generate.py

""" Generates dataset from CCEL ThML files """

import re
import os
import json
import xml.etree.ElementTree as ET
from languagemodels.embeddings import embed
from time import perf_counter


def get_refs(root):
    refs = []

    for child in root:
        if child.tag == "scripRef":
            if "osisRef" in child.attrib:
                refs.append(child.attrib["osisRef"])

    return refs


def get_text_content(root, strip_ref=False):
    """Return the plain text content of a node

    This will remove all notes, footnotes, and references
    """

    # Start with the root text prior to the first node
    text = root.text or ""

    for child in root:
        if child.text:
            if child.tag != "scripRef" or not strip_ref:
                text += child.text
        if child.tail:
            # Add the root node text between this child and the next child
            text += child.tail

    return text

def get_field(root, field):
    try:
        text = root.find(f".//{field}").text
        text = re.sub(r"\s+", " ", text)
    except:
        text = None

    return text

log = open("log.txt", "w")
jsonl = open("train.jsonl", "w")
count = 0


def get_paras():
    global count
    paras = []

    from pathlib import Path

    for filename in list(Path(".").rglob("*.xml")):
        filename = str(filename)
        if "authInfo." in filename:
            continue
        try:
            tree = ET.parse(filename)
        except:
            print("ERROR: Unable to parse:", filename)
            continue

        root = tree.getroot()

        # Skip content with copyright restrictions
        # The copyright statements aren’t consistent.
        # The ones that contain the text "public domain" (case insensitive)
        # or nothing at all should be safe.
        rights = get_field(root, "DC.Rights")

        if rights and 'public domain' not in rights.lower():
            print(f"Skipped {filename} due to copyright: {rights}", file=log, flush=True)
            continue

        title = get_field(root, "title")
        published = get_field(root, "firstPublished")
        author = get_field(root, "authorID")

        for i, p in enumerate(root.findall(".//p")):
            xml = ET.tostring(p, encoding="unicode")
            text = get_text_content(p)

            if "id" in p.attrib:
                start = perf_counter()
                refs = get_refs(p)
                emb = list(float(n) for n in embed([text])[0])
                count += 1

                row = {
                    "id": filename + ":" + p.attrib["id"],
                    "text": text,
                    "thml": xml,
                    "refs": refs,
                    "author": author,
                    "title": title,
                    "rights": rights,
                    "published": published,
                    "all-MiniLM-L6-v2": emb,
                }

                print(
                    count,
                    f"{perf_counter()-start:.2f}",
                    len(row["text"]),
                    len(row["thml"]),
                    row["id"],
                    row["all-MiniLM-L6-v2"][:2],
                    file=log,
                    flush=True,
                )

                print(json.dumps(row), file=jsonl, flush=True)

                yield row


from datasets import Dataset

ds = Dataset.from_generator(get_paras)
ds.push_to_hub("jncraton/ccel-paragraphs")