![…
…
…
…
Classification & HS Codes
XX
94
03
30
[Singapore]
…
…
…
…
Chapter
Heading
Sub-Heading
Extension 00
Section
Furniture; bedding, mattresses, mattress supports,
cushions and similar stuffed furnishings; lamps and
lighting fittings, not elsewhere specified or included;
illuminated signs, illuminated name-plates and the
like; prefabricated buildings
Other furniture and parts thereof
Wooden furniture of a kind used in offices
Wooden Furniture for Offices
Miscellaneous Manufactured Articles
Office Table made of Wood
94.03.30.00
9](https://image.slidesharecdn.com/machinelearningapproachesforfraudanalyticsincustoms12oct2020-211020140708/85/Machine-learning-approaches-for-fraud-analytics-in-Customs-8-320.jpg)
Machine learning approaches for fraud analytics in Customs
- 1. 1 Machine Learning Approaches for Fraud Analytics in Customs 35th UN/CEFACT Forum Webinar Advancements in AI towards facilitating cross border paperless trade 12 Oct 2020 Jonathan Koh Trade Facilitation Pte Ltd
- 2. 2 Challenges faced by Customs administrations • Ensuring speed and efficiency in the clearance process for an increasing volume of transactions • Managing change from a few large/bulk shipments into a large number of low- value and small shipments • Managing risks posed by limited knowledge on importers and the e-commerce supply chain • Ensuring data quality (accuracy and adequacy of the data received) • Defining the role and responsibility (liability) of e-commerce operators to assist governments (e-vendors/intermediaries) • Identifying abuse or misuse of ‘de minimus’ for illicit trade purposes (splitting of consignments/undervaluation) • Ensuring compliance with classification and origin rules • Integration of e-commerce vs traditional trade
- 4. Challenges in HS Classification Hurdles to Correct HS Classification Complex HS Classification having total of 98 Chapters, 1244 Headings, 5212 Sub- headings Gaps in Goods Classification between normal commercial terms, words and HS nomenclature Knowledgeable, Experienced Staff in order to give correct, accurate HS Code 4
- 5. Importance of Goods Classification Calculation of duties, taxes and fees is also dependent on ◦ FOB value ◦ Country of origin ◦ Certificates of origin ◦ Exchange rates ◦ Tariff concession codes Permits, Licenses or Certificates to export / import ◦ Customs also act on behalf of Other Government Authorities (OGA) ◦ Examples of permits , licenses, certificates ◦ Health permits ◦ Veterinary certificates ◦ CITES certificates ◦ Phytosanitary certificates Classification determines the regulatory rules to be applied to declarations Collection of National Trade Statistics ◦ Exports ◦ Imports ◦ Transshipments 6
- 6. HS Codes A 6-digit code consisting of three parts: • Chapter, • Heading and • Sub-Heading Arranged in a logical structure, • by section, • used to classify goods for export / import Chapter Heading Sub-Heading 1 2 3 4 5 6 7 8 Extension HS Code 7
- 7. HS Codes – Top-Level Classification SECTI ON TITLE CHAPTE RS I LIVE ANIMALS; ANIMAL PRODUCTS 01 - 05 II VEGETABLE PRODUCTS 06 - 14 III ANIMAL OR VEGETABLE FATS AND OILS AND THEIR CLEAVAGE PRODUCTS; PREPARED EDIBLE FATS; ANIMAL OR VEGETABLE WAXES 15 IV PREPARED FOODSTUFFS; BEVERAGES, SPIRITS AND VINEGAR; TOBACCO AND MANUFACTURED TOBACCO SUBSTITUTES 16 - 24 V MINERAL PRODUCTS 25 - 27 VI PRODUCTS OF THE CHEMICAL OR ALLIED INDUSTRIES 28 - 38 VII PLASTICS AND ARTICLES THEREOF; RUBBER AND ARTICLES THEREOF 39 – 40 VIII RAW HIDES AND SKINS, LEATHER, FURSKINS AND ARTICLES THEREOF; SADDLERY AND HARNESS; TRAVEL GOODS, HANDBAGS AND SIMILAR CONTAINERS; ARTICLES OF ANIMAL GUT (OTHER THAN SILK-WORM GUT) 41 - 43 IX WOOD AND ARTICLES OF WOOD; WOOD CHARCOAL; CORK AND ARTICLES OF CORK; MANUFACTURES OF STRAW, OF ESPARTO OR OF OTHER PLAITING MATERIALS; BASKETWARE AND WICKERWORK 44 - 46 X PULP OF WOOD OR OF OTHER FIBROUS CELLULOSIC MATERIAL; RECOVERED (WASTE AND SCRAP) PAPER OR PAPERBOARD; PAPER AND PAPERBOARD AND ARTICLES THEREOF 47 - 49 XI TEXTILES AND TEXTILE ARTICLES 50 - 63 ** OTHER 98 8
- 8. … … … … Classification & HS Codes XX 94 03 30 [Singapore] … … … … Chapter Heading Sub-Heading Extension 00 Section Furniture; bedding, mattresses, mattress supports, cushions and similar stuffed furnishings; lamps and lighting fittings, not elsewhere specified or included; illuminated signs, illuminated name-plates and the like; prefabricated buildings Other furniture and parts thereof Wooden furniture of a kind used in offices Wooden Furniture for Offices Miscellaneous Manufactured Articles Office Table made of Wood 94.03.30.00 9
- 9. Example of an Invoice & Goods Description 10
