AI Ethics Education and Evaluation Framework Based on Case Method Using Network Motif and Sentiment Analysis

The Case Method (CM) is coming to be celebrated as an effective means of teaching AI ethics since it can invoke critical consideration of multi-dimensional, complicated moral challenges. Quinn and Coghlan refer to its success with medical schools, with which it increases the perception of clinicians of the moral dimensions and facilitates the observant incorporation of AI instruments [6]. Kooli and others refer to the possibility of applying it both to programmes of academic study and to community programmes, and of endeavouring the ethics-focused teaching models particular to the locale [7]. Much work has been done to date, albeit much of it is such that it is limited almost exclusively to work-based cultural and social contexts and is based almost exclusively on work done to date. Most recent studies have demanded the setting up of an entire ethics code that covers the whole of the AI developmental process [8] [9]. Ashok et al. promote flexible and adaptable directives that shift in tandem with technological advances and refer to the importance of embedding innovations such as these into schooling systems more generally [10]. Many of the proposals lack an empirical foundation and are irrelevant to culture-specific applications. Focusing on the limitation, the structure subjects the case method to Japan's particular employment-searching tradition, the shukatsu, and grounds the moral discussion that takes place therein in the quotidian lives of the student population. Incorporating network analysis of the student's reflection, thus bringing the process of assessment an empirical intensity, it is able to provide a culturally sensitive paradigm of ethics teaching that covers an area noted in existing research.

The recurring subgraph patterns known as network motifs, appearing more frequently than statistically predicted, turn out to be useful analytical tools in the investigation of complex networks in such areas as, e.g., the topology of behavioral patterns in areas ranging from tourism, in which they identify how the tourists move between attractions, to trade, in which they identify the common commercial exchange that underlies the thinking of strategy [11]. Shao et al. have shown that these patterns can help track how educational changes affect thinking, revealing how changes in conversation relate to shifts in ethical views. Structural network analysis is supplemented by sentiment analysis, which extracts the emotional cues from the text data to bridge the gap. Advances in deep learning over the past few years have enormously expanded its scope of applicability in areas such as, e.g., education, marketing, and public discourse, esp. in particular. Liu et al. examined the emotional response of the student to the online learning setting [12]; in contrast, Yadav and Vishwakarma utilized the transformer structures to visualize the socio-political discourse's emotions [13]. Zhao et al. conducted the sentiment analysis and the network analysis both to identify how the emotional indicators shape the consumption pattern [14]. Here, the current structure integrates the sentimental analysis with the motif-based network analysis to identify the structure and emotional transformation of the reflection on ethics, both in terms of the former breakthrough.

The first phase of the framework contains four essential components. The process is initiated by the setting of pedagogical goals, which entails the identification of the moral dimensions of AI technology under review and the population of intended participants. Thereafter, relevant case studies are developed to fit the desired focus of instruction. These cases then become the basis upon which instructions are imparted under the case method, which invites the taking of positions on the moral dilemmas under consideration in an orderly, dialogical setting. The final step entails the systematic acquisition of the data both during and within the educational intervention period. Various methods of data acquisition may be used, ranging from the use of surveys and structured interviews to the use of sensor-based monitoring. Data must, at least, be collected twice during the whole process of instruction. Data so obtained may take different formats, ranging from quantitative scores, written accounts, tape-recorded discussions, or biometric measurements, and are all translated into either discrete or continuous formats that are amenable to further analysis. Importantly, the relationship between the participants and the data points is retained such that relational graphs that become the building blocks of the latter network modeling that follows may be generated.

The technique integrates network creation with motif analysis by mapping all relevant textual data into a network and evaluating the relevance of motifs as Figure 2 shown. The program requires textual input (T), a cohort of people (I), specified relations (R), a motif size (s), and several random graphs (M) for statistical analysis. A morphological analysis is conducted on the original textual dataset T, referred to as MA(T), which extracts morphemes from the text. The morphemes are subjected to an additional filtering process using a Filter (M, O) function with a specified goal O, resulting in a filtered subset S of the morphemes used to represent the textual data in a network [15]. The procedure starts by establishing empty sets for vertices (V) and edges (E). The algorithm subsequently examines whether a pre-existing connection exists for each individual-morpheme pair (i,m) inside the set I×S as defined by R. Upon verification of a relation, both the person and morpheme are included as vertices into set V, and an edge is established in set E, therefore forming a realized network G = (V, E). The next motif analysis starts by extracting a detailed enumeration of all interconnected subgraphs of size s inside graph G, which are then quantified and classified according to their topology and organized into a frequency set F. The program will thereafter generate M random graphs that possess the same degree-per-node based network topology G state, which are used for statistical comparison rather than representing actual or realized networks. Motif frequencies will be computed for the motifs of the random graphs in a same manner to establish comparison set C.

文本 AI 生成的内容可能不正确。

The method includes finding the average (r_t) and standard deviation (σ_t) of each motif shape from the random graphs we looked at, and then using a Z-score (Z_t) from these averages to see if the number of each motif in Graph G is different from what we expected based on those averages [16]. The algorithm creates and gives back a group of important subnetworks S, along with their frequencies (F) and Z-scores (Z), while also offering a full motif analysis that highlights important structural and statistical features in the network.

