Pooling Function

The Pooling Function type creates a pooling function to embed documents into a word space by using aggregators to combine the word embeddings. Conceptually, provided we have a representation of words, the function provides a way to represent a document by the combination of all the words it contains.

Configuration

A new function of type pooling named <id> can be created as follows:

mldb.put("/v1/functions/"+<id>, {
    "type": "pooling",
    "params": {
        "aggregators": <ARRAY [ string ]>,
        "embeddingDataset": <SqlFromExpression>
    }
})

with the following key-value definitions for params:

The aggregators specifies the type of pooling that will be performed. To do average pooling, use the avg aggregator, etc.

Input and Output Values

Functions of this type have a single input value named words which is a row, and a single output value named embedding.

Example

Suppose we have a word embedding represented by the following dataset called word_embedding:

Also suppose we have the following bag of words in the bag_of_words dataset. This can be obtained by applying the tokenize function to any text field:

Let's configure a pooling function:

mldb.put("/v1/functions/pooler", {
    "type": "pooling",
    "params": {
        "aggregators": ["avg"],
        "embeddingDataset": "word_embedding"
    }
})

We can now do the following call:

SELECT pooler({words: {*}})[embedding] as embed from bag_of_words

See also

MLDB Embedding importers

• The import.word2vec procedure type imports a Word2Vec embedding.
• The import.sentiwordnet procedure type imports a SentiWordNet embedding.