农业设备JSON数据NLP分类：自定义模型选型与优化问询

输入数据说明 #

外部供应商提供的农业设备JSON数据格式如下：

{
    "EquipmentList": [
        {
            "EquipmentClass": "1",
            "EquipmentID": "1234",  
            "EquipmentGroup": "PLG",
            "Description": "Advanced Electrical Plough, Bosch"
        },
        {
            "EquipmentID": "1235",  
            "EquipmentGroup": "MOT",
            "Description": "Small Gasoline Motor, Honda, 250 kW 1.2L"
        },
        {
            "EquipmentID": "1236",
            "Description": "Leather Grip Handle"
        }
    ]
}

其中Description和EquipmentID为必选字段，其余字段为可选。

目标标签定义 #

需要将上述数据映射到以下预定义标签：

当前尝试的问题 #

由于不同供应商描述方式差异大，正则/字符串规则维护成本高，尝试用Hugging Face Transformers的4款预训练模型做零样本分类，代码如下：

from transformers import pipeline
from transformers.utils import logging

logging.set_verbosity_error()
logger = logging.get_logger("transformers")
logger.info("INFO")
logger.warning("WARN")

# model = "bert-base-german-cased"
model = "bert-base-uncased"
# model = "xlm-roberta-base"
# model = "distilbert-base-uncased"
classifier = pipeline("zero-shot-classification", model=model)
possible_plough_motor_vendors = ["Bosch", "Honda", "Shimano", "Volkswagen"]
print("Classification Result: ", classifier(json_src_data[1]["Description"], possible_plough_motor_vendors))

但结果准确率极低，平均置信度仅约3.8%。

需求 #

希望训练自定义模型，而非BERT这类适用于长句的预训练微调模型，求针对该场景的最佳方法、库或平台。

内容的提问来源于stack exchange，提问作者Bijay Regmi