Introducing NAVER Place's AI Development Team

NAVER GLACE CIC operates O2O services such as NAVER PLACE, Reservations, MY PLACE, LINE PLACE, and SmartPlace in Korea and Japan. NAVER PLACE is a core service that connects businesses and users, built on a place database that has accumulated hundreds of millions of reviews.

The AI development team's goal is to leverage the data accumulated in GLACE to enhance service support and TAG extraction for search. Currently, the team operates 15 or more AI models across CV and NLP domains, improving actual service quality and operational stability.

The main NLP models are as follows.

• OCR Place Matching: Uses receipt OCR text as a query to find the actual visited place among millions of places. It combines NAVER CLOVA OCR with NER-based parsing, and is designed to deliver strong search performance even on text with many typos and noise.
• Menu Matching: Corrects OCR noise in menu names and adds new menus to the DB to improve search accuracy. It also operates a Sequence Labeling model that classifies raw menu text into 9 options such as Size, Price, and Temperature.

In the CV domain, models are applied directly to service operations.

• Receipt Classifier: Filters out fake receipts submitted for rewards or advertising purposes in real time.
• Food Classifier: Classifies images of 300+ types of food from Korea and Japan to increase review utility.
• Image Scoring: Scores review images on a 0–4 scale to prioritize high-quality images in the Discovery tab. On the test set, prediction accuracy for scores 3 and 4 reached 99% or higher.
• Object Detection: Filters out review images with privacy issues or inappropriate content.
• Atmosphere Classifier: Extracts 50+ atmosphere features from review images, used to help find stores with the desired atmosphere.

The project currently receiving the most focus is PlaceLM, an LLM dedicated to GLACE. Since running separate models per tag increases cost and makes it hard to extract complex semantic tags, the team chose to handle both existing tags and semantic tags together with a single model. PlaceLM is a 12.8B-scale model designed to lower operating costs while handling frequent place information updates and high traffic.

Training proceeds in three stages.

• Pretrained Language Model: Learns general knowledge from a large-scale corpus.
• Supervised Fine-Tuning: Aligns response quality using high-quality instruction data for the PLACE domain.
• Preference Optimization: Reflects human preferences to reduce odd outputs.

Pretraining used 71.23B tokens, mainly NAVER data, trained on 100 A100/V100 GPUs, taking 14 days per epoch for a total of 2 epochs completed. SFT was conducted on 622,426 documents using LoRA and QLoRA, on 8 A100 GPUs over 2 days. Preference Optimization is currently in progress, with Chosen and Rejected responses constructed from 31,523 documents.

Evaluation is divided into public tasks and internal tasks. Public tasks use NSMC, KorQuAD, and GPT-4-based G-EVAL, while internal tasks are converted into prompts and compared against GPT-4 using ROUGE. On public benchmarks, PlaceLM PLM showed the best performance among public Korean LLMs, and in GPT-4-based evaluation it performed close to Kullm's level. On internal tasks, it achieved a top Rouge-L of 0.823.

On the operations side, the team uses DVC to version-control training data, strengthening reusability and history management, and uses MLFlow tracking and model registry to quickly track training results and their reflection in the service. Going forward, plans include adding Japanese data to expand coverage, and extending PlaceLM into a VLM by using it as a Language Encoder.