Explore how NLP and language AI can address real-world challenges that affect quality of life. Projects start with a human need, then define a researchable language problem, evaluation, dataset, or prototype.
Do not start from the tool. Start from the problem: where does language create misunderstanding, friction, exclusion, or difficulty accessing a service or information?
1Social problem
2Affected users
3Language problem
4NLP task
5Data
6Annotation
7Method
8Evaluation
9Error analysis
10Prototype
11Expected impact
Transport & LogisticsAdvanced
Pre-Arrival Delivery Check
Social problem
Ambiguous or contradictory delivery notes can delay a handoff.
Who is affected?
Recipients, dispatch staff, and drivers.
Language barrier
Location references, actions, restrictions, and contact preferences may be incomplete or unclear.
Ambiguous or contradictory delivery notes can delay a handoff.
Expected impact
Potentially clearer handoffs and fewer avoidable instruction failures.
Required data
Synthetic, researcher-created, consented, or licensed delivery notes; no platform data is assumed.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potentially clearer handoffs and fewer avoidable instruction failures.
Evaluation
Precision, recall, and F1 against a human reference; errors grouped by ambiguity type.
Visitors may misinterpret entry, navigation, ticketing, or event instructions.
Expected impact
Potentially more accessible visitor guidance without replacing official instructions.
Required data
Rights-cleared visitor instructions and researcher-created comprehension cases.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potentially more accessible visitor guidance without replacing official instructions.
Evaluation
Human-rated fidelity and clarity, unresolved-question analysis, and error review.
Surplus resources and matching needs may be described with different terminology.
Expected impact
Possible support for finding reusable resources more efficiently.
Required data
Consented or licensed descriptions of surplus items and needs with relevant attributes.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Possible support for finding reusable resources more efficiently.
Evaluation
Relevant/irrelevant judgments, ranking quality where justified, and false-match analysis.
Arabic is sometimes weakly represented in children’s digital leisure activities.
Expected impact
Potentially broader access to Arabic leisure content; no developmental claim.
Required data
Researcher-authored activities and consented evaluation responses; no scraped children’s data.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potentially broader access to Arabic leisure content; no developmental claim.
Evaluation
Human rubric for fidelity, clarity, appropriateness, and failure analysis.
A person may read a public instruction but still not know the next action.
Expected impact
Potentially clearer navigation of public information, not official advice.
Required data
Public, licensed, or researcher-created service instructions and a narrow extraction schema.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potentially clearer navigation of public information, not official advice.
Evaluation
Span/category precision and recall with a complete reference, plus omission analysis.
One instruction can be clear to one audience and difficult for another.
Expected impact
Possible support for more inclusive language review without judging individual ability.
Required data
Rights-cleared instructions and researcher-defined, non-diagnostic user profiles.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Possible support for more inclusive language review without judging individual ability.
Evaluation
Human clarity/fidelity rubrics and qualitative error analysis across profiles.
Written health information may contain difficult expressions or unclear sequences.
Expected impact
Potential contribution to health-information comprehension research only.
Required data
Approved educational health texts and ethically collected comprehension judgments.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potential contribution to health-information comprehension research only.
Evaluation
Expert fidelity review, comprehension-item review, and omission/error analysis.
Literal understanding may miss politeness, formality, indirect requests, or pragmatic meaning.
Expected impact
Possible support for pragmatic-language learning and research without stereotyping groups.
Required data
Researcher-created or consented contextual utterances with situation metadata.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Possible support for pragmatic-language learning and research without stereotyping groups.
Evaluation
Annotator agreement, evidence review, and qualitative analysis of disputed cases.
Course and administrative instructions can hide conditions, requirements, or unresolved references.
Expected impact
Potentially clearer study instructions while preserving the authoritative source.
Required data
Rights-cleared academic instructions with manually reviewed requirements and conditions.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Potentially clearer study instructions while preserving the authoritative source.
Evaluation
Span/category precision and recall, omission review, and ambiguity error taxonomy.
Residents may describe places with landmarks and relations rather than official address terms.
Expected impact
Possible support for organizing community place descriptions.
Required data
Synthetic, consented, or licensed place descriptions; mapping/GIS remains external.
Core execution steps
Define the social need → Prepare rights-cleared data → Define the language task → Create a human reference → Run the tool → Evaluate and analyze errors → Review prototype risks
Expected outputs
Possible support for organizing community place descriptions.
Evaluation
Span/category precision and recall with a reference, plus relation error analysis.