Separação eficiente de estruturas de texto complexas em documentos com 50 linhas de expressões regulares

Xiao Han, CEO da Jina, compartilhou um trecho de código impressionante no GitHub da implementação central de fragmentação usada no tokenizador da Jina. O trecho de código da expressão regular tem pouco mais de 50 linhas e, ainda assim, lida com eficiência com o conteúdo de texto de todas as complexidades. O desempenho é surpreendentemente robusto.

Experiência on-line: https://jina.ai/tokenizer/

// Updated: Aug. 15, 2024
// Run: node testRegex.js testText.txt
// Used in https://jina.ai/tokenizer
const fs = require('fs');
const util = require('util');

// Define variables for magic numbers
const MAX_HEADING_LENGTH = 7;
const MAX_HEADING_CONTENT_LENGTH = 200;
const MAX_HEADING_UNDERLINE_LENGTH = 200;
const MAX_HTML_HEADING_ATTRIBUTES_LENGTH = 100;
const MAX_LIST_ITEM_LENGTH = 200;
const MAX_NESTED_LIST_ITEMS = 6;
const MAX_LIST_INDENT_SPACES = 7;
const MAX_BLOCKQUOTE_LINE_LENGTH = 200;
const MAX_BLOCKQUOTE_LINES = 15;
const MAX_CODE_BLOCK_LENGTH = 1500;
const MAX_CODE_LANGUAGE_LENGTH = 20;
const MAX_INDENTED_CODE_LINES = 20;
const MAX_TABLE_CELL_LENGTH = 200;
const MAX_TABLE_ROWS = 20;
const MAX_HTML_TABLE_LENGTH = 2000;
const MIN_HORIZONTAL_RULE_LENGTH = 3;
const MAX_SENTENCE_LENGTH = 400;
const MAX_QUOTED_TEXT_LENGTH = 300;
const MAX_PARENTHETICAL_CONTENT_LENGTH = 200;
const MAX_NESTED_PARENTHESES = 5;
const MAX_MATH_INLINE_LENGTH = 100;
const MAX_MATH_BLOCK_LENGTH = 500;
const MAX_PARAGRAPH_LENGTH = 1000;
const MAX_STANDALONE_LINE_LENGTH = 800;
const MAX_HTML_TAG_ATTRIBUTES_LENGTH = 100;
const MAX_HTML_TAG_CONTENT_LENGTH = 1000;
const LOOKAHEAD_RANGE = 100;  // Number of characters to look ahead for a sentence boundary

// Define the regex pattern
// Headings
// Citations
// List items
// Block quotes
// Code blocks
// Tables
// Horizontal rules
// Standalone lines or phrases
// Sentences or phrases
// Quoted text, parenthetical phrases, or bracketed content
// Paragraphs
// HTML-like tags and their content
// LaTeX-style math expressions
// Fallback for any remaining content
// Read the regex and test text from files

