跳到主要内容

使用 Sentence Transformers 和 MLflow 进行高级释义挖掘

下载此笔记本

通过 Sentence Transformers 和 MLflow 增强，踏上高级释义挖掘的丰富之旅。

学习目标​

  • 应用 sentence-transformers 进行高级释义挖掘。
  • 在 MLflow 中开发自定义 PythonModel 以适应此任务。
  • 在 MLflow 生态系统中有效管理和跟踪模型。
  • 使用 MLflow 的部署功能部署释义挖掘模型。

探索释义挖掘​

探索识别语义相似但文本不同的句子，这是文档摘要和聊天机器人开发等各种 NLP 应用的关键方面。

Sentence Transformers 在释义挖掘中的作用​

了解专门用于生成丰富句子嵌入的 Sentence Transformers 如何用于捕获深层语义含义并比较文本内容。

MLflow：简化模型管理和部署​

深入了解 MLflow 如何简化 NLP 模型的管理和部署流程，重点关注高效跟踪和可自定义的模型实现。

加入我们，深入了解释义挖掘，并掌握使用 MLflow 管理和部署 NLP 模型的艺术。

python
import warnings

# Disable a few less-than-useful UserWarnings from setuptools and pydantic
warnings.filterwarnings("ignore", category=UserWarning)

释义挖掘模型简介​

启动释义挖掘模型，集成 Sentence Transformers 和 MLflow 以进行高级 NLP 任务。

模型结构概述​

  • 加载模型和语料库 load_context 方法：加载 Sentence Transformer 模型和用于释义识别的文本语料库至关重要。
  • 释义挖掘逻辑 predict 方法：集成自定义逻辑进行输入验证和释义挖掘，提供可自定义的参数。
  • 排序和过滤匹配项 _sort_and_filter_matches 辅助方法：通过基于相似度分数进行排序和过滤，确保相关且唯一的释义识别。

主要特性​

  • 高级 NLP 技术：利用 Sentence Transformers 进行语义文本理解。
  • 自定义逻辑集成：展示模型行为自定义的灵活性。
  • 用户自定义选项：允许最终用户为各种用例调整匹配标准。
  • 处理效率：预编码语料库以实现高效的释义挖掘操作。
  • 稳健的错误处理：包含用于可靠模型性能的验证。

实际应用​

该模型为各种应用中的释义检测提供了强大的工具，例证了在 MLflow 框架内有效使用自定义模型。

python
import warnings

import pandas as pd
from sentence_transformers import SentenceTransformer, util

import mlflow
from mlflow.models.signature import infer_signature
from mlflow.pyfunc import PythonModel


class ParaphraseMiningModel(PythonModel):
  def load_context(self, context):
      """Load the model context for inference, including the customer feedback corpus."""
      try:
          # Load the pre-trained sentence transformer model
          self.model = SentenceTransformer.load(context.artifacts["model_path"])

          # Load the customer feedback corpus from the specified file
          corpus_file = context.artifacts["corpus_file"]
          with open(corpus_file) as file:
              self.corpus = file.read().splitlines()

      except Exception as e:
          raise ValueError(f"Error loading model and corpus: {e}")

  def _sort_and_filter_matches(
      self,
      query: str,
      paraphrase_pairs: list[tuple[float, int, int]],
      similarity_threshold: float,
  ):
      """Sort and filter the matches by similarity score."""

      # Convert to list of tuples and sort by score
      sorted_matches = sorted(paraphrase_pairs, key=lambda x: x[1], reverse=True)

      # Filter and collect paraphrases for the query, avoiding duplicates
      query_paraphrases = {}
      for score, i, j in sorted_matches:
          if score < similarity_threshold:
              continue

          paraphrase = self.corpus[j] if self.corpus[i] == query else self.corpus[i]
          if paraphrase == query:
              continue

          if paraphrase not in query_paraphrases or score > query_paraphrases[paraphrase]:
              query_paraphrases[paraphrase] = score

      return sorted(query_paraphrases.items(), key=lambda x: x[1], reverse=True)

  def predict(self, context, model_input, params=None):
      """Predict method to perform paraphrase mining over the corpus."""

