Word2Vec-Platform For AI(PAI)-阿里云帮助中心

The Word2Vec component uses a neural network to map words to K-dimensional vectors. These vectors capture semantic relationships, allowing you to perform mathematical operations that correspond to linguistic meanings. The component accepts a word column or text corpus as input and generates a word vector table and a vocabulary.

Usage notes

Connect the Word2Vec component to the output of the word frequency statistics component.

Note

The output of the word frequency statistics component is a set of triples that contain words and their counts. The Word2Vec component takes this output, converts the data into a list of single words, and then processes the entire list as a single document.

Configure the component

You can configure the Word2Vec component in one of the following ways.

Method 1: Configure in the GUI

Configure the component parameters on the pipeline page in Machine Learning Designer.

Tab	Parameter	Description
Field Settings	Select word column	The word column to use for training. For best performance, we recommend that the number of words does not exceed 10 million.
Parameter Settings	Word feature dimension	The number of dimensions for the word features. Valid values: 0 to 1,000. Default value: 100.
	Language model	The language model to use for training. Valid values: skip-gram model and CBOW model. Default value: skip-gram model.
	Word window size	The size of the word window. The value must be a positive integer. Default value: 5.
	Use random window	Specifies whether to use a random window. Enabled by default.
	Minimum word frequency	The value must be a positive integer. Default value: 5.
	Use hierarchical softmax	Specifies whether to use hierarchical softmax. Enabled by default.
	Negative sampling	The number of negative samples to use. A value of 0 disables this feature. Default value: 0.
	Downsampling threshold	The threshold for downsampling frequent words. A value of 0 disables this feature. Default value: 0.
	Initial learning rate	The value must be greater than 0. Default value: 0.025.
	Number of iterations	The value must be 1 or greater. Default value: 1.
Tuning	Number of cores	By default, the system allocates resources automatically.
Tuning	Memory size per core	By default, the system allocates resources automatically.

Method 2: Use a PAI command

Configure the component by using a PAI command. You can run a PAI command with the SQL Script component. For more information, see SQL Script.

pai -name Word2Vec
    -project algo_public
    -DinputTableName=w2v_input
    -DwordColName=word
    -DoutputTableName=w2v_output;

Parameter	Required	Description	Default
inputTableName	Yes	The name of the input vocabulary table.	None
inputTablePartitions	No	The input table partitions for tokenization. Format: `partition_name=value`. For multi-level partitions, use `name1=value1/name2=value2`. Separate multiple partitions with commas (,).	None
wordColName	Yes	The name of the word column. Each row in this column must contain a single word. The </s> string represents a newline character.	None
inVocabularyTableName	No	The output table from a wordcount operation on the input vocabulary.	The system performs a wordcount operation on the input table.
inVocabularyPartitions	No	Partitions of the output table from the word count operation.	All partitions of the table specified by inVocabularyTableName.
layerSize	No	The number of dimensions for the word features. Valid values: 0 to 1,000.	100
cbow	No	Specifies the training model. Set to 0 for the skip-gram model or 1 for the CBOW model.	0
window	No	The size of the word window. Must be a positive integer.	5
minCount	No	The minimum word frequency for truncation. Must be a positive integer.	5
hs	No	Specifies whether to use hierarchical softmax. Set to 1 to enable or 0 to disable.	1
negative	No	The number of negative samples. Must be a positive integer. A value of 0 disables this feature.	0
sample	No	The threshold for downsampling. Valid values are from 1e-3 to 1e-5. A value of 0 disables this feature.	0
alpha	No	The initial learning rate. Must be greater than 0.	0.025
iterTrain	No	The number of training iterations. Must be 1 or greater.	1
randomWindow	No	Specifies the word window type. A value of 0 indicates a fixed window size set by the window parameter. A value of 1 indicates a random window size between 1 and 5.	1
outVocabularyTableName	No	The name of the output table for the vocabulary.	None
outputTableName	Yes	The name of the output table for the word vectors.	None
lifecycle	No	The lifecycle of the output table in days. Must be a positive integer.	None
coreNum	No	The number of cores. This parameter takes effect only when memSizePerCore is also set. Must be a positive integer.	Automatically allocated
memSizePerCore	No	The memory size per core in MB. This parameter takes effect only when coreNum is also set. Must be a positive integer.	Automatically allocated

Examples

Input vocabulary:

Note
The Word2Vec component accepts input only from the word frequency statistics component.

The input vocabulary is a data preview of the sina_news_segment table (only the first 100 rows are displayed). It includes columns such as doc_id, category, title, and content. The content column is the input for the word frequency statistics.
Output result:

The first column contains the words, and the subsequent columns contain the corresponding embedding results. The layerSize parameter determines the number of dimensions.

FAQ

If the error message "Vocab size is zero! vocab_size: 0" appears, it indicates that the vocabulary is empty. To resolve this issue, decrease the value of the minCount parameter.