- 11. Applying Machine Learning (ML) for HS Goods Classification Machine Learning - Reduce costs - Reduce time - Learning capability (incremental) - Less prone to errors - Automated maintenance Conventional approach - Rule Based - High costs - Time consuming - Tedious troubleshoot - Error prone - Manual maintenance 12
- 12. Efforts in Applying ML for HS Goods Classification Brazil – The Customs Selection System through Machine Learning (SISAM) is an artificial intelligence which learns from historical import declarations and help to reduce the percentage of goods verified during customs clearance and thus reduce tax evasion and noncompliance with administrative requirements like failing to obtain consents from the ministries of health, agriculture or defence. Source : http://www.jambeiro.com.br/jorgefilho/sisam_mono_eng.pdf China – This effort proposes a text-image adaptive convolutional neural network to effectively utilize website information and facilitate the customs classification process, helping cross-border e-commerce sellers to classify goods efficiently. Source: “Customs classification for cross-border e-commerce based on text-image adaptive convolutional neural network” by Guo Li & Na Li , School of Management and Economics, Beijing Institute of Technology, Beijing The Netherlands – The effort sought to classify products descriptions with convolutional neural networks (CNNs), a deep learning technique for text classification with good success, and we obtain domain-specfiic word embeddings by iterative pre-training on context-rich, domain-specific corpora we obtain from DBpedia, scraping Wikipedia and public data sets. The corpora we create contain almost two billion tokens. Source – Masters thesis by Classifying Short Text for the Harmonized System with Convolutional Neural Networks - Jeffrey Luppes 13
- 13. Efforts in Applying ML for HS Goods Classification Singapore – This effort aims to produce an auto-categorization system is expected to accurately classify products into HS codes based on the text description of the goods declaration using background nets approach as the key technique for the development of classification engine in the system. Source : “Auto-Categorization of HS Code Using Background Net Approach” Liya Dinga et al , Institute of Systems Science, National University of Singapore 14
- 14. Applying Machine Learning (ML) for HS Goods Classification § Classify the products according to HS System based on a given description accurately and efficiently § A highly flexible and adaptable ML engine with self-learning capability from historical data, catering to continuous changing product descriptions § Able to learn from product descriptions written in different languages Commercial Descs • CHEVALIER BLANC-DE BLANCS 6X0.75LTR Classification Engine • Self- Learning Classifier HS Codes • 22041000 15
- 15. Data Integration Engine Historical Data Integration Rule Set Integration Data Selection and Trans- formation Data Cleanin g NLP C HS Classification Mechanics Core HS Classification Engine Data Source Data Integration Data Preparation Modeling Deployment Training Data DB Other Data Sources Excel CSV Text File Machine Learning C Web Service Web Application Rule Engine Ad hoc Rule Processing Rule Indexing 16
- 16. Key Features/Capabilities • By HS Code (2/4/6/8 digits) • By Chemical Abstract Number (CAS) • By Product Code, Name • By HS Alphabetical Index (HSAI) Search • Multi-language Support • Understand use of localised words & terms Localisation • Understand meaning of whole sentence in query • Ability to detect noise in query and omit in search • Detect spelling errors and Auto correct Semantics 17
- 17. Benefits For Traders • More Accurate Classification of Goods • Save Time, Costs • Faster Clearance & Approval For Government Agencies • Better Compliance from Trading Community • Better Risk Assessment • Prevention of Fraud • Better Statistics 18
- 19. Support Vector Machine (SVM) 20 Support-vector Machines (SVMs), are supervised learning models with associated learning algorithms that analyze data used for classification and regression analysis. Applications includes: • Face recognition: SVM classifies parts of the image as face and non-face and draws a square box around it. • Hypertext and text categorization: SVM will allow both hypertext and text categorization. By using training data it classifies documents into categories. It differentiates by using score generated and then compares with the threshold value. • Image classification: It provides great search accuracy for classifying the image. SVM will give better accuracy comparing with traditional query-based search algorithms. • Recognizing handwriting: We use SVM to recognize handwritten characters.
