Legal
Data Ethics
TOMCUBE's ethical principles for collecting, processing and sharing multimodal language data, including consent, de-identification and review.
These guidelines apply to every stage of designing, collecting, processing, annotating, verifying, storing and sharing multimodal language data at TOMCUBE and among our partners. We treat ethics as part of data quality: a corpus whose origin, authorization and processing boundaries cannot be explained has no research value.
1. Purpose and principles
We follow four principles:
- Respect for autonomy. Explain data use, retention and sharing scope, and respect participants’ informed choices and right to withdraw at any time.
- Beneficence and non-maleficence. Assess physical, psychological, social and reputational risks, and apply protections proportionate to those risks.
- Justice. Avoid discrimination in recruitment, benefit allocation and data processing, with additional protection for vulnerable groups.
- Minimization and accountability. Collect only data needed for the research purpose and keep auditable records of processing.
2. Scope
These guidelines cover speech, video, text, eye-tracking, physiological and motion-capture data collected in laboratories, field sites, homes, schools, medical institutions and remote settings. They cover the full lifecycle: participant recruitment, informed consent, collection, processing and annotation, quality review, storage and access, external sharing, secondary use and destruction.
3. Informed consent
Informed consent is the starting point for lawful data. We require four conditions: sufficient information, genuine understanding, voluntary participation and the ability to withdraw. Disclosure covers the project purpose and methods; data types, duration and settings; use and sharing scope; retention and destruction; potential risks and mitigations; how participants can pause or withdraw; and who to contact for the research team and ethics review.
Consent is recorded in writing or in a traceable electronic form, with the consent version, signing date and the person who explained it. Sensitive personal information, biometric features and children’s data require separate consent. If the scope changes, we re-disclose and obtain new consent from participants.
4. Children and vulnerable groups
Children, people with cognitive impairments, detained persons and participants in dependent relationships require additional protection:
- Collection involving children requires written guardian consent and, where age-appropriate, the child’s own willingness to take part.
- Sessions can be paused at any time to avoid long continuous recording, fatigue and anxiety.
- We do not record family content unrelated to the language study, and we do not publish audio/video clips that could identify a child.
- In medical, school or care settings, institution-level ethics approval and data management arrangements are completed in parallel.
- For participants in dependent relationships, recruitment and consent explanation are not conducted solely by a direct supervisor or caregiver.
5. De-identification and anonymization
We process data in three steps: remove direct identifiers, control quasi-identifiers, and assess re-identifiability. Direct identifiers include names, contact details, identity numbers, precise addresses and employer names. Quasi-identifiers include combinations of age, sex, occupation, dialect site and education. For the re-identifiability assessment, a team member who did not collect the data listens to or reads the processed sample to judge whether personal identity could still be inferred.
Speech is harder to de-identify than text because of voiceprint characteristics. Raw audio is retained only in controlled, encrypted storage; analysis copies are normalized for pitch and speaking rate and trimmed of background details; high-risk segments are provided as transcripts and feature files only. Every de-identification step records the method, operator and date, and residual re-identification risk is reported in the delivery report.
6. Minimization and purpose limitation
Research questions and required variables are defined before collection. Processing and sharing stay within the approved and authorized scope and do not expand because data is “conveniently available”. Without explicit authorization, research data are not used for model training, product development, advertising or other commercial purposes. At project close, data are destroyed or anonymized as agreed, with a certificate of processing.
7. Ethics review process
All projects involving human participants pass ethics review before starting. Review combines written materials and panel discussion, focusing on risks and benefits, the consent process, data management and protection of vulnerable groups.
| Stage | Review focus | Output |
|---|---|---|
| Initiation | Purpose, participant scope, initial risk assessment | Ethics application form |
| Design | Collection list, consent form, data management plan | Final protocol and consent form |
| Pre-collection | Site, equipment, staff qualifications, contingency plan | Collection readiness checklist |
| Processing | De-identification plan, access rights, sharing scope | Processing record and access log |
| Pre-delivery | Re-identifiability assessment, authorization chain | Ethics compliance statement |
| Close-out | Data destination, destruction or archiving | Close-out and destruction record |
Projects that fail review do not proceed to collection. If unforeseen risks emerge during collection, work pauses and review is repeated.
8. Field conduct
Fieldwork follows a “explain, check, then record” sequence: confirm participants understand and have signed the consent; check equipment, venue and privacy conditions so unrelated people are not present; announce the start and end of recording; and allow participants to stop and request deletion at any time. Collection staff complete ethics training, sign confidentiality undertakings and do not store raw material on personal devices.
9. Annotation and processing
Annotators access only the minimum data needed for their task, sign confidentiality agreements and complete ethics training. If annotators encounter content involving personal privacy, illegality or potential harm, they stop and escalate. External annotation teams sign data processing agreements defining access scope, retention and liability, and are audited periodically.
10. Secondary use and licensing chain
When corpora are shared externally, the authorization chain must be complete: participant consent covers the sharing purpose; the commissioning party has the right to sublicense; and the sharing agreement defines purpose, scope, term and prohibitions on re-identification. We review applicants’ intended use, require commitments not to attempt re-identification or sublicense to third parties, and require citation of the data source in outputs.
11. Sensitive data and biometrics
Physiological signals, eye-tracking trajectories, facial video and voiceprints are sensitive personal information requiring separate consent and stricter access control. Biometric data are not used for identity authentication, are not matched against external databases, and are not stored in plaintext in the cloud.
12. Researcher and annotator welfare
Ethics protects those who execute research as well as participants. We provide psychological support channels for annotators, avoid prolonged exposure to high-emotion content, define workload and rest arrangements, and offer rotation and briefing for projects involving traumatic content.
13. Violations and reporting
Any team member who becomes aware of a violation can report it confidentially or openly through the ethics contact. We commit to starting a review within 10 business days, pausing affected projects where necessary, and giving the reporter feedback on the outcome. Members or partners who violate these guidelines may be required to complete training, remediate, or end the collaboration; suspected illegal conduct is referred to the relevant authorities.
14. Contact and updates
Questions, suggestions or reports about these guidelines can be sent to our ethics contact at admin@tomcube.com, phone +86 10 8585 8518, or Room 247, 2F, East Building, No. 015 Shui Niufang, Dougezhuang Township, Chaoyang District, Beijing. We review these guidelines annually and publish the current version on our website.
Guidelines version: v2.0. Last updated: November 1, 2026. Effective: November 15, 2026.