A human-centered, retrieval-augmented pipeline that turns moments in familiar stories into grounded questions about real-world child safety.
Child safety education is important, yet abstract rules can be difficult for young children to understand and remember. Parents may also lack the time, specialized knowledge, or confidence to provide safety instruction consistently. SafeQA investigates whether familiar story events can become situated prompts for discussing how a child should respond in a comparable real-life situation.
Seven parents of children aged 5-9 participated in semi-structured interviews. The findings shaped three requirements: integrate safety learning into recurring everyday activities; anchor abstract guidance in simple, imaginable situations; and use an agent to assist, rather than assume unlimited parent time or expertise.
Safety learning should recur naturally instead of depending only on occasional formal instruction.
Concrete story scenes give children an accessible mental model for otherwise abstract risks.
Automation can help parents surface relevant material while preserving adult involvement.
Interview seven parents about current practices, constraints, and design needs.
Create 141 structured safety entries and annotate 157 story-knowledge-QA pairs.
Rank curated knowledge against each story passage.
Condition GPT-3 on the story, retrieved knowledge, and selected examples.
Run offline module tests, then compare complete outputs with parents.
The formative study and final parent evaluation involved the same seven participants. The project therefore treats their feedback as an exploratory design signal, not population-level evidence.
The architecture separates factual grounding from language generation. A learned ranker connects a story passage to curated safety entries; a frozen GPT-3 model then receives that evidence through a structured few-shot prompt. This design makes the intended lesson explicit instead of relying only on knowledge implicit in the language model.
Sentence-BERT embeds story sections and safety entries. Separate MLP towers learn task-specific representations. A gradient-reversal discriminator discourages overfitting to the small set of individual stories, while pairwise BPR loss teaches relevant knowledge to rank above sampled negatives.
The prompt combines the task definition, child-safety objective, story-to-real-life guidance, retrieved knowledge, and up to two selected demonstrations. Examples first match the knowledge label and are then selected or supplemented by story similarity under the model’s token limit.
These thesis figures expose the two mechanisms behind the result: cross-story knowledge retrieval and knowledge-conditioned few-shot generation.
Leave-one-out cross-validation tested whether the ranker generalized to a held-out story. It outperformed TF-IDF and Sentence-BERT cosine similarity across Hit@K, NDCG, precision, and recall. Removing adversarial learning reduced every reported metric, providing an ablation check on that design choice.
| Model | Hit@1 | Hit@6 | NDCG | Precision | Recall |
|---|---|---|---|---|---|
| TF-IDF | 0.000 | 0.259 | 0.124 | 0.043 | 0.046 |
| Sentence-BERT + cosine | 0.500 | 0.667 | 0.565 | 0.178 | 0.221 |
| NCF, no adversarial loss | 0.467 | 0.800 | 0.642 | 0.339 | 0.412 |
| NCF + adversarial loss | 0.533 | 0.833 | 0.670 | 0.372 | 0.466 |
Across all 157 annotated pairs, leave-one-out evaluation compared the knowledge-grounded few-shot configuration with a zero-shot GPT-3 baseline. Few-shot prompting improved both lexical overlap and semantic similarity without updating GPT-3 parameters.
The same seven parents reviewed five story cases. Each case contained three QA pairs from the proposed pipeline and three from a zero-shot baseline. Questionnaires measured preference, technical quality, and child-safety value; follow-up interviews examined usefulness, knowledge matching, comparison with existing education, adoption intent, and future interaction needs.
Parents saw story-grounded QA as a useful supplement for sparking interest and reinforcing safety awareness in daily life. They did not consider a single generated exchange a substitute for systematic professional education.
They also wanted richer back-and-forth interaction, child-only and parent-child modes, and more visual or embodied presentation. These findings define product directions rather than capabilities already implemented in the thesis prototype.
The exploratory sample included seven parents recruited through network and snowball sampling. Future studies should include more families, children, educators, and child-safety professionals.
Expand and professionally validate the 141-entry corpus and the 157-pair dataset before relying on generated material in higher-stakes settings.
Develop guided multi-turn interaction, including a child-facing mode and a co-reading mode in which caregivers select or adapt suggested prompts.
Position the agent as a contextual supplement, expose retrieved evidence, and give caregivers control rather than presenting generated QA as professional safety instruction.
Research design, interviews, corpus and dataset construction, model and prototype implementation, experiments, analysis, and thesis writing.
Provided domain and research-method guidance, critical feedback, and academic supervision throughout the thesis.