Bilingual Named Entity Recognition for Cargo Logistics Using Model Distillation

ibs_software · Other · 2026 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Bilingual cargo logistics entity extraction

IBS Software needed a production NER system that could extract 23 cargo logistics entity types from thousands of daily English and Japanese emails while keeping accuracy, latency, and inference cost suitable for real-time operations.

Constraints:

• emails arrive in both English and Japanese
• the system must identify 23 entity types
• manual intervention slowed operations
• inference cost mattered at production volume
• Japanese commodity descriptions and missing word boundaries made entity boundaries harder

Architecture

IBS used Amazon Bedrock managed model distillation to transfer a Nova Pro teacher model into a cheaper Nova Lite student model. The production pipeline ingests .eml files in S3, uses Lambda for preprocessing, invokes the distilled Bedrock model for NER, applies confidence and validation rules, and stores structured extracted entities in DynamoDB.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  cargo-email-source[Cargo email source] --> amazon-s3-ingestion[Amazon S3 ingestion]
  amazon-s3-ingestion --> aws-lambda-preprocessing[AWS Lambda preprocessing]
  aws-lambda-preprocessing --> bedrock-nova-lite-endpoint[Bedrock Nova Lite endpoint]
  distillation-training-set[Distillation training set] --> bedrock-nova-lite-endpoint
  bedrock-nova-lite-endpoint --> post-processing-rules[Post-processing rules]
  post-processing-rules --> dynamodb-output[DynamoDB output]

Components

Data flow

Cargo emails arrive as .eml files in S3. Lambda extracts email body and metadata, the Bedrock endpoint runs the distilled Nova Lite model, the model returns 23 entity types with confidence scores, post-processing applies validation rules and thresholds, and structured JSON entities are written to DynamoDB.

Why this architecture

Managed distillation let IBS keep most of the teacher model's extraction quality while cutting inference cost and avoiding custom open-source training/deployment infrastructure. Serverless AWS components fit the email workflow and allowed the team to add validation and human review around the distilled model.

Models used: Amazon Nova Pro · Amazon Nova Lite

Scale

• The source says the Cargo system processes thousands of bilingual cargo logistics emails daily.
  “IBS Software’s Cargo system processes thousands of bilingual cargo logistics email messages daily.”article, stated
• The model extracts 23 entity types across English and Japanese.
  “The goal was to build a bilingual NER system capable of extracting 23 entity types from cargo logistics email messages written in English and Japanese.”article, stated
• The team annotated 500 messages split across 350 English and 150 Japanese emails.
  “* Annotated 500 cargo email messages (350 English, 150 Japanese) with 23 entity types. * Configured Amazon Bedrock distillation with custom hyperparameters.”article, stated
• The production pipeline processes messages in under 2 seconds.
  “The entire pipeline processes email messages in under 2 seconds with 95.085 percent accuracy, meeting our real-time processing requirements.”article, stated

Alternatives considered (roads not taken)

Use open-source distillation frameworks such as PyTorch and TextBrewer

Rejected because: They created complexity around bilingual distillation, hyperparameters, infrastructure, and production workflow integration.

Serve the larger Nova Pro teacher model directly

Rejected because: The Nova Lite student retained most of the teacher performance while reducing inference cost 14x.

Trust model output without post-processing

Rejected because: The team added rules and confidence thresholds for known formats and low-confidence predictions.

Patterns

model-distillation (model-ops)

A Nova Pro teacher was distilled into a Nova Lite student for cheaper NER inference.

confidence-thresholding (quality-control)

Low-confidence entity predictions are flagged for human review.

human-in-the-loop-gate (governance)

Low-confidence predictions route to human review rather than being accepted automatically.

structured-output-contract (interface-design)

The pipeline emits structured JSON containing extracted cargo entities.

Tradeoffs

Teacher/student model choice

Chose distilled Nova Lite over direct Nova Pro production inference.

Rationale: The student retained 98 percent of teacher performance while reducing cost 14x.

Managed training platform

Chose Amazon Bedrock Model Distillation over PyTorch/TextBrewer custom pipelines.

Rationale: Managed infrastructure simplified training, monitoring, evaluation, and workflow integration.

Gotchas (what bit them)

Open-source distillation added too much production friction.

Root cause: Bilingual token-level distillation, hyperparameter tuning, infrastructure, and workflow integration were all hard to manage manually.

Resolution: Pivot to Amazon Bedrock managed distillation.

Japanese extraction accuracy lagged English.

Root cause: Kanji combinations, absent spaces, fewer Japanese examples, and embedded multi-line entities made boundaries ambiguous.

Resolution: Augment Japanese data and add post-processing plus confidence thresholds.

Evaluation

IBS evaluated the teacher and student models with F1-score on a manually annotated bilingual test set, monitored training loss, analyzed language-specific error gaps, and checked real-time processing latency.

Metrics: 95.085 percent F1-Score · 98 percent teacher performance retained · 14x cost reduction · loss 0.05 to 0.008 · under 2 seconds processing

Outcomes

• The distilled model achieved 95.085 percent F1-score.
• Operational cost fell 14x versus the larger teacher model.
• The production pipeline met under-2-second real-time processing requirements.

The system processes thousands of daily bilingual cargo logistics emails and extracts 23 entity types.