parse_guardian

topic_segmentation.news_dataset.parse_guardian

Parse saved pages in data/pages/guardianint_subheadings/ into data/outputs/guardian_segments.json.

is_noise_heading(text: str) -> bool

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def is_noise_heading(text: str) -> bool:
    t = text.lower().strip()
    return len(t) < 3 or any(pat in t for pat in NOISE_PATTERNS)

is_skip_container(el) -> bool

Identify subtrees to exclude from article content.

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def is_skip_container(el) -> bool:
    """Identify subtrees to exclude from article content."""
    # Rich-link figures (related article widgets)
    if el.name == "figure" and el.get("data-spacefinder-role") == "richLink":
        return True
    if el.name == "aside" and el.get("data-gu-name") != "pullquote":
        return True
    if el.name in ("form", "nav"):
        return True
    return False

is_link_only_list(el) -> bool

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def is_link_only_list(el) -> bool:
    if el.name not in ("ul", "ol"):
        return False
    items = el.find_all("li", recursive=False)
    if not items:
        return False
    for li in items:
        li_text = li.get_text(" ", strip=True)
        a_text = "".join(a.get_text(" ", strip=True) for a in li.find_all("a"))
        if not li_text or not a_text:
            return False
        if li_text == a_text:
            continue
        if len(a_text) / len(li_text) >= 0.6:
            continue
        return False
    return True

is_trailing_element(el) -> bool

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def is_trailing_element(el) -> bool:
    if el.name in ("h2", "h3", "h4"):
        t = el.get_text(" ", strip=True).lower().rstrip(":")
        if t in TRAILING_HEADINGS:
            return True
        if any(t.startswith(p) for p in TRAILING_HEADING_PREFIXES):
            return True
    text_lower = el.get_text(" ", strip=True).lower()
    if any(text_lower.startswith(p) for p in TRAILING_TEXT_PREFIXES):
        return True
    return False

clean_inner_html(el) -> str

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def clean_inner_html(el) -> str:
    el_copy = copy.deepcopy(el)
    # Replace decorative SVG quote marks with a real quote character
    for svg in el_copy.find_all("svg"):
        svg.replace_with("\u201c")
    for tag in el_copy.find_all(True):
        if tag.name == "script":
            tag.decompose()
            continue
        # Preserve bullet markers
        if tag.name == "span" and tag.get("data-dcr-style") == "bullet":
            tag.attrs = {"data-dcr-style": "bullet"}
            continue
        for attr in STRIP_ATTRS:
            tag.attrs.pop(attr, None)
        if tag.name == "a" and (tag.get("href") or "").startswith("/"):
            tag["href"] = "https://www.theguardian.com" + tag["href"]
            tag["target"] = "_blank"
            tag["rel"] = "noopener"
    return el_copy.decode_contents().strip()

collect_elements(container) -> list

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def collect_elements(container) -> list:
    seen_ids: set = set()
    elements: list = []

    def walk(node):
        for el in node.children:
            if not hasattr(el, "name") or el.name is None:
                continue
            if is_skip_container(el):
                continue
            eid = id(el)
            if eid in seen_ids:
                continue
            seen_ids.add(eid)
            if el.name in BLOCK_TAGS:
                # Recurse into figure wrappers containing headings.
                if el.name == "figure" and el.find(["h2", "h3", "h4"]):
                    walk(el)
                elif not is_link_only_list(el):
                    elements.append(el)
            else:
                walk(el)

    walk(container)
    return elements

process_article(aid: str, raw_html: str, meta: dict)

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def process_article(aid: str, raw_html: str, meta: dict):
    soup = BeautifulSoup(ftfy.fix_text(raw_html), "html.parser")

    body = soup.find(attrs={"data-gu-name": "body"})
    if not body:
        return None

    # Title
    h1 = soup.find("h1")
    title = h1.get_text(" ", strip=True) if h1 else meta.get("title", "")

    # Standfirst — skip if it's a pure link (newsletter CTA) or boilerplate
    standfirst_text = None
    sf_div = soup.find(attrs={"data-gu-name": "standfirst"})
    if sf_div:
        # Try <p> first, fall back to the whole standfirst div text
        p = sf_div.find("p")
        sf_el = p if p else sf_div
        text = sf_el.get_text(" ", strip=True)
        if text:
            # Skip if entire text is inside a link (e.g. "Sign up now!")
            a_text = "".join(a.get_text(" ", strip=True) for a in sf_el.find_all("a"))
            is_pure_link = a_text and a_text == text
            is_cta = text.lower().startswith("sign up") or "join the conversation" in text.lower()
            if not is_pure_link and not is_cta:
                standfirst_text = text

