Extracting Error Messages from Documents

[Gitclop]

If i want to extract Error messages from human written reports, should i train the NER component?
Error messages can be multiple paragraphs long and are usualy indicated by the user stating something like "we get the following message: " or "we are seeing this error ..."

[svlandeg]

I would advice to try out a spancat model instead. spaCy's ner implementation is tuned towards recognizing actual named entities - short phrases that are names of specific people or locations. The spancat implementation allows to recognize longer and potentially overlapping spans. See also our blog post here. Good luck!

[Gitclop]

I am now creating the training examples, but get the Error: Invalid span: span is not from the same doc.

I have a list of predefined Errormessages that i use to mark the start and end of the span:

def annotate_text(row):
    error = row['Error']
    text = row['Text']
    tag = 'FEHLERMELDUNG'
    spans = {}
    span_key = "sc"

    index = 0
    while error in text[index:]:
        start_index = text[index:].find(error) + index
        end_index = start_index + len(error)
        if span_key not in spans:
            spans[span_key] = []
        spans[span_key].append((start_index, end_index, tag))
        index = end_index

    return {'spans': spans}

This returs a dictionary in the Format:
{'spans': {'sc': [(75, 129, 'FEHLERMELDUNG'), (1578, 1632, 'FEHLERMELDUNG')]}}
and adds it to the dataframe.

data['Annotated_Span'] = data.apply(annotate_text, axis=1)
if st.button("Spacy Format"):

    nlp = cs_spacy().load_model_parser()
    train_data = []
    db = DocBin()

    for index, row in data.iterrows():

        annotated_text = row['Annotated_Span']
        doc = nlp.make_doc(row['Text'])
        train_data.append(Example.from_dict(doc, annotated_text))

[Gitclop]

I think i got it and am currently training the model. Its still tricky though.

def annotate_text(df):

    list_of_docs = []

    for index, row in df.iterrows():
        error = row['Error']
        text = row['Text']
        tag = 'FEHLERMELDUNG'
        span_key = "sc"
        doc = nlp(text)
        doc_spans = []

        start_index = 0

        while start_index < len(text):
            error_start_index = text.find(error, start_index)

            if error_start_index == -1:
                break
            error_end_index = error_start_index + len(error)

            token_start_index = None
            for i, token in enumerate(doc):
                if token.idx == error_start_index:
                    token_start_index = i
                    break

            if token_start_index is not None:
                token_end_index = token_start_index + len(error.split()) 
                doc_spans.append(Span(doc, token_start_index, token_end_index, tag))

            start_index = error_end_index
        doc.spans[span_key] = doc_spans

        list_of_docs.append(doc)

    doc_bin = DocBin(docs=list_of_docs)
    doc_bin.to_disk("./SpanCat/train_span.spacy")