USAMO/segment_script/segment_usamo.py

# -----------------------------------------------------------------------------
# Author: Marina
# Date: 2024-11-15
# -----------------------------------------------------------------------------
''' Script to segment md files in en-usamo, en-tstst, en-tst, en-jmo folder using regex. 
To run: 
`python segment_usamo.py`
'''

import json
from pathlib import Path
import warnings
warnings.filterwarnings("ignore", category=DeprecationWarning)

import os 
import re
import pandas as pd
from rapidfuzz import fuzz


# "## §0 Problems" -> match
section_re = re.compile(r"^#{1,2}\s(?:Contents|Problem|§[\d.]+.*)")
# section_re = re.compile(r"^#{1,2}\s(?:§[\d.]+.*)")

# "## §1.3 USAMO 2024/3, proposed by Krit Boonsiriseth" ->3
# "## §2.3 TSTST 2011/6" ->6
# "## §1.1 TSTST 2021/1, proposed by Holden Mui" -> 1
# "## §1.2 USA TST 2020/2, proposed by Merlijn Staps" -> 2
# "# §1.3 USA TST 2019/3, proposed by Nikolai Beluhov" -> 3
solution_label_re = re.compile(
    r"^#{1,2}\s§[\d.]+\s[A-Za-z0-9 ]+\s\d{4}/(\d+)(?:,\s.*)?$"
)

# "1. Prove that the average of the numbers $n" -> match
# "2 . For any nonempty set $S$ of real numbers,"" -> match
problem_re = re.compile(r"^(\d+)\s?\.\s(.*(?:\n\s+.*)*)")

# "## Problem statement extra text" -> match
# "Problem statement" -> match
# "##Problem statement (missing space)" -> match
solution_re = re.compile(r"^#{0,2}\s?Problem statement\b.*$")

# not actually used, only for debugging
pattern_debug = re.compile(
    r"^[【『\\]*.*?\b(First|Second|Third|Fourth|Fifth|Sixth|Seventh|Eighth|Ninth|Tenth|Complex|Inversion|Synthetic|One|Another|Solution)\b.*\b(solution|approach|proof)\b.*",
    re.IGNORECASE
)

# "Solution 1" -> match
# "【 Solution 1" -> match
solution_split_re1 = re.compile(r"\bSolution\s[1-9]\b")

# "First solution" -> match
# "【 Third approach" -> match
# "【 Second approach" -> match
solution_split_re2 = re.compile(r"\b(First|Second|Third|Fourth|Fifth|Sixth|Seventh|Eighth|Ninth|Synthetic)\b\s+(solution|approach|proof)\b")

# catch special cases (ugly but it works). To generate them run script with `DEBUG = True``
# and identify in console which special_cases should be catched as first line of a solution
DEBUG = False
special_cases = [
"【 First short solution, by Jeffrey Kwan. Let $p_{0",
"II Second longer solution using an invariant. Visu",
"【 Complex solution (Evan Chen). Toss on the comple",
"Second (longer) solution. If one does not notice t",
"『 Second calculation approach (along the lines of ",
"T Outline of second approach (by convexity, due t",
"I Inversion solution submitted by Ankan Bhattacha",
"【 Complex numbers approach with Apollonian circles",
" A second solution. Both lemmas above admit varia",
"【 A third remixed solution. We use Lemma I and Lem",
"【I A fourth remixed solution. We also can combine ",
"I First grid-based solution. The following solutio",
"Another short solution. Let $Z$ be on line $B D E$",
"【 Most common synthetic approach. The solution hin",
"\\ First \"local\" solution by swapping two points. L",
"Second general solution by angle chasing. By Rei",
"Third general solution by Pascal. Extend rays $A",
"【 Second length solution by tangent lengths. By $t",
"【 Angle chasing solution. Note that $(B D A)$ and",
"【 Harmonic solution (mine). Let $T$ be the point o",
"【 Pascal solution (Zuming Feng). Extend ray $F D$",
"『 A spiral similarity approach (Hans $\mathbf{Y u}",
"ब The author's original solution. Complete isoscel",
"l Evan's permutation-based solution. Retain the n",
"I Original proposer's solution. To this end, let's",
"【 Cartesian coordinates approach with power of a p",
"【 Cartesian coordinates approach without power of", 
"I III-advised barycentric approach (outline). Use",
"【 Approach using difference of squares (from autho",
"【 Divisibility approach (Aharshi Roy). Since $p q-",
"Solution with Danielle Wang: the answer is that $|",
"【 Homothety solution (Alex Whatley). Let $G, N, O$",
"【 Power of a point solution (Zuming Feng, official",
"【 Solution by Luke Robitaille. Let $Q$ be the seco",
"ๆ Solution with coaxial circles (Pitchayut Saengru",
"【 Solution to generalization (Nikolai Beluhov). We",
"【 Approach by deleting teams (Gopal Goel). Initial",
"【 Approach by adding colors. For a constructive al",
"【 Solution using spiral similarity. We will ignore",
"『 Barycentric solution (by Carl, Krit, Milan). We",
"I A Menelaus-based approach (Kevin Ren). Let $P$ b",
"【 Barycentric solution. First, we find the coordin",
"【 Angle chasing solution (Mason Fang). Obviously $",
"【 Inversive solution (Kelin Zhu). Invert about $A$",
"l The one-liner. ",
" The external power solution. We distinguish betw",
"Cauchy-Schwarz approach. Apply Titu lemma to get",
"đ Cauchy-Schwarz approach. The main magical claim ",
"『 Alternate solution (by proposer). Let $L$ be dia"
]


