Multilingual Schema Grounding

Skill: multilingual schema grounding

1. Capability Definition & Real Case

• Professional Definition: The ability to map natural language intent to relational schema and logic across languages with divergent syntactic orders (e.g., SVO to SOV) and complex morphological structures. It requires resolving agglutinative suffixes (case endings, plurality) that carry logical operators, navigating non-linear word sequences where the action verb is terminal, and mitigating tokenization fragmentation in morphologically rich, low-resource languages.
• Dimension Hierarchy: Robustness and Adaptation->Interface Adaptation->multilingual schema grounding

Real Case

[Case 1]

• Initial Environment: A localized database contains a table 'Kullanicilar' (Users) with a column 'yas' (age). The environment enforces strict Turkish ASCII-only identifiers to prevent encoding errors.
• Real Question: Yaşı 20’den büyük kullanıcıları getir.
• Real Trajectory: 1. Identify the verb 'getir' (get/bring) at the end of the sentence as the SELECT intent. 2. Ground the subject 'Yaşı' to the column yas. 3. Analyze the phrase '20’den büyük' where the ablative suffix '-den' combined with 'büyük' (greater) signifies the > operator logic. 4. Map the entity 'kullanıcıları' to the table Kullanicilar.
• Real Answer: SELECT * FROM Kullanicilar WHERE yas > 20;
• Why this demonstrates the capability: This case demonstrates structural divergence handling. In English (SVO), the logic is linear: 'Get (SELECT) users (FROM) older than (>) 20'. In Turkish (SOV), the SELECT action is at the very end and the comparison operator is distributed across a suffix and an adjective, forcing the agent to deconstruct non-linear syntax.

[Case 2]

• Initial Environment: A Turkish-localized aviation database where table/column names use snake_case Turkish terms (e.g., 'ucus_numarasi' for flight_number).
• Real Question: Uçuş numarası 'TK101' olan uçuşun kalkış saatini gösterin.
• Real Trajectory: 1. Locate the terminal verb 'gösterin' (show) for the SELECT clause. 2. Identify the target attribute 'kalkış saatini' and map it to kalkis_saati through semantic translation. 3. Resolve 'uçuş numarası' as the filtering column ucus_numarasi. 4. Preserve the quoted literal 'TK101' exactly for the WHERE predicate as it appears in the database.
• Real Answer: SELECT kalkis_saati FROM ucuslar WHERE ucus_numarasi = 'TK101';
• Why this demonstrates the capability: It highlights the agent's ability to maintain 'Canonical Identity' for literals and identifiers across a localized interface. The agent must translate the natural language intent into localized schema items while ensuring the literal values (which are not translated in the DB) remain untouched, preventing value-linking failures common in multilingual tasks.

[Case 3]

• Initial Environment: An enterprise database where 'Evidence' fields provide domain-specific formulas in Turkish to resolve ambiguous terms like 'productivity rate'.
• Real Question: En yüksek verimlilik oranına sahip personeli listeleyin.
• Real Trajectory: 1. Read the evidence hint: 'verimlilik oranı = tamamlanan_is / calisilan_saat'. 2. Identify 'listeleyin' as the SELECT command for personnel names. 3. Calculate the 'verimlilik oranı' (productivity rate) as a derived scalar math expression using the base columns identified in the evidence. 4. Apply a DESC sort and LIMIT 1 because 'en yüksek' (highest) signifies a ranking objective.
• Real Answer: SELECT personel_adi, CAST(tamamlanan_is AS REAL) / calisilan_saat AS verim FROM personeller ORDER BY verim DESC LIMIT 1;
• Why this demonstrates the capability: This demonstrates 'Evidence-centric reasoning' within a multilingual context. The agent must interpret localized hints provided in the prompt to derive complex SQL logic (division) that isn't explicitly named in the schema, bridging the gap between Turkish business terms and raw numeric columns.

Pipeline Execution Instructions

To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:

1. Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds: references/EXPLORATION.md

2. Phase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: references/SELECTION.md

3. Phase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: references/SYNTHESIS.md