      # Validate and extract the query input
      if isinstance(model_input, pd.DataFrame):
          if model_input.shape[1] != 1:
              raise ValueError("DataFrame input must have exactly one column.")
          query = model_input.iloc[0, 0]
      elif isinstance(model_input, dict):
          query = model_input.get("query")
          if query is None:
              raise ValueError("The input dictionary must have a key named 'query'.")
      else:
          raise TypeError(
              f"Unexpected type for model_input: {type(model_input)}. Must be either a Dict or a DataFrame."
          )

      # Determine the minimum similarity threshold
      similarity_threshold = params.get("similarity_threshold", 0.5) if params else 0.5

      # Add the query to the corpus for paraphrase mining
      extended_corpus = self.corpus + [query]

      # Perform paraphrase mining
      paraphrase_pairs = util.paraphrase_mining(
          self.model, extended_corpus, show_progress_bar=False
      )

      # Convert to list of tuples and sort by score
      sorted_paraphrases = self._sort_and_filter_matches(
          query, paraphrase_pairs, similarity_threshold
      )

      # Warning if no paraphrases found
      if not sorted_paraphrases:
          warnings.warn("No paraphrases found above the similarity threshold.", UserWarning)

      return {sentence[0]: str(sentence[1]) for sentence in sorted_paraphrases}

准备用于释义挖掘的语料库​

通过创建和准备多样化的语料库来为释义挖掘奠定基础。

语料库创建​

  • 定义一个 corpus，其中包含来自各种主题的句子，包括太空探索、人工智能、园艺等。这种多样性使模型能够跨广泛的主题识别释义。

将语料库写入文件​

  • 语料库将保存到名为 feedback.txt 的文件中，这反映了大型数据处理中的常见做法。
  • 此步骤还为在释义挖掘模型中高效处理语料库做准备。

语料库的重要性​

语料库是模型查找语义相似句子的关键数据集。其多样性确保了模型在各种用例中的适应性和有效性。

python
corpus = [
  "Exploring ancient cities in Europe offers a glimpse into history.",
  "Modern AI technologies are revolutionizing industries.",
  "Healthy eating contributes significantly to overall well-being.",
  "Advancements in renewable energy are combating climate change.",
  "Learning a new language opens doors to different cultures.",
  "Gardening is a relaxing hobby that connects you with nature.",
  "Blockchain technology could redefine digital transactions.",
  "Homemade Italian pasta is a delight to cook and eat.",
  "Practicing yoga daily improves both physical and mental health.",
  "The art of photography captures moments in time.",
  "Baking bread at home has become a popular quarantine activity.",
  "Virtual reality is creating new experiences in gaming.",
  "Sustainable travel is becoming a priority for eco-conscious tourists.",
  "Reading books is a great way to unwind and learn.",
  "Jazz music provides a rich tapestry of sound and rhythm.",
  "Marathon training requires discipline and perseverance.",
  "Studying the stars helps us understand our universe.",
  "The rise of electric cars is an important environmental development.",
  "Documentary films offer deep insights into real-world issues.",
  "Crafting DIY projects can be both fun and rewarding.",
  "The history of ancient civilizations is fascinating to explore.",
  "Exploring the depths of the ocean reveals a world of marine wonders.",
  "Learning to play a musical instrument can be a rewarding challenge.",
  "Artificial intelligence is shaping the future of personalized medicine.",
  "Cycling is not only a great workout but also eco-friendly transportation.",
  "Home automation with IoT devices is enhancing living experiences.",
  "Understanding quantum computing requires a grasp of complex physics.",
  "A well-brewed cup of coffee is the perfect start to the day.",
  "Urban farming is gaining popularity as a sustainable food source.",
  "Meditation and mindfulness can lead to a more balanced life.",