def add_content(current_dict):
    if not current_dict["lines"] or not current_dict["label"] :
        return
    text_str = " ".join(current_dict["lines"]).strip()
    entry =  {"label": current_dict["label"]}
    if current_dict["class"] == "problem":
        entry["problem"] = text_str
        current_dict["problems"].append(entry)
    elif current_dict["class"] == "solution":
        entry["solution"] = text_str
        entry["solution_lines"] = current_dict["lines"]
        current_dict["solutions"].append(entry)


def parse(file):
    content = file.read_text(encoding="utf-8")
    current = {
        "label": None,
        "class": None,
        "lines": [],
        "problems": [],
        "solutions": []
    }
    for line in content.splitlines(): 
        if match := section_re.match(line): # match a section
            add_content(current)
            if "problems" in line.lower(): # match problem section
                current["class"] = "problem"
            elif sub_match:= solution_label_re.match(line): # match solutions section, extract label for join
                current["class"] = "other"
                current["label"] = sub_match.group(1)
            elif match := solution_re.match(line): # match solutions subsection
                current["class"] = "solution"
            else:
                current["class"] = "other"
            current["lines"] = []
        elif match := problem_re.match(line): # match a problem
            if current["class"] == "solution": # handle wrong problem match
                current["lines"].append(line)
            else:
                add_content(current)
                label, text = match.groups()
                current["label"] = label
                current["lines"] = [text]
        else:
            if current["class"]=="solution" or current["class"]=="problem":
                current["lines"].append(line)
    add_content(current)
    problems_df = pd.DataFrame(current["problems"])
    solutions_df = pd.DataFrame(current["solutions"])
    return problems_df, solutions_df


def parse_solution(lines):
    """parses lines of a solution, finds multiple solutions and splits them"""
    solutions = []
    current = []
    for line in lines:
        if match := solution_split_re1.search(line): # match a solution (case1)
            solutions.append(" ".join(current).strip())
            current = [line]
        elif match := solution_split_re2.search(line):  # match a solution (case1)
            solutions.append(" ".join(current).strip())
            current = [line]
        elif any(case.lower() in line.lower() for case in special_cases): # match a solution (handle special_case)
            solutions.append(" ".join(current).strip())
            current = [line]
        elif  any(case.lower() in line[:50].lower() for case in ["solution", "approach", "proof"]):
            if DEBUG:
                if not any(case.lower() in line[:50].lower() for case in ["remark", "proof.", "proof", "approaches", "solutions"]):
                    print(line[:50])
        else:
            current.append(line)
    solutions.append(" ".join(current).strip())
    return solutions

