GEO用户意图智能分析系统

import torch
from transformers import BertTokenizer, BertForSequenceClassification

class IntentClassifier:
    def __init__(self, model_path):
        self.tokenizer = BertTokenizer.from_pretrained(model_path)
        self.model = BertForSequenceClassification.from_pretrained(model_path)
        self.model.eval()

    def predict(self, text, company_embedding):
        inputs = self.tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
        with torch.no_grad():
            outputs = self.model(**inputs)
            logits = outputs.logits
            # 融合企业特征向量（简化）
            probs = torch.softmax(logits, dim=-1)
        return probs.numpy()  # [认知, 质疑, 决策]

import math
from sklearn.metrics.pairwise import cosine_similarity

def calculate_distinctiveness(term, brand_docs, all_docs):
    p_brand = brand_docs.get_term_freq(term) / brand_docs.total_terms
    p_all = all_docs.get_term_freq(term) / all_docs.total_terms
    return max(0, math.log(p_brand / p_all))

def calculate_representativeness(term, topic_vectors):
    # 使用LDA主题向量，计算最大余弦相似度
    term_vec = get_term_vector(term)
    return max(cosine_similarity([term_vec], topic_vectors)[0])

def calculate_naturalness(term):
    # 使用GPT-2计算困惑度
    perplexity = gpt2_model.perplexity(term)
    return 1 / (1 + perplexity)

def calculate_search_value(term):
    search_volume = keyword_tool.get_monthly_search_volume(term)
    commercial_weight = get_commercial_weight(term)  # 价格类词权重更高
    return math.log(search_volume + 1) * commercial_weight

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Embedding

def build_lstm_model(vocab_size, embedding_dim=64, lstm_units=128):
    model = Sequential([
        Embedding(vocab_size, embedding_dim, input_length=12),
        LSTM(lstm_units, return_sequences=True),
        LSTM(lstm_units),
        Dense(3)  # 预测未来3个月
    ])
    model.compile(optimizer='adam', loss='mse')
    return model