About this example
This page is an example research proposal in computer science (natural language processing applied to online safety), written at the level of a master's dissertation proposal that could be extended into a doctoral project. It proposes a system that flags online videos whose spoken content may be harmful, so that human moderators can review them first. Each part of the proposal is followed by a "What to notice" note explaining what the passage does and why.
It is an illustrative teaching example, not a real student's work. Do not submit any part of it as your own work; your university's academic integrity rules apply.
The example: background, problem and aims
1. Background and problem
Video-sharing platforms host very large volumes of user-generated content, and a small proportion of it contains harassment, hate speech, threats or incitement. Platforms rely on a combination of user reports, automated filters and human moderators. Automated filters for text are well established, but much of the harmful material in videos is spoken rather than written, and spoken language differs from written comments in vocabulary, structure and noise. A practical question is therefore whether speech can be converted to text accurately enough, and classified reliably enough, to help moderators decide what to review first.
This proposal addresses only the spoken channel. Harmful visual content, such as violent or sexual imagery, cannot be detected from a transcript, and a complete system would need to combine audio, visual and textual signals (Baltrušaitis et al., 2019). Limiting the scope in this way makes the project feasible and allows its contribution to be measured precisely.
2. Aim and research questions
The aim is to develop and evaluate a transcript-based classifier that ranks videos by the likelihood that their speech contains harmful content, as a triage aid for human review. The research questions are:
- How accurately does automatic speech recognition transcribe the speech in a sample of user-generated videos, and how do transcription errors affect classification?
- How well do a lexical baseline and a fine-tuned transformer classifier identify harmful speech segments in transcripts, on channels not seen during training?
- Do error rates differ across dialects and speaker groups represented in the evaluation set?
The example: literature and rationale
3. Related work
Text classification for abusive and toxic language has moved from lexical features with linear models to pretrained transformer models, whose self-attention architecture captures context across a whole passage (Vaswani et al., 2017). Three findings from the wider literature shape this proposal. First, models trained on one source of text often perform worse on another, so a classifier trained on written comments cannot be assumed to work on transcribed speech. Second, language models can encode social biases, including associations between dialects or identity terms and offensiveness, which can lead to unequal error rates across groups (Gallegos et al., 2024). Third, machine-learning studies in security-related settings are prone to recurring pitfalls, including sampling bias, data snooping, inappropriate baselines and evaluation only in laboratory conditions (Arp et al., 2022). Detection of harmful content is also an adversarial, open-world problem in which the content being detected changes in response to detection, a long-standing difficulty for machine learning in security (Sommer & Paxson, 2010).
The example: method
4.1 Data
A sample of publicly available videos will be drawn from channels covering a range of topics, with the sampling frame and dates recorded. Audio will be extracted and segmented, and transcripts produced by an open-source speech-recognition system. A random subset of segments will be transcribed manually to measure word error rate (RQ1). Two trained annotators will label each transcript segment as harmful or not harmful under a written codebook with definitions and examples for each category; agreement will be measured and disagreements resolved by discussion. Existing written-comment corpora will be used only as a comparison condition, to show how much performance changes when a model trained on written text is applied to speech.
4.2 Models and evaluation
Two classifiers will be compared: a baseline using word and character n-gram features with logistic regression, and a pretrained transformer model fine-tuned on the annotated segments. Segment scores will be aggregated to a video-level ranking. To avoid leakage, all segments from the same channel will be placed in the same split, so the model is never tested on speakers it has already seen (Kapoor & Narayanan, 2023). Because harmful segments are expected to be a small minority, performance will be reported with precision, recall and the area under the precision-recall curve rather than accuracy alone (Saito & Rehmsmeier, 2015). Hyperparameter search budgets and variation across random seeds will be reported (Dodge et al., 2019), and error rates will be compared across dialect and speaker groups where the sample allows (RQ3).