  "The popularity of podcasts has revolutionized audio storytelling.",
  "Space exploration continues to push the boundaries of human knowledge.",
  "Wildlife conservation is essential for maintaining biodiversity.",
  "The fusion of technology and fashion is creating new trends.",
  "E-learning platforms have transformed the educational landscape.",
  "Dark chocolate has surprising health benefits when enjoyed in moderation.",
  "Robotics in manufacturing is leading to more efficient production.",
  "Creating a personal budget is key to financial well-being.",
  "Hiking in nature is a great way to connect with the outdoors.",
  "3D printing is innovating the way we create and manufacture objects.",
  "Sommeliers can identify a wine's characteristics with just a taste.",
  "Mind-bending puzzles and riddles are great for cognitive exercise.",
  "Social media has a profound impact on communication and culture.",
  "Urban sketching captures the essence of city life on paper.",
  "The ethics of AI is a growing field in tech philosophy.",
  "Homemade skincare remedies are becoming more popular.",
  "Virtual travel experiences can provide a sense of adventure at home.",
  "Ancient mythology still influences modern storytelling and literature.",
  "Building model kits is a hobby that requires patience and precision.",
  "The study of languages opens windows into different worldviews.",
  "Professional esports has become a major global phenomenon.",
  "The mysteries of the universe are unveiled through space missions.",
  "Astronauts' experiences in space stations offer unique insights into life beyond Earth.",
  "Telescopic observations bring distant galaxies within our view.",
  "The study of celestial bodies helps us understand the cosmos.",
  "Space travel advancements could lead to interplanetary exploration.",
  "Observing celestial events provides valuable data for astronomers.",
  "The development of powerful rockets is key to deep space exploration.",
  "Mars rover missions are crucial in searching for extraterrestrial life.",
  "Satellites play a vital role in our understanding of Earth's atmosphere.",
  "Astrophysics is central to unraveling the secrets of space.",
  "Zero gravity environments in space pose unique challenges and opportunities.",
  "Space tourism might soon become a reality for many.",
  "Lunar missions have contributed significantly to our knowledge of the moon.",
  "The International Space Station is a hub for groundbreaking space research.",
  "Studying comets and asteroids reveals information about the early solar system.",
  "Advancements in space technology have implications for many scientific fields.",
  "The possibility of life on other planets continues to intrigue scientists.",
  "Black holes are among the most mysterious phenomena in space.",
  "The history of space exploration is filled with remarkable achievements.",
  "Future space missions could unlock the mysteries of dark matter.",
]

# Write out the corpus to a file
corpus_file = "/tmp/feedback.txt"
with open(corpus_file, "w") as file:
  for sentence in corpus:
      file.write(sentence + "
")

设置释义挖掘模型​

为 Sentence Transformer 模型与 MLflow 集成做好准备，以利用其释义挖掘能力。

加载 Sentence Transformer 模型​

  • 初始化 all-MiniLM-L6-v2 Sentence Transformer 模型，该模型非常适合生成适用于释义挖掘的句子嵌入。

准备输入示例​

  • 创建 DataFrame 作为输入示例，以说明模型将处理的查询类型，有助于定义模型的输入结构。

保存模型​

  • 将模型保存到 /tmp/paraphrase_search_model 以便移植，并在使用 MLflow 进行部署时方便加载。

定义构件和语料库路径​

  • 将已保存模型和语料库的路径指定为 MLflow 中的构件，这对于模型日志记录和复现至关重要。

为签名生成测试输出​

  • 生成示例输出，说明模型在释义挖掘方面的预期输出格式。

创建模型签名​

  • 使用 MLflow 的 infer_signature 定义模型的输入和输出模式，并添加 similarity_threshold 参数以实现推理灵活性。
python
# Load a pre-trained sentence transformer model
model = SentenceTransformer("all-MiniLM-L6-v2")