    # Hero image from data-gu-name="media"
    hero_img_src = None
    media_div = soup.find(attrs={"data-gu-name": "media"})
    if media_div:
        pic = media_div.find("picture")
        if pic:
            img_tag = pic.find("img")
            if img_tag:
                hero_img_src = img_tag.get("src", "")

    sentences, html_parts, labels, title_labels = [], [], [], []

    if standfirst_text:
        sentences.append(standfirst_text)
        html_parts.append({"tag": "p", "html": standfirst_text})
        labels.append(0)
        title_labels.append(1)

    if hero_img_src:
        sentences.append("")
        html_parts.append({"tag": "figure", "html": f'<img src="{hero_img_src}" style="max-width:100%">'})
        labels.append(0)
        title_labels.append(0)

    first_elem = not (standfirst_text or hero_img_src)

    for el in collect_elements(body):
        if is_trailing_element(el):
            break

        # Include figures containing a picture.
        if el.name == "figure":
            pic = el.find("picture")
            if pic:
                img = pic.find("img")
                if img:
                    src = img.get("src", "")
                    if src:
                        # Caption
                        figcap = el.find("figcaption")
                        cap_text = ""
                        if figcap:
                            cap_text = figcap.get_text(" ", strip=True)
                        cap_html = f'<figcaption>{cap_text}</figcaption>' if cap_text else ""
                        sentences.append(cap_text)
                        html_parts.append({"tag": "figure", "html": f'<img src="{src}" style="max-width:100%">{cap_html}'})
                        labels.append(0)
                        title_labels.append(0)
                        first_elem = False
            continue

        # Replace decorative SVGs with quote character before extracting text
        for svg in el.find_all("svg"):
            svg.replace_with("\u201c")
        text = el.get_text(" ", strip=True)
        if not text:
            continue

        if el.name in ("h2", "h3", "h4") and is_noise_heading(text):
            continue
        if any(text.lower().startswith(p) for p in SKIP_PARA_PREFIXES):
            continue

        is_heading = el.name in ("h2", "h3", "h4") and len(text) <= 200

        sentences.append(text)
        html_parts.append({"tag": el.name, "html": clean_inner_html(el)})
        if is_heading:
            labels.append(0 if first_elem else 1)
            title_labels.append(1)
        else:
            labels.append(0)
            title_labels.append(0)
        first_elem = False

    # Remove trailing headings without following content.
    while sentences and title_labels[-1] == 1:
        sentences.pop()
        html_parts.pop()
        labels.pop()
        title_labels.pop()

    if not sentences:
        return None

    return {
        "articleId": aid,
        "outlet": "GuardianInt",
        "title": title,
        "url": meta.get("url", ""),
        "sentences": sentences,
        "html_parts": html_parts,
        "labels": labels,
        "title_labels": title_labels,
        "num_sections": labels.count(1) + 1,
    }

main()

Source code in src/topic_segmentation/news_dataset/parse_guardian.py

def main():
    with open(ROOT / "data/articles.json", encoding="latin-1") as f:
        all_meta = {a["articleId"]: a for a in json.load(f)}

    results = []
    seen_content: set = set()
    for html_file in sorted((ROOT / "data/pages/guardianint_subheadings").glob("*.html")):
        aid = html_file.stem
        raw = html_file.read_text(encoding="utf-8")
        rec = process_article(aid, raw, all_meta.get(aid, {}))
        if not rec:
            continue
        content = (rec.get("title", "") + "".join(rec["sentences"]))
        if content in seen_content:
            continue
        seen_content.add(content)
        results.append(rec)

    results = [r for r in results if r["num_sections"] >= 2]
    print(f"Processed {len(results)} articles")

    (ROOT / "data/outputs").mkdir(parents=True, exist_ok=True)
    seg_path = ROOT / "data/outputs/guardian_segments.json"
    with open(seg_path, "w", encoding="utf-8") as f:
        json.dump(results, f, ensure_ascii=False, indent=2)
    print(f"Saved {seg_path}")