- 20. Natural Language Processing 21 Text mining is an artificial intelligence (AI) technology that uses natural language processing (NLP) to transform the free (unstructured) text in documents and databases into normalized, structured data suitable for analysis or to drive machine learning (ML) algorithms
- 21. Anatomy of an Auto-Classifier Named Entity Extraction Lexical Processing Feature- Vector Extension Statistical Learner Training Corpus Goods Descrip- tions Goods Classification + Declaration Attributes Statistical KB Named Entity Extraction Lexical Processing Feature- Vector Extension Statistical Classifier Test Corpus Goods Descrip- tions Declaration Attributes Goods Classifi- cation Trade Ontology Trade Lexicon Training Classifying
- 22. Results
- 23. 24 Test Examples : 1.264 carton disposable medical products: - blood bag Able to catch the meaning of whole sentence other than just keywords Semantic Understanding
- 24. 25
- 25. Semantic Understanding 26 Examples: 1.ETHYL GLYCOL, INDUSTRIAL GRADE INV:00082 ◦ Able to detect “noise” within description
- 26. 27
- 27. CAS Number Detection 28 Test Examples: 1.96392-96-0 ç, 2.41183-64-6, 3.80809-81-0 Able to understand and classify Chemical Abstracts Service (CAS) number instantly
- 28. 29
- 29. Product Code 30 Test Examples: 1.ZDP0AW0PM00 ç 2.HFT0PA 3.ZCP0AB0PN00 Able to understand and classify Product code instantly
- 30. 31
- 31. Product name 32 Test Examples: 1.KINOHIMITSU HEALTH PAD ç , 2.OSIM massage equipment and product Able to classify even with product brand name
- 32. 33
- 33. Product name 34 Test Examples: 1.KINOHIMITSU HEALTH PAD , 2.OSIM massage equipment and product ç Able to classify even with product brand name
- 34. 35
- 35. Localization 36 Test Examples: 1.Chicken satay ç 2.“Old Town” White Coffee 3 in 1 Able to understand local words and terms
- 36. 37
- 37. Localization 38 Test Examples: 1.Chicken satay 2.“Old Town” White Coffee 3 in 1 ç Able to understand local words and terms
- 38. 39
- 39. Auto-spelling Correction 40 Test Examples: 1.Fuji bpple ç, 2.cottno, 3.massag chair Able to correct different wrong spelling
- 40. 41
- 41. 42
- 42. Classification with Vague Descriptions 43 Test Examples: Heading Level Description: water pump Able to understand vague information and classify depend on heading level description
- 43. 44
- 44. 45
- 45. Demo: classification with vague descriptions 46 Test Examples: Chapter level Description: 1.Cottons ß, 2.Zinc alloy, 3.glasses, 4. cereals Able to understand chapter level information and give the relevent result
- 46. 47
- 47. 48
- 48. Declaration Validation Help flag declarations with wrong or insufficiently clear description: ◦ By reading the description submitted by declarants. For example: ◦ By matching the description entered by declarants vs. those in the supporting documents. Help flag suspicious declarations: ◦ By reading the description submitted by declarants and determine whether the HS Code entered was misclassified. This will result in better risk assessment as well as better trade statistics. 49
- 49. Post Clearance Audit During Post Clearance Audit, Customs officers may run supporting documents through ALICE as a form of validation against old declarations: ◦ ALICE will alert the officers if the supporting document’s goods description does not match the goods being declared This will help in Post Clearance Audit work. 50
- 50. Valuation ALICE could be deployed as a valuation engine: ◦ ALICE will learn the acceptable range of value for a HS Code from various countries and later on flag those declarations which are out of the acceptable range. ◦ For example, if a machinery from Germany usually cost $ Xx per unit, and the same machinery from China cost $y per unit, ALICE will flag a declaration of the machinery from Germany with value of $ Yy. This will help in Customs valuation and prevent revenue lost. 51
- 51. Abnormality Detection ALICE could be deployed as an engine for abnormality detection. For example: ◦ ALICE will learn the import or export pattern associated with a trader and it will be able to flag out abnormality if the trader trade something very different from previously. For example IBM importing live fish. This will help in Customs risk assessment. 52
- 52. Future of CBT – Where are the Game Changers?
