大语言模型如何为企业解锁新机遇（第二部分）

Part 2 Contributors:Annamalai ChockalingamAnkur PatelShashank VermaTiffany Yeung Table of Contents Overview ....................................................................................................................................................3Traditional NLP Tasks Performed by Large Language Models...................................................................5Content Generation..............................................................................................................................6Summarization......................................................................................................................................7Translation ............................................................................................................................................9Classification .......................................................................................................................................11Chatbot Support .................................................................................................................................12Case Study: Korea Telecom X NeMo Megatron ......................................................................................15 Overview Most enterprises need to perform many language-related tasks everyday. They need to organize documents based on common themes(text classification), write emails(content generation), remove harmful posts from their online communities(toxicity classification),and much more. Applications powered by large language models can help enterprises automate theseand many other tasks, helping them to streamline their operations, decrease expenses, and increaseproductivity. Alternatively, enterprises can use LLM-powered apps to provide innovative and improved services toclients or strengthen customer relationships. For example, enterprises could provide customersupport via AI companions or use sentiment analysis apps to extract valuable customer insights. This section will examine both applications of LLMs–in internalandexternal operations by reviewingthe six LLM-powered text apps listed in Table 1. It’s worth noting that large language models can be applied to other content categories besides text.They’re currently widely used for speech-, image-, and video-related tasks, such as image generationor video classification. Some of the applications presented in this document may be capable ofperforming such tasks as well. However, the following sections primarily aim to familiarize enterprises withtext-relatedLLMsolutions. They’ll explain what makes LLMs superior to traditional solutions, such asrule-basedsystems or manual approaches, as well as show how enterprises can implement them depending ontheir industry. This should clarify the immense value of LLMs for enterprises, as well as inspireorganizations to develop their own LLM variants. Traditional NLP Tasks Performed by LargeLanguage Models Traditional NLP tasks can be thought of as the foundational and morebasic iterations of contemporary use cases.Chatbot support, for example, can beconsidered a sophisticated combination of two traditional NLP tasks, text generation and question-answering. These and some other basic NLP tasks are listed in Table 2. This brief, albeit incomplete overview of traditional NLP tasks may help with understanding how tofine-tune proprietary models or select well-suited base models, a topic further explored inPart 3:How Enterprises Can Build Large Language Models. It should also be helpful in understanding themechanics behind more advanced use cases analyzed in the remaining sections. Content Generation Content generation refers to producing any type of content, from social media posts to annualreports. Most enterprises need to create significant amounts of various types of content for a widerange of purposes and for bothinternalandexternaluse. Table 3 list common content types created for internal and external use within enterprises. At the same time, enterprises face multiple challenges that can hinder their content-generationefforts and make it harder to achieve tangible results. Main content-related issues include: >Increasing content production demands.Today’s digital landscape requires businesses toconsistently produce vast amounts of content.This is underscored by the fact that 4.4 million blogpostswere publisheddaily in 2022, with 69% of businessesplanningto increase their contentmarketing budgets in 2023. Enterprises need to keep up with these standards to stay competitivebut may not be able to do so solely via manual methods, especially as other businesses startembracingAI content generation. >Increasing email composition demands.Reading and writing emailsconsumesas much as 28% ofworkweeks, or 13 hours every week, that employees could spend on more important tasks.Additionally, email composition is also one of the major stressors within enterprises, withapproximately 92% o