Python针对不同文本长度的处理方案总结与对比_Python

直接上代码+注释

有意尝试可交流

效果正在验证中。

1.短文本处理(<500tokens)

from sentence_transformers import sentencetransformer
model = sentencetransformer('all-minilm-l6-v2')  # 384维小型模型

def process_short(text):
    """直接全文本编码"""
    return model.encode(text, convert_to_tensor=true)

# 示例
short_text = "自然语言处理的基础概念"  # 长度约15 tokens
vector = process_short(short_text)

2. 中长文本处理 (500-2000 tokens)

from langchain_text_splitters import recursivecharactertextsplitter

def process_medium(text):
    """重叠分块策略"""
    splitter = recursivecharactertextsplitter(
        chunk_size=500,
        chunk_overlap=50,
        separators=["\n\n", "\n", "。", "！", "？"]
    )
    chunks = splitter.split_text(text)
    return [model.encode(chunk) for chunk in chunks]

# 示例
medium_text = "机器学习发展历史...（约1500字）"  # 约1800 tokens
chunk_vectors = process_medium(medium_text)

3. 长文本处理 (2000-20000 tokens)

import spacy

def process_long(text):
    """语义分块+摘要增强"""
    # 加载语义分割模型
    nlp = spacy.load("zh_core_web_sm") 
    doc = nlp(text)
    
    # 按段落分割
    chunks = [sent.text for sent in doc.sents]
    
    # 生成章节摘要
    summary_model = sentencetransformer('uer/sbert-base-chinese-nli')
    summaries = [summary_model.encode(chunk[:200]) for chunk in chunks]
    
    return chunks, summaries

# 示例
long_text = "人工智能技术白皮书...（约2万字）"  # 约20000 tokens
text_chunks, summary_vecs = process_long(long_text)

4. 超长文本处理 (20000-200000 tokens)

import faiss
import numpy as np

class hierarchicalindex:
    def __init__(self):
        # 两级索引结构
        self.summary_index = faiss.indexflatl2(384)
        self.chunk_index = faiss.indexivfpq(
            faiss.indexflatl2(384), 384, 100, 16, 8
        )
        self.metadata = []

    def add_document(self, text):
        # 生成段落级摘要
        chunks, summaries = process_long(text)
        
        # 构建索引
        summary_vecs = np.array(summaries).astype('float32')
        chunk_vecs = np.array([model.encode(c) for c in chunks]).astype('float32')
        
        self.summary_index.add(summary_vecs)
        self.chunk_index.add(chunk_vecs)
        self.metadata.extend(chunks)

    def search(self, query, k=5):
        # 先检索摘要层
        query_vec = model.encode(query).astype('float32')
        _, sum_indices = self.summary_index.search(np.array([query_vec]), 10)
        
        # 精搜相关块
        target_chunks = [self.chunk_index.reconstruct(i) for i in sum_indices]
        target_chunks = np.array(target_chunks).astype('float32')
        _, chunk_indices = self.chunk_index.search(target_chunks, k)
        
        return [self.metadata[i] for i in chunk_indices]

# 使用示例
hindex = hierarchicalindex()
hindex.add_document("某领域技术文档...（约15万字）")  # 约200000 tokens
results = hindex.search("深度学习在医疗影像的应用")

5. 海量文本处理 (>200000 tokens)

import dask.dataframe as dd
from dask.distributed import client

def process_extreme(file_path):
    """分布式处理方案"""
    client = client(n_workers=4)  # 启动dask集群
    
    # 分块读取
    df = dd.read_parquet(file_path, chunksize=100000)  
    
    # 并行编码
    df['vector'] = df['text'].map_partitions(
        lambda s: s.apply(model.encode),
        meta=('vector', object)
    )
    
    # 构建分布式索引
    df.to_parquet("encoded_data.parquet", engine="pyarrow")
    
# 示例（处理100万条文本）
process_extreme("massive_data.parquet")

性能优化对照表=

文本长度	处理策略	索引类型	响应时间	内存消耗
<500	直接编码	flatindex	<10ms	1mb
2000	重叠分块	ivf+pq	50-100ms	50mb
20000	语义分块+摘要索引	二级索引	200-500ms	300mb
200000	层次化索引	ivfopq+productquant	1-2s	2gb
>200000	分布式处理	分片索引	10s+	集群资源