SEC提出書類のセクション抽出と比較

# Text Data Extraction: Structuring Unstructured Financial Documents


# Text Data Extraction: Structuring Unstructured Financial Documents

**Chapter 4: Fundamental and Alternative Data**
**Docker image**: `ml4t`
**Section Reference**: Section 4.5 (Using Text Data for NLP Features)

## Purpose

An annual report is a quarter of a million characters of which a model wants perhaps a
twentieth. The Securities and Exchange Commission prescribes what a 10-K must contain and
numbers the parts, so the document has a structure; what it does not have is markup that a
parser can rely on, because the numbering appears in the table of contents, in cross-references
inside the narrative, and again as the actual section heading, in whatever typography the filer
chose.

Getting from a filing to a section, reliably, over thousands of documents, is the whole job of
this notebook. Everything downstream in the book - sentiment scoring, topic models, embeddings -
assumes that job was done properly, and the failure mode is silent: an extractor that returns
the table of contents entry instead of the section returns a plausible-looking short string.

## Learning Objectives

After completing this notebook, you will be able to:

- Name the items of a 10-K a research pipeline usually wants, and say where the same content
  sits in a 10-Q.
- Strip HTML from a filing without destroying the paragraph boundaries a later stage needs.
- Locate a section's boundaries when its item number appears many times in the document, and
  explain why the first match is the wrong one.
- Check an extraction for the two failures it can have - returning nothing, and running past
  the section's end - and report the rate rather than assuming success.
- Measure how much of a section changed between two consecutive filings, which is the input to
  the change-detection signals in Chapter 10.

## Prerequisites

This notebook reads live filings from EDGAR, which requires an identifying `User-Agent`. It is
not an API key and there is nothing to sign up for: put your own name and email on the
`EDGAR_IDENTITY` line of the `.env` file in the repository root, before you start Jupyter.

```
EDGAR_IDENTITY=Jane Doe jane@example.org
```

## Cross-References

- **Related**: [`02_sec_filing_explorer`](02_sec_filing_explorer.ipynb) (reaching filings through the library)
- **Downstream**: `10_text_feature_engineering/09_filing_text_signals.py` (signals built on filing text)

## Key Concepts

- **10-K**: the annual report, filed once a year and the most complete narrative disclosure.
- **10-Q**: the quarterly report, shorter, and with the same content under different item
  numbers.
- **Item 1A, Risk Factors**: the risks management is required to disclose. The largest narrative
  block in a modern filing and the most mined.
- **Item 7, Management's Discussion and Analysis**: management's account of the period's
  results, in its own words.
- **Accession number**: the identifier the SEC assigns to a filing. Unique, permanent, and the
  right primary key for a text dataset.

```python
"""Text Data Extraction - extract and structure high-value text blocks from SEC filings for NLP analysis."""

import os
import re
import warnings

# EdgarTools re-exports three legacy HTML helpers and each one raises a DeprecationWarning as it
# is imported. They are the same three every time and say nothing about the filings.
warnings.filterwarnings("ignore", category=DeprecationWarning, module="edgar")

import plotly.express as px
import polars as pl
from bs4 import BeautifulSoup
from edgar import Company, set_identity

from utils.paths import get_output_dir
from utils.style import COLORS, show_plotly_with_alt
```

The company and form choose which filings are read. Two consecutive filings of the same form
are fetched, because the second is what makes the change measurement in Part 6 possible; one
filing alone can be extracted from but not compared.

```python
TICKER = "AAPL"  # any filer with a long 10-K history works here
FORM = "10-K"  # switch to "10-Q" to exercise the quarterly item numbering
N_FILINGS = 2  # consecutive filings to read, most recent first
MIN_SECTION_WORDS = 100  # an extraction shorter than this is treated as a failure
MAX_SECTION_WORDS = 50_000  # and one longer than this as having run past its boundary
```

```python
OUTPUT_DIR = get_output_dir(4, "sec_text")
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
```

## 1. What a filing is made of

A 10-K is a numbered sequence of items whose meaning is fixed by regulation. Three of them carry
almost all the narrative a research pipeline wants.