- 53. Auto-Classification Copyright 2012 Private & Confidential 54
- 54. Auto-Classification Needs … Corpus (collection of texts) of Classified Goods Descriptions with relevant attributes ◦ Training Statistical Learner ◦ Building Knowledge-Bases & Lexicons Knowledge Base of WCO HS Codes ◦ Hierarchy of HS Coding Concepts (Section, Chapter, Headings, Sub-Headings, etc.) ◦ Country-Specific Extensions to HS Codes Specialized Lexicon ◦ Concepts used in Trade and Goods Descriptions ◦ Names (Proper Nouns), e.g. “Intel”, “Herman Miller”, etc. Country of Destination Specialised Ontology of Concepts ◦ Explicit relations (e.g. is-a, part-of, etc.) between words / concepts Natural Language Processing Capability ◦ Named Entity Extraction ◦ Expansion of Extracted Concepts “Feature-Vectors” using specialized lexicon and ontology, and country of destination Statistically derived associations between feature vectors and Goods Classifications
- 55. Deploying Auto-Classification Needs Deployment / Operational Scenarios ◦ Suggest Goods Classifications rather than authoritatively provide them ◦ Embed in Goods Classification Wizards that provide other helpful services ◦ Deploy Classification Wizards as a SOA API that can be invoked from within Private Sector IT systems’ workflows Zero Declaration Clearance Regulatory Regime ◦ Training ◦ Enlisting Community Support ◦ Supporting the Community ◦ Communications Changes to existing regulations to … ◦ Set up equipped & certified private entities that embed wizards ◦ Legalize non-declaration via certified entities ◦ Specify inspections and corrective / enforcement actions regulators may carry out post (implicit) declaration using source documents only 56
- 56. Current Process Customs Documents 1. Customs Declaration 2. Invoice 3. Packing List 4. Bill of Lading 1 Invoice 2 Packing List 3 Bill of Lading 4 Customs Declaration Carrier Shipper Freight Forwarder (Customs Broker) Customs 57
- 57. AutoClass Generate HS Codes 1 Invoice 2 Packing List 3 Bill of Lading 4 Customs Declaration Possible Solution 1 Trader (Exporter/ Importer) Trader 1. Scan Doc 2. E-Doc e.g. XML 3. Web forms Customs Pre- Declaration Copyright 2012 Private & Confidential 58 Source Doc as Declaration
- 58. AutoClass Generate HS Codes 1 Customs Declaration Possible Solution 2 Trader (Exporter/ Importer) Verification (document inspection) upon request 2 Invoice 3 Packing List 4 Bill of Lading Trader Scan Doc E-Doc e.g. XML Web forms Customs Pre- Declaration Simplified Semi-Automatic Declaration 59
- 59. Future ? Proposal: Eliminate Declarations (in the Normal Case) ◦ 21st Century Regulatory Regime presuming ◦ Cross-Sectoral (Public-Private) collaboration ◦ Underlying eGovernment systems able to accept and process source trade / logistics documents in lieu of formal declarations (in the normative case) Key Assumptions ◦ Private Sector IT systems ◦ In active use ◦ will embed Public Sector Controls in their normal workflows ◦ Public Sector IT (eGov) systems will be able to ◦ Accept (i.e. interoperate with) Private Sector IT systems ◦ Process source documents’ – including their free-text goods descriptions Copyright
- 60. Steps towards Auto-Classification Copyright Step 1 Collect Classified (w/ HS Codes) Goods Description Corpus (Copora) Step 2 Build PoC to classify goods using commercial descriptions – including national classification codes Step 3 Provide traders with “Fail Soft” automated assistance (suggest HS Codes) plus other value-added services (e.g. giving regulatory advice) Step 5 Allow traders to check goods classification before “legally” submitting declarations Step 6 Within 10 years, there may be no need for traders to check declarations, unless there is a need to prompt for more information
- 61. 62 2017 Private & Confidential – Trade Facilitation Pte Ltd Trade Facilitation Pte Ltd 160 PAYA LEBAR ROAD #02-04 , ORION @ PAYALEBAR SINGAPORE 409022 Jonathan Koh Managing Director TATTSEN@GMAIL.COM +65 82330321 Thank you.