def find_mult_solutions(solutions_df):
    """apply parse_solution to all df"""
    solutions_df["solution"] = solutions_df["solution_lines"].apply(lambda v: parse_solution(v))
    solutions_df = solutions_df.drop(columns=["solution_lines"])
    solutions_df = solutions_df.explode('solution', ignore_index=True)
    return solutions_df


def join(problems_df, solutions_df):
    pairs_df = problems_df.merge(solutions_df, on=["label"], how="outer")
    return pairs_df


def clean(pairs_df):
    '''removes the problem statement from the solution in an approximate way'''
    def find_closest_char(s, i, char):
        left = s.rfind(char, 0, i)  # Find the last '.' before index i
        right = s.find(char, i)    # Find the first '.' after or at index i
        if left == -1 and right == -1:
            return None  # No '.' found
        elif left == -1:  # No '.' on the left
            return right
        elif right == -1:  # No '.' on the right
            return left
        else:  # Closest '.' on either side
            return left if abs(i - left) <= abs(i - right) else right
    def remove_approx_match(row, threshold=90):
        problem = row["problem"]
        solution = row["solution"]
        similarity = fuzz.partial_ratio(problem, solution)
        if similarity >= threshold:
            i = find_closest_char(solution, len(problem), problem[-1])
            if i is not None:
                solution = solution[i+1:]
        return solution
    pairs_df["solution"] = pairs_df.apply(remove_approx_match, axis=1)
    return pairs_df    


def process_mult_solutions(pairs_df):
    '''in case of multiple solutions, prepend common text to all solutions'''
    def prepend_to_solution(group):
        if len(group) == 1:
            return group
        first_row = group.iloc[0]
        comment = f"{first_row['solution']}"
        group = group.iloc[1:].copy()
        group["solution"] = group["solution"].apply(lambda x: f"{comment} {x}")
        return group
    pairs_df = pairs_df.groupby("label", group_keys=False).apply(prepend_to_solution).reset_index(drop=True)
    return pairs_df


def add_metadata(pairs_df, year, tier, resource_path):
    pairs_df['year'] = year
    pairs_df['tier'] = tier # according to omnimath 
    pairs_df['exam'] = 'USAMO'
    pairs_df['metadata'] = [{"resource_path": resource_path}] * len(pairs_df)
    pairs_df['problem_type'] = None
    pairs_df.rename(columns={'label': 'problem_label'}, inplace=True)
    return pairs_df[['year', 'tier', 'problem_label', 'problem_type', 'exam', 'problem', 'solution', 'metadata']]


def write_pairs(file_path, pairs_df):
    pairs_df = pairs_df.replace({pd.NA: None, pd.NaT: None, float("nan"): None})
    pairs_dict = pairs_df.to_dict(orient="records")
    output_text = ""
    for pair in pairs_dict:
        output_text += json.dumps(pair, ensure_ascii=False) + "\n"
    file_path.write_text(output_text, encoding="utf-8")


if __name__ == "__main__":
    project_root = Path(__file__).parent.parent.parent

    problem_count = 0
    solution_count = 0

    tier = "T1"
    compet_base_path = Path(__file__).resolve().parent.parent
    compet_md_path = compet_base_path / "md"
    seg_output_path = compet_base_path / "segmented"

    for md_file in compet_md_path.glob("**/*.md"):
        year = re.search(r"\d{4}", md_file.name).group()
        output_file = seg_output_path / md_file.relative_to(compet_md_path).with_suffix(".jsonl")
        output_file.parent.mkdir(parents=True, exist_ok=True)

        problems, solutions = parse(md_file)
        solutions = find_mult_solutions(solutions)
        pairs_df = join(problems, solutions)
        pairs_df = clean(pairs_df)
        pairs_df = process_mult_solutions(pairs_df)
        pairs_df = add_metadata(
            pairs_df, year, tier, output_file.relative_to(project_root).as_posix()
        )
        problem_count += len(problems)
        solution_count += len(pairs_df)
        write_pairs(output_file, pairs_df)

    print(f"problem count: {problem_count}")
    print(f"solution count: {solution_count}")

# en-usamo
# problem count: 174
# solution count: 203