| Item | Section | What it holds |
|------|---------|---------------|
| 1 | Business | What the company does, its segments and its markets |
| 1A | Risk Factors | The risks management is required to disclose |
| 7 | Management's Discussion and Analysis | Management's account of the period's results |
| 7A | Quantitative and Qualitative Disclosures | Market risk exposures |
| 8 | Financial Statements | The audited financials |

The numbering is **form-specific**, and this is the single most common source of a silently
empty extraction. In a 10-Q, management's discussion is Item 2 of Part I, not Item 7, and the
risk factors, if updated at all, are Item 1A of Part II. An extractor that hardcodes the 10-K
numbers returns nothing on a 10-Q and says nothing about why.

```python
SECTIONS = {
    "10-K": {
        "business": {
            "starts": [r"^\s*ITEM\s*1\.?\s*[-–—]?\s*BUSINESS"],
            "ends": [r"^\s*ITEM\s*1A\b", r"^\s*ITEM\s*2\b"],
        },
        "risk_factors": {
            "starts": [r"^\s*ITEM\s*1A\.?\s*[-–—]?\s*RISK\s*FACTORS"],
            "ends": [r"^\s*ITEM\s*1B\b", r"^\s*ITEM\s*2\b"],
        },
        "mda": {
            # Filers use both the straight and the curly apostrophe in "Management's".
            "starts": [r"^\s*ITEM\s*7\.?\s*[-–—]?\s*MANAGEMENT(?:'|’)?S?\s*DISCUSSION"],
            "ends": [r"^\s*ITEM\s*7A\b", r"^\s*ITEM\s*8\b"],
        },
    },
    "10-Q": {
        "risk_factors": {
            "starts": [r"^\s*ITEM\s*1A\.?\s*[-–—]?\s*RISK\s*FACTORS"],
            "ends": [r"^\s*ITEM\s*1B\b", r"^\s*ITEM\s*2\b"],
            "after": r"PART\s*II",
        },
        "mda": {
            "starts": [r"^\s*ITEM\s*2\.?\s*[-–—]?\s*MANAGEMENT(?:'|’)?S?\s*DISCUSSION"],
            "ends": [r"^\s*ITEM\s*3\b"],
        },
    },
}
FLAGS = re.IGNORECASE | re.MULTILINE | re.DOTALL
```

## 2. The filings

EDGAR requires every request to identify its sender, and rejects placeholder addresses. The
fetch is read-only and rate-limited by the library.

```python
identity = os.environ.get("EDGAR_IDENTITY")
if not identity:
    raise RuntimeError(
        "EDGAR_IDENTITY is not set. The SEC requires a real User-Agent - your "
        "name and email - on every EDGAR request, and blocks placeholder "
        "addresses. It is not an API key and there is nothing to sign up for.\n"
        "Put your own name and email on the EDGAR_IDENTITY line of the .env "
        "file in the repository root:\n"
        "    EDGAR_IDENTITY=Jane Doe jane@example.org\n"
        ".env is read once, when the process starts, so then restart this "
        "notebook's kernel (Kernel -> Restart Kernel). On the Docker path, stop "
        "Jupyter Lab and run `docker compose up ml4t` again - `docker compose "
        "restart` keeps the environment the container was created with."
    )
set_identity(identity)

# `amendments=False` keeps 10-K/A out of the list. An amendment covers a period the original
# already covered and often omits the narrative items entirely, so a pair drawn without this
# filter can be two filings of the same period presented as consecutive ones.
filings = Company(TICKER).get_filings(form=FORM, amendments=False)[:N_FILINGS]
documents = [
    {
        "cik": filing.cik,
        "company_name": filing.company,
        "accession_no": filing.accession_no,
        "form": filing.form,
        "filing_date": filing.filing_date,
        "accepted_at": filing.acceptance_datetime,
        "period_end": filing.period_of_report,
        "text": filing.text(),
    }
    for filing in filings
]

pl.DataFrame(
    [
        {
            "accession_no": d["accession_no"],
            "filing_date": d["filing_date"],
            "period_end": d["period_end"],
            "characters": len(d["text"]),
        }
        for d in documents
    ]
)
```

The acceptance timestamp is worth carrying alongside the filing date. A filing accepted after
the close is public that evening and tradeable the next morning, and a strategy that acts on the
filing date alone is a few hours early on every after-hours submission.

