Doc2Vec向量计算方法、size参数含义及单句向量维度疑问解析

Great question! Let’s break this down step by step so it makes sense.

What does the size parameter mean? #

The size parameter (in older Gensim versions; rebranded as vector_size in newer releases) directly defines the dimensionality of the document vector the model outputs. When you set it to 100, you’ll get a 100-dimensional numerical vector; set it to 200, and you’ll get a 200-dimensional one. Think of this as the "length" of the number list that represents your sentence/document. It’s a hyperparameter you choose: higher dimensions can capture more nuanced semantic details, but they also mean heavier computation and a higher risk of overfitting to your training data.

How Doc2Vec computes document vectors #

Doc2Vec relies on two core training architectures, both learning document vectors through predictive tasks. Here’s a plain-language breakdown of each:

1. PV-DM (Distributed Memory) #

Think of this as "using the document’s context to predict words." Every document gets its own unique vector during training. When the model trains, it combines this document vector with the vectors of nearby words in the text, then tries to predict a target word within a set window. Over thousands of iterations, the model adjusts the document vector (and word vectors) to make these predictions as accurate as possible. By the end of training, the document vector has learned to capture the overall meaning of the document—because it’s been tuned to work seamlessly with the words in that document.

2. PV-DBOW (Distributed Bag of Words) #

This is the simpler, faster architecture: "using the document to predict random words from it." Instead of combining the document vector with nearby words, the model uses only the document vector to predict randomly sampled words from the document. Again, over training, the document vector is refined to get better at these predictions. This approach works particularly well with larger datasets.

For your 5-word sentence example: no matter how short or long the input is, the size parameter dictates how many dimensions the final document vector will have. The model doesn’t just average word vectors (a simpler baseline method)—it learns a unique vector specific to the document, optimized for the predictive task it’s trained on. These vectors aren’t arbitrary numbers; they’re the result of the model identifying patterns in your training data to represent each document’s unique semantic identity.

内容的提问来源于stack exchange，提问作者Yash Ghorpade