- 1Sample videos and record sampling frame
- 2Extract and segment audio
- 3Automatic speech recognition; measure word error rate on a manual subset
- 4Annotate transcript segments with a codebook
- 5Train baseline and transformer classifiers on channel-level splits
- 6Rank videos for human review and analyse errors by group
| Question | Measure | Why this measure |
|---|---|---|
| RQ1 transcription quality | Word error rate on a manually transcribed subset | Shows how much noise the classifier receives |
| RQ2 classification | Precision, recall, precision-recall AUC on unseen channels | Harmful content is rare; accuracy would be misleading |
| RQ2 robustness | Change in performance from written-text training to transcript training | Tests the domain-shift problem directly |
| RQ3 fairness | False positive and false negative rates by group | Unequal errors would harm some speakers more than others |
The example: ethics, timeline and contribution
5. Ethical considerations
Although the videos are public, the people in them have not consented to research, and transcripts may contain personal information. The project will follow established internet research ethics guidance, store only the data needed, remove names and other identifiers from transcripts, and not quote material that could identify a speaker (franzke et al., 2020). Annotators will be exposed to distressing material, so the protocol will limit session length, allow withdrawal from categories of content, and provide access to support. Approval will be sought from the university research ethics committee before data collection, and the platform's terms of service will be checked.
6. Timeline
| Months | Work |
|---|---|
| 1 to 2 | Ethics approval; codebook drafting and pilot annotation |
| 3 to 4 | Video sampling, audio extraction, speech recognition and word error rate measurement |
| 5 to 6 | Full annotation and agreement analysis |
| 7 to 8 | Model training and evaluation; subgroup error analysis |
| 9 to 10 | Writing up, with a final month for revision |
7. Expected contribution
The study will provide a measured account of how well transcript-based classification supports the triage of harmful spoken content, how much speech-recognition errors and domain shift reduce performance, and whether errors fall unequally on particular groups. A doctoral extension would add the visual channel and test multimodal models, evaluate the effect of the ranking on moderators' decisions, and study how performance changes over time as content adapts to detection.
How to adapt this to your own study
- State the decision or problem your system supports, and narrow the scope to what your method can detect.
- Write research questions that can be answered with evidence, not a list of processing steps.
- Use training data that matches the data your system will meet, or test the mismatch explicitly.
- Design splits that prevent leakage between training and test data.
- Choose evaluation measures that suit the class balance and the purpose.
- Plan for bias, ethics and annotator welfare from the start.
- If you build on a published design, cite it and say clearly what your study adds.
Common mistakes in this kind of proposal
- Reproducing a published design without acknowledging it.
- Training on one kind of text (written comments) and claiming results for another (speech).
- Claiming to detect all harmful videos with a method that sees only one channel.
- Comparing accuracy figures from studies of different tasks.
- Reporting accuracy alone when the positive class is rare.
- Treating ethics as a formality when the data involve real people and distressing material.
| What examiners look for | Where this example does it | Check in your own draft |
|---|---|---|
| A clear, bounded problem | Section 1 | Does your scope match what your method can do? |
| Answerable research questions | Section 2 | Could each question be answered with your data? |
| A review that informs the design | Section 3 | Is every point in your review used later? |
| A sound evaluation design | Section 4.2 and Table 1 | Have you prevented leakage and chosen fitting metrics? |
| Specific ethical safeguards | Section 5 | Are your safeguards specific to your data and people? |
| Feasibility | Table 2 | Is the plan realistic for your time and resources? |
| Originality and attribution | Sections 3 and 7 | Have you credited prior designs and stated what is new? |
Key takeaways
- Define the decision your system supports and scope the project to what the method can observe.
- Test the gap between training data and real data rather than assuming it away.
- Split data by source to prevent leakage, and report precision and recall for rare classes (Kapoor & Narayanan, 2023; Saito & Rehmsmeier, 2015).
- Check for unequal error rates across groups (Gallegos et al., 2024).
- Credit any published design you build on and say what your study adds.
