本文为数字化转型网(szhzxw.cn）为大模型专题系列文章之一，本文讲述了大语言模型的三大技术路线是什么？大语言模型在客户联络领域的应用价值有哪些？
一、大模型的三大技术路线
大模型的三大技术路线是哪三大路线？
大型语言模型研究的发展有三条技术路线:Bert模式、GPT模式、混合模式。其中国内大多采用混合模式，多数主流大型语言模型走的是GPT技术路线，直到2022年底在GPT-3.5的基础上产生了ChatGPT.
可以看到，到2019年后，Bet路线基本上就没有什么标志性的新模型出现了，而GPT技术路线趋于繁荣，从Bet往GPT走，模型越来越大，做的事越来越通用。
大型语言模型按照从数据到知识来划分，数据可分为通用数据和领域数据，知识分为语言知识和世界知识，从任务类型来划分，可以分为单一任务和多任务、理解类和生成类。
1、Bert模式
Bert模式有两阶段(双向语言模型预训练+任务Fine-tuning)，适用于理解类、做理解类、某个场景的具体任务，专而轻。
2、GPT模式
GPT模式是由两阶段到一阶段(单向语言模型预训练+zero shot prompt/lnstruct)，比较适合生成类任务、多任务，重而通.
3、T5模式
T5模式将两者的方法结合，有两阶段(单向语言模型预训练+Fine-tuning) 。张俊林称这种模式”形似GPT，神似Bert”，生成和理解都行，从效果上看较适合理解类任务，国内很多大型语言模型采用这种模式。
目前的研究结论是，如果模型规模不是特别大，面向单一领域的理解类任务，适合用T5模式。做生成类任务时，用GPT模式效果最好。
如果单独考虑zero-shot，GPT模式效果最好，如果在预训练后引入多任务ine-tuning，则T5模式效果好。不过张俊林认为这个结论存疑，因为目前的实验Encoder-Decoder都是Decoder-only参数量的两倍。
综合来看，当前几乎所有参数规模超过千亿的大型语言模型都采取GPT模式张俊林分析可能的原因有三点:
1、Encoder-Decoder里的双向attention，损害zero shot能力;
2、Encoder-Decoder结构在生成oken时，只能对Encoder高层做attentin， Decoder-only结构在生成Token时以逐层Attention，信息更细粒度:
3、Encoder-Decoder训练“中间填空”，生成最后单词Next Token，存在不一致性，Decoder-onlv结构训练和生成方式一致。
二、大语言模式在客户联络领域的应用价值
大语言模式在客户联络领域的应用价值有哪些？
1.提升自动回复能力
大模型的工作逻辑始于对大量数据的持续训练。分、持续的训练，能够使大模型具备更加精准的语义理解能力和更强大的自然语言生成能力。基于已经训练成熟的大模型，智能客服系统的开发就有了更加坚实的底层支撑。它可以根据用户输入的问题提供快速和准确的响应，快速解决问题节省了客服团队大量的时间和资源，提高客户体验和满意度。
2.强化意图识别能力
智能客服能否处理复杂问题，在行业内有一个通用的指标，就是意图识别的准确率。
观察客户联络领域所处的现状，大部分是把简单、重复、流程性的问题，交给机器人处理，复杂的、需要情感关怀的问题，则交由人工客服处理而传统的智能客服在意图理解方面的能力，仍然相对薄弱。
ChatGPT的泛化为我们提供了处理复杂问题的新思路.其于对文本语音、图像等多模态数据的分析，大模型对于意图识别的准确性进一步提升。借助大模型，智能客服能够有效结合用户的历史对话当前沟通内容等上下文语境，更准地识别出用户的需求和意图。同时，借助大模型所具备的深度学习能力，进行更加智能化的问答推荐，进而有效赋能企业的业务咨询、留资引导、服务应答等环节。
3.优化人机交互体验
传统机器人在处理复杂场景的时候，往往应变能力不够灵活.一旦用户问的问题在知识库里没有，或者超出了预设的流程，机器人就无法很好地应对了。但是，基于大模型超强的知识库，上述情况就缓解了很多。以ChatGPT为例来看，大模型的深度应用也开创了客户使用体验的新范本，其丰富的参数和强大的内容生成能力，能够支持智能客服实现更加个性化的问答回复，而非过往千篇一律的机械式问答。
4.丰富实际应用场景
基于大模型所提供的底层能力，智能客服的渗透力和应用场景也将在未来得到进一步延伸。
ChatGPT的应用目前已经有相对确定的场景可以落地了，如扮演人工客服与客户沟通专业知识，提供专业的问答知识建议、对沟通记录进行质检标记主动分析座席工作行为、发起产品推介、闲聊寒暄以及更“人性化”的引导留资等。
此外，在情绪关怀方面也有很大的应用前景。
更多有关大模型相关文章，例如大模型资料、大模型成功案例、大模型方案等可在数字化转型网大模型专栏里进行阅读；如果您需要大模型相关资料包报可点击蓝字填写问卷与我们取得联系；更多精彩内容可进入读者群与百万读者一同交流~
大模型专题栏目文章（四）：大语言模型的三大技术路线、大语言模型在客户联络领域的应用价值。翻译：
This article is one of a series of articles on Digital Transformation network (szhzxw.cn) for Big model. This article tells what are the three technical routes of big language models? What are the application values of large language models in the field of customer contact?
