Skip to main content
Version: 0.94

snorkelflow.rich_docs

Utilities to manipulate Rich Document structures for labeling functions and operators in Snorkel Flow.

Labeling functions that use rich documents​

When rich document information is available, the RichDoc object is available for you to write labeling functions over. We also provide a RichDocWrapper class that provides additional helpers and derived properties for a given RichDoc object. You can import these classes and use them to operate over the serialized RichDoc object.

from snorkelflow.lfs.lfs import labeling_function

@labeling_function(name="rich_doc_lf")
def rich_doc_lf(x):
    from snorkelflow.rich_docs import RichDocCols, RichDoc
    from rich_doc_wrapper import RichDocWrapper

    # Deserialize the RichDoc object
    rd = x[RichDocCols.PAGE_DOCS].apply(lambda x: RichDoc.from_page_docs(x))
    # Wrap RichDoc object and text in RichDocWrapper class
    rdw = RichDocWrapper(rd)

    # Get the left bounding box coordinate for the given span
    span_left = rdw.get_span_ngram(x["char_start"], x["char_end"]).left
    # Label INVALID if span bounding box begins too far to the left
    if span_left < 1100:
        return "INVALID"
    return "UNKNOWN"

sf.add_code_lf(node, rich_doc_lf, label="INVALID")

Classes

HVLines(dfs_horz, dfs_vert)

Serializable wrapper for lists of horizontal and vertical lines in image.

RichDoc(pages, areas, pars, lines, words[, ...])

An object representing a document with rich formatting preserved.

RichDocCols()

Base class that specifies Rich Doc columns.

RichDocList(rich_docs)

Serializable wrapper for list of RichDoc's.

Serializable()

Interface for types that have a custom serialization / deserialization function.

TextClusters(word_to_cluster, df_clusters)

Serializable wrapper for horizontal clusters of words.

Exceptions

MissingRichDocException

An Exception raised when a RichDoc was expected and could not be found.