embed_text_with_model
Use language models to calulate an embedding for each text in provided column.
An embedding vector is a numerical representation of a text, such that different numerical components of the vector
capture different dimensions of the text’s meaning. Embeddings can be used, for example, to calculate the semantic similarity
between pairs of texts. See link_embeddings,
for example, to create a network of texts connected by similarity.
In this step, embeddings of texts are calculated using pre-trained neural language models, especially those using the popular transformer architecture (e.g. Bert-based models).
Things to keep in mind
- Unlike
embed_text, which uses a different, appropriate spaCy model for each language in the text column, this step will always use a single model only to calculate embeddings. This means the model should be multilingual if you have mixed languages, and that otherwise you need to choose the correct model for your (single) language. - Each model will be downloaded on the fly before processing the text. This adds a little lag to its execution time (the bigger the model the longer the download), though for a sufficient number of texts the time spent downloading should not be significant. Note also, however, that the download, and therefore this step, may fail if the servers of its publisher are not responsive.
- Since this step potentially supports tens if not hundreds of different models, we cannot provide support or advice on specific models.
Usage
The following example shows how the step can be used in a recipe.
To calculate embeddings using a multilingual sentence-bert model (from sentence-transformers):
To calculate embeddings using a multilingual sentence-bert model (from sentence-transformers):
General syntax for using the step in a recipe. Shows the inputs and outputs the step is expected to receive and will produce respectively. For futher details see sections below.
Inputs & Outputs
The following are the inputs expected by the step and the outputs it produces. These are generally
columns (ds.first_name), datasets (ds or ds[["first_name", "last_name"]]) or models (referenced
by name e.g. "churn-clf").
Configuration
The following parameters can be used to configure the behaviour of the step by including them in
a json object as the last “input” to the step, i.e. step(..., {"param": "value", ...}) -> (output).
Embed texts using a Sentence-BERT model. Models in this collection (also known as sentence-transformers) have been trained specifically for semantic similarity, i.e. for the purpose of comparing the meaning of texts. Individual models in this collection can be found here: https://www.sbert.net/docs/pretrained_models.html. They differ in terms of the language they have been trained on; their size (the bigger the better usually, but also the slower); as well as their purpose or intended area of application (e.g. it has a specific model to embed scientific publications).
A specific Sentence-BERT model name. To find a model appropriate for your data or task, check the website of the Sentence-BERT model collection.
Whether text embedding vectors should be normalized (to lengths of 1.0). This may make similarity calculations easier. E.g. we can then use the dot product as a similarity “metric”, instead of the usual cosine angle (which not all downstream functions may support).
How many texts to push through the model at the same time. Greater values usually mean faster processing (if supported by the model), but also greater use of memory.
Values must be in the following range:
Embed texts using a Sentence-BERT model. Models in this collection (also known as sentence-transformers) have been trained specifically for semantic similarity, i.e. for the purpose of comparing the meaning of texts. Individual models in this collection can be found here: https://www.sbert.net/docs/pretrained_models.html. They differ in terms of the language they have been trained on; their size (the bigger the better usually, but also the slower); as well as their purpose or intended area of application (e.g. it has a specific model to embed scientific publications).
A specific Sentence-BERT model name. To find a model appropriate for your data or task, check the website of the Sentence-BERT model collection.
Whether text embedding vectors should be normalized (to lengths of 1.0). This may make similarity calculations easier. E.g. we can then use the dot product as a similarity “metric”, instead of the usual cosine angle (which not all downstream functions may support).
How many texts to push through the model at the same time. Greater values usually mean faster processing (if supported by the model), but also greater use of memory.
Values must be in the following range:
Embed texts using a model from the Hugging Face hub. Any pytorch or tensorflow model in HF’s hub can be used as long as its output contains a last hidden state. Note however, that using the output embedding of an arbitrary transformer is not always useful, and specifically may not be approriate for sentence similarity. Rather, these embeddings usually represent the input for downstream classification tasks instead. A sentence-bert or universal sentence encoder may be more appopriate in most cases.
A specific Hugging Face model name.
To find a model appropriate for your data or task, browse the Hugging Face model hub.
Note that the name of a model should include the name of the organization if applicable (e.g.
"cardiffnlp/" in the example below).
Whether text embedding vectors should be normalized (to lengths of 1.0). This may make similarity calculations easier. E.g. we can then use the dot product as a similarity “metric”, instead of the usual cosine angle (which not all downstream functions may support).
How many texts to push through the model at the same time. Greater values usually mean faster processing (if supported by the model), but also greater use of memory.
Values must be in the following range:
How individual “word” embeddings should be combined. The output of a transformer contains embeddings for individual words (or sentence pieces, sub-word character sequences etc.). This parameter determines how these are combined to create a single vector representing the whole text. This can be the mean of individual vectors or the (component-wise) maximum (currently pooling doesn’t take the attention mask into account).
Values must be one of the following:
meanmax