## 3. Cleaning

`filing.text()` already returns text rather than markup, but a pipeline that fetches documents
directly gets HTML, so the conversion belongs here. The one thing to preserve is the paragraph
boundary: a converter that joins everything with spaces destroys the unit that the change
detection in Part 6 compares.

```python
def html_to_text(html: str, drop_tables: bool = False) -> str:
    """Convert filing HTML to text, keeping one newline per block element."""
    if not html:
        return ""
    soup = BeautifulSoup(html, "html.parser")
    for element in soup(["script", "style"]):
        element.decompose()
    if drop_tables:
        # Financial tables inside a narrative section are numbers, not prose; dropping them is
        # right for sentiment work and wrong for anything reading the figures.
        for table in soup.find_all("table"):
            table.decompose()
    return soup.get_text(separator="\n")


def clean_text(text: str) -> str:
    """Remove filing furniture and normalize whitespace without flattening paragraphs."""
    if not text:
        return ""
    text = text.replace("\r\n", "\n").replace("\r", "\n")
    # Page furniture, anchored to whole lines so a mention inside a sentence survives.
    text = re.sub(r"(?im)^\s*table of contents\s*$", "", text)
    text = re.sub(r"(?im)^\s*page\s+\d+\s*$", "", text)
    text = re.sub(r"(?im)^\s*\d+\s*of\s*\d+\s*$", "", text)
    text = re.sub(r"https?://\S+", "", text)
    text = re.sub(r"[_=\-]{3,}", " ", text)
    text = re.sub(r"[•●◦▪]", " ", text)
    text = "\n".join(re.sub(r"[ \t]+", " ", line).strip() for line in text.split("\n"))
    return re.sub(r"\n{3,}", "\n\n", text).strip()
```

## 4. Finding a section

The naive approach is to search for the item heading and take everything to the next one. It
fails on the first line of every filing, because the item heading appears in the table of
contents before it appears as a section, and it appears again wherever the narrative
cross-references another item.

Counting the matches in a real document is the quickest way to see the size of the problem.

```python
first = documents[0]
for section, config in SECTIONS[FORM].items():
    matches = [
        m.start()
        for pattern in config["starts"]
        for m in re.finditer(pattern, first["text"], FLAGS)
    ]
    print(f"{section}: {len(matches)} places match the heading pattern")
```

Every section matches in more than one place, and only one of those places is the section. What
separates them is **span**: a table of contents entry is followed by the next contents entry a
line or two later, while the real heading is followed by the whole section before the next item
begins. Taking the candidate with the longest span to its own end pattern picks the section
without needing to know where the table of contents ends, and it keeps working on a filer whose
contents page is formatted differently.

```python
def extract_section(text: str, section: str, form: str) -> str | None:
    """The longest span between this section's heading and the item that follows it."""
    config = SECTIONS[form][section]

    # A 10-Q's risk factors live in Part II, and Part I has an Item 1A of its own meaning
    # something else. Where a section declares a marker, the search starts after it.
    offset, haystack = 0, text
    if marker := config.get("after"):
        if found := re.search(marker, text, FLAGS):
            offset, haystack = found.start(), text[found.start() :]

    starts = sorted(
        {
            offset + m.end()
            for pattern in config["starts"]
            for m in re.finditer(pattern, haystack, FLAGS)
        }
    )
    if not starts:
        return None

    best = None
    for start in starts:
        end = len(text)
        for pattern in config["ends"]:
            if found := re.search(pattern, text[start:], FLAGS):
                end = start + found.start()
                break
        if best is None or (end - start) > (best[1] - best[0]):
            best = (start, end)
    return clean_text(text[best[0] : best[1]])
```

