be colored green or red to indicate support or opposition, respectively, to the conclusion. It was decided that premises and critical questions would be rendered in the same way in the argument diagram (i.e. as text boxes attached to the conclusion) since they both could be used to support or weaken the conclusion. The factors that we considered most important for the student to consider were listed as premises, while other factors were listed as critical questions.