1 引言

当今世界上有太多的文本信息，几乎没有人能够拥有足够的时间去阅读所有我们想了解的东西。但令人感到欣喜的是，目前LLM在文本概括任务上展现了强大的水准，也已经有不少团队将这项功能插入了自己的软件应用中。
本章节将介绍如何使用编程的方式，调用API接口来实现“文本概括”功能。
首先，我们需要OpenAI包，加载API密钥，定义getCompletion函数。

  1. import openai
  2. import os
  3. OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
  4. openai.api_key = OPENAI_API_KEY
  6. def get_completion(prompt, model="gpt-3.5-turbo"): 
  7.     messages = [{"role": "user", "content": prompt}]
  8.     response = openai.ChatCompletion.create(
  9.         model=model,
  10.         messages=messages,
  11.         temperature=0, # 值越低则输出文本随机性越低
  12.     )
  13.     return response.choices[0].message["content"]

2 单一文本概括Prompt实验

这里我们举了个商品评论的例子。对于电商平台来说，网站上往往存在着海量的商品评论，这些评论反映了所有客户的想法。如果我们拥有一个工具去概括这些海量、冗长的评论，便能够快速地浏览更多评论，洞悉客户的偏好，从而指导平台与商家提供更优质的服务。
输入文本

  1. prod_review = """
  2. Got this panda plush toy for my daughter's birthday, \
  3. who loves it and takes it everywhere. It's soft and \ 
  4. super cute, and its face has a friendly look. It's \ 
  5. a bit small for what I paid though. I think there \ 
  6. might be other options that are bigger for the \ 
  7. same price. It arrived a day earlier than expected, \ 
  8. so I got to play with it myself before I gave it \ 
  9. to her.
  10. """

输入文本（中文翻译）

  1. prod_review_zh = """
  2. 这个熊猫公仔是我给女儿的生日礼物，她很喜欢，去哪都带着。
  3. 公仔很软，超级可爱，面部表情也很和善。但是相比于价钱来说，
  4. 它有点小，我感觉在别的地方用同样的价钱能买到更大的。
  5. 快递比预期提前了一天到货，所以在送给女儿之前，我自己玩了会。
  6. """

2.1 限制输出文本长度

我们尝试限制文本长度为最多30词。

  1. prompt = f"""
  2. Your task is to generate a short summary of a product \
  3. review from an ecommerce site. 
  5. Summarize the review below, delimited by triple 
  6. backticks, in at most 30 words. 
  8. Review: ```{prod_review}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. Soft and cute panda plush toy loved by daughter, but a bit small for the price. Arrived early.
  1. prompt = f"""
  2. 你的任务是从电子商务网站上生成一个产品评论的简短摘要。
  4. 请对三个反引号之间的评论文本进行概括，最多30个词汇。
  6. 评论: ```{prod_review_zh}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. 可爱软熊猫公仔，女儿喜欢，面部表情和善，但价钱有点小贵，快递提前一天到货。

2.2 关键角度侧重

有时，针对不同的业务，我们对文本的侧重会有所不同。例如对于商品评论文本，物流会更关心运输时效，商家更加关心价格与商品质量，平台更关心整体服务体验。
我们可以通过增加Prompt提示，来体现对于某个特定角度的侧重。
侧重于运输

  1. prompt = f"""
  2. Your task is to generate a short summary of a product \
  3. review from an ecommerce site to give feedback to the \
  4. Shipping deparmtment. 
  6. Summarize the review below, delimited by triple 
  7. backticks, in at most 30 words, and focusing on any aspects \
  8. that mention shipping and delivery of the product. 
  10. Review: ```{prod_review}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. The panda plush toy arrived a day earlier than expected, but the customer felt it was a bit small for the price paid.
  1. prompt = f"""
  2. 你的任务是从电子商务网站上生成一个产品评论的简短摘要。
  4. 请对三个反引号之间的评论文本进行概括，最多30个词汇，并且聚焦在产品运输上。
  6. 评论: ```{prod_review_zh}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. 快递提前到货，熊猫公仔软可爱，但有点小，价钱不太划算。

可以看到，输出结果以“快递提前一天到货”开头，体现了对于快递效率的侧重。
侧重于价格与质量

  1. prompt = f"""
  2. Your task is to generate a short summary of a product \
  3. review from an ecommerce site to give feedback to the \
  4. pricing deparmtment, responsible for determining the \
  5. price of the product.  
  7. Summarize the review below, delimited by triple 
  8. backticks, in at most 30 words, and focusing on any aspects \
  9. that are relevant to the price and perceived value. 
  11. Review: ```{prod_review}

“””

response = get_completion(prompt) print(response)

  1. ```
  2. The panda plush toy is soft, cute, and loved by the recipient, but the price may be too high for its size compared to other options.
  1. prompt = f"""
  2. 你的任务是从电子商务网站上生成一个产品评论的简短摘要。
  4. 请对三个反引号之间的评论文本进行概括，最多30个词汇，并且聚焦在产品价格和质量上。
  6. 评论: ```{prod_review_zh}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. 可爱软熊猫公仔，面部表情友好，但价钱有点高，尺寸较小。快递提前一天到货。

可以看到，输出结果以“质量好、价格小贵、尺寸小”开头，体现了对于产品价格与质量的侧重。

2.3 关键信息提取

在2.2节中，虽然我们通过添加关键角度侧重的Prompt，使得文本摘要更侧重于某一特定方面，但是可以发现，结果中也会保留一些其他信息，如价格与质量角度的概括中仍保留了“快递提前到货”的信息。有时这些信息是有帮助的，但如果我们只想要提取某一角度的信息，并过滤掉其他所有信息，则可以要求LLM进行“文本提取(Extract)”而非“文本概括(Summarize)”。