## 5. The dataset, and whether the extraction worked

One row per filing and section, keyed on the accession number. The word count is what the
quality check reads: an extraction can fail by returning nothing, by returning a fragment - the
table of contents entry, if the span rule went wrong - or by running past its own end and
swallowing the next item.

```python
records = [
    {
        "cik": document["cik"],
        "company_name": document["company_name"],
        "accession_no": document["accession_no"],
        "form": document["form"],
        "filing_date": document["filing_date"],
        "accepted_at": document["accepted_at"],
        "period_end": document["period_end"],
        "section": section,
        "text": extracted or "",
        "word_count": len(extracted.split()) if extracted else 0,
    }
    for document in documents
    for section in SECTIONS[FORM]
    for extracted in [extract_section(document["text"], section, FORM)]
]
dataset = pl.DataFrame(records).with_columns(
    quality=pl.when(pl.col("word_count") == 0)
    .then(pl.lit("nothing extracted"))
    .when(pl.col("word_count") < MIN_SECTION_WORDS)
    .then(pl.lit("too short"))
    .when(pl.col("word_count") > MAX_SECTION_WORDS)
    .then(pl.lit("ran past the boundary"))
    .otherwise(pl.lit("ok")),
    # A section that swallowed the next item carries that item's heading inside it.
    boundary_leak=pl.when(pl.col("section") == "mda")
    .then(pl.col("text").str.contains(r"(?im)^\s*ITEM\s*8\b"))
    .when(pl.col("section") == "risk_factors")
    .then(pl.col("text").str.contains(r"(?im)^\s*ITEM\s*1B\b"))
    .otherwise(False),
)

print(f"Extractions attempted: {len(dataset)}")
print(f"Extractions the quality check passed: {(dataset['quality'] == 'ok').sum()}")
print(f"Extractions carrying the next item's heading: {int(dataset['boundary_leak'].sum())}")
dataset.select("accession_no", "filing_date", "section", "word_count", "quality", "boundary_leak")
```

Reporting the rate rather than eyeballing one extraction is the point. Over two filings it is a
formality; over a corpus of thousands it is the only way to know that a formatting change at one
filer has not silently emptied a column, and the same three columns scale unchanged.

```python
latest_sections = dataset.filter(pl.col("accession_no") == documents[0]["accession_no"]).sort(
    "word_count", descending=True
)
labels = {"business": "Business", "risk_factors": "Risk Factors", "mda": "MD&A"}
fig = px.bar(
    latest_sections.with_columns(label=pl.col("section").replace(labels)).to_pandas(),
    x="word_count",
    y="label",
    orientation="h",
    title="Risk factors are several times longer than any other narrative section",
    labels={"word_count": "Words extracted", "label": ""},
    color_discrete_sequence=[COLORS["blue"]],
    text="word_count",
)
fig.update_traces(textposition="outside", cliponaxis=False)
fig.update_layout(
    height=320,
    yaxis=dict(autorange="reversed"),
    xaxis_range=[0, float(latest_sections["word_count"].max()) * 1.18],
    margin=dict(l=120, r=60),
)
show_plotly_with_alt(
    fig,
    "Horizontal bar chart of the words extracted from each narrative section of the most recent "
    "filing. The risk factors bar is several times longer than the other two, which are of "
    "similar length to each other.",
)
```

The proportions are why risk factors are the most mined section in the corpus. They are also why
the section is hard to use: a block that long changes a little every year for reasons that have
nothing to do with the business, so the signal is in what changed rather than in what it says.

