Design and Implementation of the 58.com Intelligent Article Writing Robot

The background explains that traditional public‑relations copywriting is labor‑intensive and costly, prompting the development of an intelligent writing robot for 58.com’s PR department to automate article creation across multiple business lines.

The generation pipeline consists of three steps: material acquisition (including manually written templates and structured data), body generation, and title generation. Materials such as opening/closing sentences and templates are stored in a material library, while structured attributes (e.g., car brand, model, displacement) are fetched via API and inserted into the templates.

For title generation, the team first collected a dataset of human‑written titles and corresponding bodies, training a seq2seq encoder‑decoder model with attention and BeamSearch decoding. BLEU was used as the evaluation metric. The first‑version model produced fluent titles but often mismatched key slot values (e.g., wrong car brand).

To address this, the second‑version model introduced slot‑name replacement: during training, actual slot values in titles were replaced with placeholder names, and at inference time the placeholders are swapped back with the correct values. This approach retained fluency while greatly improving consistency between title and body.

Experimental results show that although both models have relatively low BLEU scores (first version 0.3103, second version 0.2725), the second version achieves better alignment with the article content, which is more important for the PR use case. The second model is now deployed online and serves as a reference for human writers.

Body generation follows a template‑based method: each article class has multiple long‑text templates containing slots for structured data. After filling slots with the appropriate values and inserting 4–6 relevant images, the final article body is assembled.

In summary, the 58.com intelligent writing robot combines deep‑learning title generation and template‑driven body construction to produce thousands of articles weekly, reaching millions of reads. Future work includes incorporating click‑through and recommendation metrics into the title model and expanding the system to more business categories.