automatic-post-editing

# Skill: automatic-post-editing ## 1. Capability Definition & Real Case * **Professional Definition**: Automatic Post-Editing (APE) is the capability of an agent to refine and correct a machine-generated translation hypothesis (draft) by using the original source text as the ground truth. It focuses on identifying and rectifying semantic drift (where the meaning shifts away from the source), syntactic errors, terminology inconsistencies, and stylistic awkwardness in the candidate output to produce a high-fidelity, professional-grade translation. * **Dimension Hierarchy**: Translation Quality Maintenance and Refinement -> Post-Editing and Error Correction -> automatic-post-editing ### Real Case **[Case 1]** * **Initial Environment**: A translation environment where the agent is provided with an English source sentence and a candidate Chinese translation that was produced by a lower-quality machine translation model. * **Real Question**: Improve the following machine-generated translation from English to Chinese. Correct errors and generate a more accurate translation. English: 'unfazed, Matteo hires an assassin to bomb the resort to create chaos and mayhem.' Chinese (Draft): '马特奥不慌不忙地雇用了一名刺客来轰炸制造混乱和混乱的手段.' * **Real Trajectory**: The agent first parses the English source and identifies 'resort' as a key entity missed in the draft. It notices that the Chinese draft redundantly repeats 'chaos' ('混乱和混乱') and ends with an nonsensical 'means/method' ('手段'). It subsequently removes the repetition, identifies 'resort' as '度假村', and restructures the sentence for better flow. * **Real Answer**: 马特奥不慌不忙地雇用了一名刺客来轰炸度假村，从而引起混乱。 * **Why this demonstrates the capability**: This case demonstrates post-editing by showing how the agent identifies specific failures in a draft—such as entity omission and lexical repetition—and uses the source text to reconstruct a version that is semantically faithful and stylistically superior. --- **[Case 2]** * **Initial Environment**: A professional medical translation scenario where a raw translation of a clinical report needs refinement to meet domain-specific terminology standards. * **Real Question**: Improve the following machine-generated translation from English into Chinese: English: 'The median age of the 30 patients was 56.5 (28-80) years old.' Chinese (Draft): '30名病人的中间年龄是56.5（28-80）岁。' * **Real Trajectory**: The agent analyzes the medical source and identifies that 'median age' is a term of art. It observes that the draft uses the colloquial '中间年龄' (middle age) instead of the professional '中位年龄' (median age) and uses a generic classifier '名' for patients. It performs a substitution for the professional term and adjusts the classifier to '例' for a formal medical register. * **Real Answer**: 30例PCL患者中位年龄56.5（28-80）岁。 * **Why this demonstrates the capability**: This case proves the agent's ability to move beyond literal correction to domain-specific refinement. It shows the agent can map general terms in a draft to specific professional standards based on the source context, which is a core requirement of expert post-editing. --- **[Case 3]** * **Initial Environment**: A translation validation task for an Indian language (Hindi) search query where an LLM-generated draft contains a semantic drift error regarding statistical frequency. * **Real Question**: The following Hindi translation was generated by a model. Please post-edit it against the English source to ensure semantic and statistical accuracy. English: 'What blood type is most common?' Hindi (Draft): 'किस खून के प्रकार सबसे आम है?' * **Real Trajectory**: The agent identifies the source 'most common' as referring to statistical frequency in a population. It observes that the Hindi draft uses the word 'आम' (aam), which normally translates to 'ordinary' or 'commonplace' in a general sense, rather than the statistical 'most frequent'. The agent corrects the draft to use 'सबसे अधिक बार' (most frequent/occur most often) to align with search intent and linguistic norms for technical data. * **Real Answer**: कौन सा रक्त प्रकार सबसे अधिक बार होता है * **Why this demonstrates the capability**: This demonstrates the ability to detect and fix 'semantic drift'—a common failure in neural translation where a word's high-frequency literal meaning (ordinary) is used instead of its context-specific meaning (frequent). The agent uses the source intent to select the precise Indic language term of art. ## 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`