## 6. What changed between the two filings

The same section from two consecutive filings can be compared directly. Two measurements are
worth having and they answer different questions. **Jaccard similarity** over the vocabulary
says how much of the language is shared, which is a measure of drift. The **paragraphs present
in the new filing and absent from the old** are the additions, and those are the ones that name
a risk management decided to start disclosing.

```python
def vocabulary_overlap(old: str, new: str) -> dict:
    """Shared vocabulary between two documents, and what each has that the other does not."""
    old_words, new_words = set(old.lower().split()), set(new.lower().split())
    union = old_words | new_words
    return {
        "jaccard_similarity": len(old_words & new_words) / len(union) if union else 0.0,
        "words_only_in_new": len(new_words - old_words),
        "words_only_in_old": len(old_words - new_words),
    }


def added_paragraphs(old: str, new: str, min_characters: int = 200) -> list[str]:
    """Paragraphs of the new document that do not appear verbatim in the old one."""
    old_paragraphs = {p.strip().lower() for p in old.split("\n\n")}
    return [
        paragraph.strip()
        for paragraph in new.split("\n\n")
        if len(paragraph.strip()) >= min_characters
        and paragraph.strip().lower() not in old_paragraphs
    ]
```

```python
CHANGE_SECTION = "risk_factors"
newer, older = documents[0]["accession_no"], documents[1]["accession_no"]
comparable = dataset.filter(
    (pl.col("section") == CHANGE_SECTION)
    & (pl.col("quality") == "ok")
    & ~pl.col("boundary_leak")
    & pl.col("accession_no").is_in([older, newer])
)

# An extraction the quality check rejected must not enter the comparison. An empty older section
# scores zero similarity and marks every paragraph of the newer one as an addition, which reads
# as a company rewriting its risk factors rather than as a failed extraction.
if comparable.height < 2:
    additions = []
    print(f"Both filings' {CHANGE_SECTION} did not pass the quality check; no comparison made.")
else:
    texts = {row["accession_no"]: row["text"] for row in comparable.iter_rows(named=True)}
    overlap = vocabulary_overlap(texts[older], texts[newer])
    additions = added_paragraphs(texts[older], texts[newer])
    print(f"Comparing {CHANGE_SECTION} between {older} and {newer}")
    for name, value in overlap.items():
        print(f"  {name}: {value:.3f}" if isinstance(value, float) else f"  {name}: {value}")
    print(f"  paragraphs in the newer filing that are not verbatim in the older: {len(additions)}")
```

A high vocabulary overlap with a substantial number of changed paragraphs is the normal result,
and it is why the vocabulary measure alone is a poor change detector: a filer can rewrite a
paragraph entirely without introducing a single new word.

The paragraph comparison is exact rather than fuzzy, so a single reworded clause makes a
paragraph count as new. That over-reports for a change-detection signal and under-reports
nothing, which is the safer direction for a screen a human reads; a production version scores
similarity between paragraphs rather than testing them for equality.

```python
for i, paragraph in enumerate(additions[:3], 1):
    print(f"--- changed paragraph {i} ---")
    print(paragraph[:400] + ("..." if len(paragraph) > 400 else ""))
    print()
```

## 7. Saving the dataset

Parquet, keyed on the accession number, with the timestamps and the section name beside the
text. The raw filing text is deliberately not stored alongside it: it is reproducible from the
accession number, and it is twenty times the size.

```python
output_file = OUTPUT_DIR / "sec_filing_sections.parquet"
dataset.write_parquet(output_file)
print(f"Wrote {len(dataset)} rows to {output_file}")
pl.read_parquet(output_file).select("accession_no", "section", "word_count", "quality")
```

## Key Takeaways

1. The item numbers are form-specific. Management's discussion is Item 7 of a 10-K and Item 2 of
   a 10-Q's Part I, so an extractor built for one silently returns nothing on the other. Select
   the pattern set from the form type rather than defaulting to it.
2. An item heading appears several times in a filing - in the table of contents, in
   cross-references, and once as the section. The first match is almost never the right one, and
   the span to the next item is what distinguishes them: a contents entry runs a line, a section
   runs thousands.
3. Preserve paragraph boundaries through the cleaning. They are the unit a change detector
   compares, and a converter that joins blocks with spaces destroys them irrecoverably.
4. Report the extraction rate, not an example. The failure is silent - a plausible short string
   where a section should be - and the only thing that catches it at corpus scale is counting
   the empty, the short, and the ones carrying the next item's heading.
5. Key the dataset on the accession number and carry the acceptance timestamp. A company and a
   date do not identify a filing when an amendment exists, and a filing accepted after the close
   is not tradeable until the next session.
6. Vocabulary overlap and changed paragraphs measure different things. A filer can rewrite a
   paragraph without adding a word, so a signal built on vocabulary alone misses exactly the
   rewrites that matter.
![notebook output](figures/p1_1.png)