First, the three major technical routes of large models
What are the three major technical routes of large models?
The development of large-scale language model research has three technical routes :Bert model, GPT model and mixed model. Most of them adopt mixed mode in China, and most of the mainstream large language models take the GPT technology route, until the end of 2022, ChatGPT is produced on the basis of GPT-3.5.
It can be seen that after 2019, the Bet route basically has no iconic new model, and the GPT technology route tends to prosper, from Bet to GPT, the model is getting bigger and bigger, and the things are becoming more and more common.
Large language models can be divided from data to knowledge, data can be divided into general data and domain data, knowledge can be divided into language knowledge and world knowledge, and task types can be divided into single task and multi-task, understanding class and generation class.
Bert mode
Bert mode has two stages (bidirectional language model pre-training + task Fine-tuning), which is suitable for understanding the class, doing the specific task of understanding the class, and a certain scene.
GPT mode
GPT mode is from two stages to one stage (one-way language model pre-training +zero shot prompt/lnstruct), more suitable for generating class tasks, multi-tasks, heavy and pass.
T5 mode
The T5 mode combines the two approaches and has two phases (one-way language model pre-training +Fine-tuning). Zhang Junlin said that this model “looks like GPT, like Bert”, both generation and understanding, from the effect point of view is more suitable for understanding tasks, many large language models adopt this model.
The current research conclusion is that if the model size is not particularly large, it is suitable to use T5 model for understanding tasks in a single domain. GPT mode works best when doing class generation tasks.
GPT mode works best if zero-shot is considered alone, and T5 mode works best if multi-tasking ine tuning is introduced after pre-training. However, Zhang Junlin believes that this conclusion is doubtful, because the current experiment Encoder-Decoder has twice the number of Decoder-only parameters.
In summary, at present, almost all large-scale language models with more than 100 billion parameters adopt the GPT model. There are three possible reasons:
- Two-way attention in Encoder Decoder damages zero shot ability;
- When generating oken, Encoder-Decoder structure can only attentin the Encoder high layer. When generating Token, decoder-only structure pays Attention layer by layer and the information is more granular:
- Encoder-Decoder trains “fill in the blank in the middle” and generates the last word Next Token. There is inconsistency, and the training and generation methods of decoder-ONLV structure are consistent.
Second, the application value of large language patterns in the field of customer contact
What are the application values of large language patterns in the field of customer contact?
Improve automatic response capability
The working logic of large models begins with continuous training on large amounts of data. Fractional and continuous training can make the large model have more accurate semantic understanding ability and more powerful natural language generation ability. Based on the large model that has been trained and mature, the development of intelligent customer service system has a more solid underlying support. It can provide fast and accurate responses based on user input problems, and quickly resolving problems saves the customer service team a lot of time and resources, improving customer experience and satisfaction.
Strengthen the ability of intention recognition
Whether intelligent customer service can handle complex problems, there is a common indicator in the industry, that is, the accuracy rate of intention recognition.
Observing the current situation in the field of customer contact, most of the simple, repetitive and procedural problems are handed over to robots, and complex problems requiring emotional care are handed over to manual customer service.
The generalization of ChatGPT provides us with new ideas for dealing with complex problems. In the analysis of multimodal data such as text, speech and image, the accuracy of large model for intention recognition is further improved. With the help of the large model, intelligent customer service can effectively combine the context of the user’s historical dialogue and current communication content, and more accurately identify the user’s needs and intentions. At the same time, with the deep learning ability of the large model, more intelligent Q&A recommendation is carried out, and then effectively empower the business consultation, investment guidance, service response and other links of the enterprise.
Optimize human-computer interaction experience
Traditional robots are not flexible enough in dealing with complex scenes. Once a user asks a question that is not in the knowledge base, or outside the preset process, the robot cannot respond well. However, based on the strong knowledge base of large models, the above situation is alleviated a lot. Taking ChatGPT as an example, the deep application of the large model has also created a new paradigm for customer experience, with its rich parameters and powerful content generation capabilities, which can support intelligent customer service to achieve more personalized Q&A responses, rather than the past, the same mechanical Q&A.
Enrich practical application scenarios
Based on the underlying capabilities provided by the large model, the penetration and application scenarios of intelligent customer service will also be further extended in the future.
At present, the application of ChatGPT has relatively certain scenarios that can be implemented, such as playing the role of human customer service and customer communication expertise, providing professional Q&A knowledge suggestions, marking the communication record, actively analyzing the work behavior of agents, launching product introductions, chatting and more “humanized” guidance and information.
In addition, it also has great application prospects in emotional care.
More articles about big model, such as big model data, big model success stories, big model solutions, etc., can be read in the Big Model column of the Digital Transformation network; If you need a large model related information package, you can click the blue character to fill in the questionnaire and contact us; More exciting content can enter the readership group and communicate with millions of readers.
💬 评论 (0)