# Create an input example DataFrame
input_example = pd.DataFrame({"query": ["This product works well. I'm satisfied."]})

# Save the model in the /tmp directory
model_directory = "/tmp/paraphrase_search_model"
model.save(model_directory)

# Define the path for the corpus file
corpus_file = "/tmp/feedback.txt"

# Define the artifacts (paths to the model and corpus file)
artifacts = {"model_path": model_directory, "corpus_file": corpus_file}

# Generate test output for signature
# Sample output for paraphrase mining could be a list of tuples (paraphrase, score)
test_output = [{"This product is satisfactory and functions as expected.": "0.8"}]

# Define the signature associated with the model
# The signature includes the structure of the input and the expected output, as well as any parameters that
# we would like to expose for overriding at inference time (including their default values if they are not overridden).
signature = infer_signature(
  model_input=input_example, model_output=test_output, params={"similarity_threshold": 0.5}
)

# Visualize the signature, showing our overridden inference parameter and its default.
signature
inputs: 
['query': string]
outputs: 
['This product is satisfactory and functions as expected.': string]
params: 
['similarity_threshold': double (default: 0.5)]

创建实验​

我们创建一个新的 MLflow 实验，以便我们要将模型记录到的运行不会记录到默认实验，而是具有其自己的上下文相关条目。

python
# If you are running this tutorial in local mode, leave the next line commented out.
# Otherwise, uncomment the following line and set your tracking uri to your local or remote tracking server.

# mlflow.set_tracking_uri("http://127.0.0.1:8080")

mlflow.set_experiment("Paraphrase Mining")
<Experiment: artifact_location='file:///Users/benjamin.wilson/repos/mlflow-fork/mlflow/docs/source/llms/sentence-transformers/tutorials/paraphrase-mining/mlruns/380691166097743403', creation_time=1701282619556, experiment_id='380691166097743403', last_update_time=1701282619556, lifecycle_stage='active', name='Paraphrase Mining', tags={}>

使用 MLflow 记录释义挖掘模型​

使用 MLflow 记录自定义释义挖掘模型，这是模型管理和部署的关键步骤。

启动 MLflow 运行​

  • 启动 MLflow 运行，在 MLflow 框架内创建模型日志记录和跟踪的全面记录。

在 MLflow 中记录模型​

  • 使用 MLflow 的 Python 模型日志记录功能将自定义模型集成到 MLflow 生态系统中。
  • 为模型提供唯一名称，以便在 MLflow 中轻松识别。
  • 记录已实例化的释义挖掘模型，以及输入示例、模型签名、构件和 Python 依赖项。

模型日志记录的成果和优势​

  • 将模型注册到 MLflow 中，以便进行简化的管理和部署，提高其可访问性和可跟踪性。
  • 确保跨部署环境的模型可复现性和版本控制。
python
with mlflow.start_run() as run:
  model_info = mlflow.pyfunc.log_model(
      name="paraphrase_model",
      python_model=ParaphraseMiningModel(),
      input_example=input_example,
      signature=signature,
      artifacts=artifacts,
      pip_requirements=["sentence_transformers"],
  )
Downloading artifacts:   0%|          | 0/11 [00:00<?, ?it/s]
2023/11/30 15:41:39 INFO mlflow.store.artifact.artifact_repo: The progress bar can be disabled by setting the environment variable MLFLOW_ENABLE_ARTIFACTS_PROGRESS_BAR to false
Downloading artifacts:   0%|          | 0/1 [00:00<?, ?it/s]

模型加载和释义挖掘预测​

通过使用 MLflow 加载模型并执行预测，说明释义挖掘模型的实际应用。

加载模型进行推理​

  • 利用 MLflow 的 load_model 函数检索模型并为推理做好准备。
  • 使用模型在 MLflow 注册表中的唯一 URI 定位和加载模型。

执行释义挖掘预测​

  • 使用模型的 predict 方法进行预测，应用模型类中嵌入的释义挖掘逻辑。
  • 传递一个具有设置好的 similarity_threshold 的代表性查询，以在语料库中查找匹配的释义。