To complement the structural insights derived from network motif analysis, we integrated sentiment analysis to explore the emotional undertones present in participants’ written reflections. While motif analysis maps the structural logic of ethical reasoning, sentiment analysis sheds light on how learners emotionally engage with ethical dilemmas related to AI. Taken together, these methods offer a dual perspective—one that captures both the cognitive architecture and the affective dimensions of ethical reflection. This integrative approach resonates with emerging research emphasizing the critical role of emotional engagement in effective ethics education [17].

Our analytical workflow began with conventional text preprocessing procedures, including the removal of extraneous characters and morphological analysis, followed by the extraction of semantically meaningful terms. The processed textual data were then analyzed using a transformer-based sentiment model fine-tuned on Japanese language corpora. Each response was assigned a sentiment score on a scale from 0 (strongly negative) to 1 (strongly positive). By aggregating these scores—for instance, comparing pre- and post-intervention responses to Question 6—we were able to observe shifts in emotional tone over the course of the learning process, in line with established practices in AI education and the broader social sciences [18].

In the final step of the AEEE framework, the answers from participants to questions Q6 and Q7 (before and after the intervention) were turned into four text-based networks (Figure 3). We used morphological parsing to find 1,491 morphemes, subtracting 16 for punctuation marks; in total, we created 5,918 valid word tokens. Each response was shown as a graph connecting participants to individual morphemes. After the intervention, the Q6 network grew from 515 nodes and 1,418 edges to 1,023 nodes and 2,789 edges, showing more variety and complexity in the language used. Each response was represented as a graph with participants connected to individual morphemes. Following the intervention, the Q6 network increased from 515 nodes and 1,418 edges to 1,023 nodes and 2,789 edges, indicating a greater degree of lexical diversity and complexity of discourse. Furthermore, the level of engagement did not diminish as the average node degree and network density were consistent. Q7 also demonstrated a similar increase, demonstrating that the modifications of the intervention had the effect of increasing expressive complexity without disrupting the overall linguistic equilibrium.

图片包含 灯光, 交通, 游戏机 AI 生成的内容可能不正确。

Next, we analyzed the frequency of patterns for groups of three and four (Figure 4), finding six patterns (M1–M6). Motif 1 and Motif 3 illustrated individual lexical variety; Motif 2 and Motif 6 highlighted shared morphemes that referred to topical consensus or focal participants; Motif 4 illustrated convergence (partial overlap); and Motif 5 indicated agreement (strong consensus). After the intervention, both frequencies increased for Motif 2, M4, and M5. For example, Q6 increased from 0.056 to 0.072 for M2 and conjecture 40-60% for M4 overall. While occurrences of contextually accurate M5 were rare, the simple increase in occurrences indicated a higher degree of inclusion within the participants. Individually, M1 and M3 decreased slightly, with Q6's M1 declining from 0.944 to 0.928; this decrease may be due to a more precise application of lexical choices in greater occurrences related to task structure and student expression, rather than narrative fragmentation. The significance of the troop motif was previously identified and confirmed through z-scores generated for randomized null models (Figure 5). Significantly higher z-scores confirmed M4 and M5 increased structural relevance as numbers increased, while M3 and M6 decreased, revealing less fragmentation. The Q6 response displayed greater precision, and Q7 responses appeared more unified, further revealing the effectiveness of the case method approach as a pedagogical vehicle for task-specific ethical expression. The AEEE frame provided insight into the structural and pedagogical changes observed.

图表, 条形图 AI 生成的内容可能不正确。

图表 AI 生成的内容可能不正确。

Following the investigation of sentiment at the level of words, we determined the average sentiment score for each dataset to better understand participants' overall emotional framing. As can be seen in Table Y, the average sentiment for Q6 responses increased from 0.129 before the intervention to 0.245 following the intervention. This suggests movement towards more positively framed language about ethical dilemmas, in other words, decreased reliance on negations and negatively framed concerns.

Sentiment polarity trends (Figure 6) similarly suggested a general increase in scores for keywords post-intervention. In Q6, for instance, "Problem" increased from 0.067 to 0.181, and in Q7, "Decision" increased from 0.289 to 0.643. The Q7 response received the highest sentiment response of 0.938, signifying a substantial contribution to positively positioned emotional framing. On the other hand, "Utilization" rated comparatively lower in sentiment at 0.070 after Q6, suggesting possible uncertainty or continued caution. The quantities of sentiment visible in the bar chart indicated diminishing terms focused on deviations and a shift towards more strategic and evaluative language.

Although it is difficult to determine whether cumulative word choice shifts were due to part of the "after" responses with the use of emotive terms in the form of sentiment towards ethical decision-making, sentiment increased for recurring words with an apparent upward trend overall, particularly in Q7 (Figure 6), suggesting case-based discussion does influence emotional framing. After responses, the five highest-rated words were Approval = 0.938, Implementation = 0.672, Decision = 0.643, Student = 0.672, and Problem = 0.591. These indicate not only changes in word choice, however, but also an overall more comprehensive emotional shift in participants' responses to ethical dilemmas. Figure 6 depicts a distinct upward trend in sentiment for commonly used words, particularly in Q7 after, which further supports the influence of case-based discussion on emotional framing. Following the intervention, word ratings from most to least highly rated words were Approval (0.938), Implementation (0.672), Decision (0.643), Student (0.672), and Problem (0.591). These results demonstrate not only changes in word choice but also a more general change in thinking about ethics and emotionality.