代码之家  ›  专栏  ›  技术社区  ›  Michał Ziobro

Swift自然语言和CoreML:如何提高NLTagger的读卡能力

  •  0
  • Michał Ziobro  · 技术社区  · 6 年前

    我正在使用自然语言框架在信用卡上查找个人姓名。

    "Citi 6011 1111 1111 1117 07/25 ELON MUSK Discover Debit"
    

    我已经厌倦了在这样的字符串中简单地查找.personalName标签

    let tagger = NLTagger(tagSchemes: [.nameType])
        tagger.string = text
        tagger.setLanguage(.english, range: range)
    
        var fullNames = [String]()
    
        // 1) personalName
        let options : NLTagger.Options = [.omitPunctuation, .omitWhitespace, .omitOther, .joinNames]
        let foundTags = tagger.tags(in: range, unit: .word, scheme: .nameType, options: options)
    
        foundTags.forEach { (tag, tokenRange) in
            if tag == .personalName {
                let name = text[tokenRange]
                fullNames.append(String(name))
            }
        }
    

    因此,我尝试在macOS playeground中使用CreateML创建CoreML模型,该模型将用于标记此类字符串中的自定义持卡人标签。

    现在我有这样一个相对简单的案例

    [
        {
            "tokens": ["Inteligo", "4242 4242 4242 4242", "09/21", "JOHN SMITH", "VISA", "Debit"],
            "labels": ["ORG", "NONE", "NONE", "CARDHOLDER", "ORG", "NONE"]
        },
        {
            "tokens": ["mBank", "4000 0566 5566 5556", "10/23", "STEVE JOBS", "VISA", "Debit"],
            "labels": ["ORG", "NONE", "NONE", "CARDHOLDER", "ORG", "NONE"]
        },
        {
            "tokens": ["ING", "5555 5555 5555 4444", "03/22", "EMMA WATSON", "mastercard", "Debit"],
            "labels": ["ORG", "NONE", "NONE", "CARDHOLDER", "ORG", "NONE"]
        },
        {
            "tokens": ["Bank of America", "3782 822463 10005", "05/24", "JULIA ROBERTS", "American Express", "Debit"],
            "labels": ["ORG", "NONE", "NONE", "CARDHOLDER", "ORG", "NONE"]
        },
        {
            "tokens": ["Citi", "6011 1111 1111 1117", "07/25", "ELON MUSK", "Discover", "Debit"],
            "labels": ["ORG", "NONE", "NONE", "CARDHOLDER", "ORG", "NONE"]
        }
    ]
    

    然后,我从中创建CoreML模型

    import Foundation
    import CreateML
    
    let trainingData = try MLDataTable(contentsOf:
        Bundle.main.url(forResource: "data", withExtension: "json")!)
    
    let model = try MLWordTagger(trainingData: trainingData, tokenColumn: "tokens", labelColumn: "labels")
    
    try model.write(to: URL(fileURLWithPath: "#path/to/Desktop/creditcardtagger.mlmodel"))
    

    let tagger = CreditCardTagger.shared
            tagger.string = text
            tagger.setLanguage(.english, range: range)
    tagger.enumerateTags(in: range, unit: .word, scheme: CreditCardTagger.Scheme, options: options) { (tag, tokenRange) -> Bool in
    
                if tag == CreditCardTagger.cardHolderTag {
                    let cardHolder = text[tokenRange]
                    print("CARD HOLDER: \(cardHolder)")
                }
                return true
            }
    

    但我认为我用于培训的数据是不够的。你知道覆盖大多数信用卡案例需要多少这样的数据记录吗?

    0 回复  |  直到 6 年前
    推荐文章