解释模型输出​

  • 查看与查询语义相似的句子列表，突出模型在释义识别方面的能力。
  • 分析相似度分数，以了解查询和语料库句子之间的语义相关程度。

结论​

此演示验证了释义挖掘模型在实际场景中的有效性，并强调了其在内容推荐、信息检索和对话式人工智能中的效用。

python
# Load our model by supplying the uri that was used to save the model artifacts
loaded_dynamic = mlflow.pyfunc.load_model(model_info.model_uri)

# Perform a quick validation that our loaded model is performing adequately
loaded_dynamic.predict(
  {"query": "Space exploration is fascinating."}, params={"similarity_threshold": 0.65}
)
{'Studying the stars helps us understand our universe.': '0.8207424879074097',
'The history of space exploration is filled with remarkable achievements.': '0.7770636677742004',
'Exploring ancient cities in Europe offers a glimpse into history.': '0.7461957335472107',
'Space travel advancements could lead to interplanetary exploration.': '0.7090306282043457',
'Space exploration continues to push the boundaries of human knowledge.': '0.6893945932388306',
'The mysteries of the universe are unveiled through space missions.': '0.6830739974975586',
'The study of celestial bodies helps us understand the cosmos.': '0.671358048915863'}

结论：见解和潜在增强​

在完成本教程时，让我们回顾一下使用 Sentence Transformers 和 MLflow 实现释义挖掘模型的旅程。我们已成功构建并部署了一个能够识别语义相似句子的模型，展示了 MLflow 的 PythonModel 实现的灵活性和强大功能。

主要收获​

  • 我们学习了如何将高级 NLP 技术（特别是释义挖掘）与 MLflow 相集成。这种集成不仅增强了模型管理，还简化了部署和可扩展性。
  • MLflow 中 PythonModel 实现的灵活性是一个中心主题。我们亲眼看到它如何允许将自定义逻辑纳入模型的 predict 函数，以满足释义挖掘等特定 NLP 任务的需求。
  • 通过我们的自定义模型，我们探索了句子嵌入、语义相似性和语言理解细微差别的动态。这种理解在从内容推荐到对话式人工智能的广泛应用中至关重要。

增强释义挖掘模型的想法​

虽然我们的模型是一个坚实的起点，但可以在 predict 函数中进行几项增强，使其更强大、功能更丰富

  1. 上下文过滤器：引入基于上下文线索或特定关键字的过滤器，以进一步细化搜索结果。此功能允许用户将释义缩小到最相关的那些，以适应其特定上下文或主题。

  2. 情感分析集成：集成情感分析以按情感基调对释义进行分组。这在客户反馈分析等应用中尤其有用，在这些应用中，理解情感与内容同等重要。

  3. 多语言支持：将模型扩展到支持多种语言的释义挖掘。此增强将极大地扩展模型在全球或多语言环境中的适用性。

与向量数据库的可扩展性​

  • 超越静态文本文件作为语料库，更具可扩展性和实际应用的方法将涉及将模型连接到外部向量数据库或内存存储。
  • 预计算的嵌入可以存储和更新在这些数据库中，从而适应实时内容生成，而无需重新部署模型。这种方法将极大地提高模型在实际应用中的可扩展性和响应能力。

最终想法​

构建和部署释义挖掘模型的旅程既富有启发性又实用。我们已经看到 MLflow 的 PythonModel 如何提供一个灵活的画布来构建自定义 NLP 解决方案，以及如何利用 Sentence Transformers 深入挖掘语言的语义。

本教程只是一个开始。在释义挖掘和整个 NLP 领域，还有巨大的探索和创新潜力。我们鼓励您在此基础上进行构建，尝试增强功能，并继续突破 MLflow 和高级 NLP 技术所能实现的界限。