  1. prompt = f"""
  2. Your task is to extract relevant information from \ 
  3. a product review from an ecommerce site to give \
  4. feedback to the Shipping department. 
  6. From the review below, delimited by triple quotes \
  7. extract the information relevant to shipping and \ 
  8. delivery. Limit to 30 words. 
  10. Review: ```{prod_review}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. "The product arrived a day earlier than expected."
  1. prompt = f"""
  2. 你的任务是从电子商务网站上的产品评论中提取相关信息。
  4. 请从以下三个反引号之间的评论文本中提取产品运输相关的信息，最多30个词汇。
  6. 评论: ```{prod_review_zh}

“””

response = get_completion(prompt) print(response)

  1. ```python
  2. 快递比预期提前了一天到货。

3 多条文本概括Prompt实验

在实际的工作流中，我们往往有许许多多的评论文本，以下展示了一个基于for循环调用“文本概括”工具并依次打印的示例。当然，在实际生产中，对于上百万甚至上千万的评论文本，使用for循环也是不现实的，可能需要考虑整合评论、分布式等方法提升运算效率。

  1. review_1 = prod_review 
  3. # review for a standing lamp
  4. review_2 = """
  5. Needed a nice lamp for my bedroom, and this one \
  6. had additional storage and not too high of a price \
  7. point. Got it fast - arrived in 2 days. The string \
  8. to the lamp broke during the transit and the company \
  9. happily sent over a new one. Came within a few days \
  10. as well. It was easy to put together. Then I had a \
  11. missing part, so I contacted their support and they \
  12. very quickly got me the missing piece! Seems to me \
  13. to be a great company that cares about their customers \
  14. and products. 
  15. """
  17. # review for an electric toothbrush
  18. review_3 = """
  19. My dental hygienist recommended an electric toothbrush, \
  20. which is why I got this. The battery life seems to be \
  21. pretty impressive so far. After initial charging and \
  22. leaving the charger plugged in for the first week to \
  23. condition the battery, I've unplugged the charger and \
  24. been using it for twice daily brushing for the last \
  25. 3 weeks all on the same charge. But the toothbrush head \
  26. is too small. I’ve seen baby toothbrushes bigger than \
  27. this one. I wish the head was bigger with different \
  28. length bristles to get between teeth better because \
  29. this one doesn’t.  Overall if you can get this one \
  30. around the $50 mark, it's a good deal. The manufactuer's \
  31. replacements heads are pretty expensive, but you can \
  32. get generic ones that're more reasonably priced. This \
  33. toothbrush makes me feel like I've been to the dentist \
  34. every day. My teeth feel sparkly clean! 
  35. """
  37. # review for a blender
  38. review_4 = """
  39. So, they still had the 17 piece system on seasonal \
  40. sale for around $49 in the month of November, about \
  41. half off, but for some reason (call it price gouging) \
  42. around the second week of December the prices all went \
  43. up to about anywhere from between $70-$89 for the same \
  44. system. And the 11 piece system went up around $10 or \
  45. so in price also from the earlier sale price of $29. \
  46. So it looks okay, but if you look at the base, the part \
  47. where the blade locks into place doesn’t look as good \
  48. as in previous editions from a few years ago, but I \
  49. plan to be very gentle with it (example, I crush \
  50. very hard items like beans, ice, rice, etc. in the \ 
  51. blender first then pulverize them in the serving size \
  52. I want in the blender then switch to the whipping \
  53. blade for a finer flour, and use the cross cutting blade \
  54. first when making smoothies, then use the flat blade \
  55. if I need them finer/less pulpy). Special tip when making \
  56. smoothies, finely cut and freeze the fruits and \
  57. vegetables (if using spinach-lightly stew soften the \ 
  58. spinach then freeze until ready for use-and if making \
  59. sorbet, use a small to medium sized food processor) \ 
  60. that you plan to use that way you can avoid adding so \
  61. much ice if at all-when making your smoothie. \
  62. After about a year, the motor was making a funny noise. \
  63. I called customer service but the warranty expired \
  64. already, so I had to buy another one. FYI: The overall \
  65. quality has gone done in these types of products, so \
  66. they are kind of counting on brand recognition and \
  67. consumer loyalty to maintain sales. Got it in about \
  68. two days.
  69. """
  71. reviews = [review_1, review_2, review_3, review_4]
  1. for i in range(len(reviews)):
  2.     prompt = f"""
  3.     Your task is to generate a short summary of a product \ 
  4.     review from an ecommerce site. 
  6.     Summarize the review below, delimited by triple \
  7.     backticks in at most 20 words. 
  9.     Review: ```{reviews[i]}
  1. """
  2. response = get_completion(prompt)
  3. print(i, response, "\n")
  1. ```python
  2. 0 Soft and cute panda plush toy loved by daughter, but a bit small for the price. Arrived early. 
  4. 1 Affordable lamp with storage, fast shipping, and excellent customer service. Easy to assemble and missing parts were quickly replaced. 
  6. 2 Good battery life, small toothbrush head, but effective cleaning. Good deal if bought around $50. 
  8. 3 The product was on sale for $49 in November, but the price increased to $70-$89 in December. The base doesn't look as good as previous editions, but the reviewer plans to be gentle with it. A special tip for making smoothies is to freeze the fruits and vegetables beforehand. The motor made a funny noise after a year, and the warranty had expired. Overall quality has decreased.

尝试一下！