const chunkRegex = new RegExp(
    "(" +
    // 1. Headings (Setext-style, Markdown, and HTML-style, with length constraints)
    `(?:^(?:[#*=-]{1,${MAX_HEADING_LENGTH}}|\\w[^\\r\\n]{0,${MAX_HEADING_CONTENT_LENGTH}}\\r?\\n[-=]{2,${MAX_HEADING_UNDERLINE_LENGTH}}|<h[1-6][^>]{0,${MAX_HTML_HEADING_ATTRIBUTES_LENGTH}}>)[^\\r\\n]{1,${MAX_HEADING_CONTENT_LENGTH}}(?:</h[1-6]>)?(?:\\r?\\n|$))` +
    "|" +
    // New pattern for citations
    `(?:\\[[0-9]+\\][^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}})` +
    "|" +
    // 2. List items (bulleted, numbered, lettered, or task lists, including nested, up to three levels, with length constraints)
    `(?:(?:^|\\r?\\n)[ \\t]{0,3}(?:[-*+•]|\\d{1,3}\\.\\w\\.|\\[[ xX]\\])[ \\t]+(?:(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[\\r\\n]|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))` +
    `(?:(?:\\r?\\n[ \\t]{2,5}(?:[-*+•]|\\d{1,3}\\.\\w\\.|\\[[ xX]\\])[ \\t]+(?:(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[\\r\\n]|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?)))` +
    `{0,${MAX_NESTED_LIST_ITEMS}}(?:\\r?\\n[ \\t]{4,${MAX_LIST_INDENT_SPACES}}(?:[-*+•]|\\d{1,3}\\.\\w\\.|\\[[ xX]\\])[ \\t]+(?:(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[\\r\\n]|$))|(?:\\b[^\\r\\n]{1,${MAX_LIST_ITEM_LENGTH}}\\b(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?)))` +
    `{0,${MAX_NESTED_LIST_ITEMS}})?)` +
    "|" +
    // 3. Block quotes (including nested quotes and citations, up to three levels, with length constraints)
    `(?:(?:^>(?:>|\\s{2,}){0,2}(?:(?:\\b[^\\r\\n]{0,${MAX_BLOCKQUOTE_LINE_LENGTH}}\\b(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:\\b[^\\r\\n]{0,${MAX_BLOCKQUOTE_LINE_LENGTH}}\\b(?=[\\r\\n]|$))|(?:\\b[^\\r\\n]{0,${MAX_BLOCKQUOTE_LINE_LENGTH}}\\b(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))\\r?\\n?){1,${MAX_BLOCKQUOTE_LINES}})` +
    "|" +
    // 4. Code blocks (fenced, indented, or HTML pre/code tags, with length constraints)
    `(?:(?:^|\\r?\\n)(?:\`\`\`|~~~)(?:\\w{0,${MAX_CODE_LANGUAGE_LENGTH}})?\\r?\\n[\\s\\S]{0,${MAX_CODE_BLOCK_LENGTH}}?(?:\`\`\`|~~~)\\r?\\n?` +
    `|(?:(?:^|\\r?\\n)(?: {4}|\\t)[^\\r\\n]{0,${MAX_LIST_ITEM_LENGTH}}(?:\\r?\\n(?: {4}|\\t)[^\\r\\n]{0,${MAX_LIST_ITEM_LENGTH}}){0,${MAX_INDENTED_CODE_LINES}}\\r?\\n?)` +
    `|(?:<pre>(?:<code>)?[\\s\\S]{0,${MAX_CODE_BLOCK_LENGTH}}?(?:</code>)?</pre>))` +
    "|" +
    // 5. Tables (Markdown, grid tables, and HTML tables, with length constraints)
    `(?:(?:^|\\r?\\n)(?:\\|[^\\r\\n]{0,${MAX_TABLE_CELL_LENGTH}}\\|(?:\\r?\\n\\|[-:]{1,${MAX_TABLE_CELL_LENGTH}}\\|){0,1}(?:\\r?\\n\\|[^\\r\\n]{0,${MAX_TABLE_CELL_LENGTH}}\\|){0,${MAX_TABLE_ROWS}}` +
    `|<table>[\\s\\S]{0,${MAX_HTML_TABLE_LENGTH}}?</table>))` +
    "|" +
    // 6. Horizontal rules (Markdown and HTML hr tag)
    `(?:^(?:[-*_]){${MIN_HORIZONTAL_RULE_LENGTH},}\\s*$|<hr\\s*/?>)` +
    "|" +
    // 10. Standalone lines or phrases (including single-line blocks and HTML elements, with length constraints)
    `(?:^(?:<[a-zA-Z][^>]{0,${MAX_HTML_TAG_ATTRIBUTES_LENGTH}}>)?(?:(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?:[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?=[\\r\\n]|$))|(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?=[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))(?:</[a-zA-Z]+>)?(?:\\r?\\n|$))` +
    "|" +
    // 7. Sentences or phrases ending with punctuation (including ellipsis and Unicode punctuation)
    `(?:(?:[^\\r\\n]{1,${MAX_SENTENCE_LENGTH}}(?:[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:[^\\r\\n]{1,${MAX_SENTENCE_LENGTH}}(?=[\\r\\n]|$))|(?:[^\\r\\n]{1,${MAX_SENTENCE_LENGTH}}(?=[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.\\.\\.|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))` +
    "|" +
    // 8. Quoted text, parenthetical phrases, or bracketed content (with length constraints)
    "(?:" +
    `(?<!\\w)\"\"\"[^\"]{0,${MAX_QUOTED_TEXT_LENGTH}}\"\"\"(?!\\w)` +
    `|(?<!\\w)(?:['\"\`'"])[^\\r\\n]{0,${MAX_QUOTED_TEXT_LENGTH}}\\1(?!\\w)` +
    `|\\([^\\r\\n()]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}(?:\\([^\\r\\n()]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}\\)[^\\r\\n()]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}){0,${MAX_NESTED_PARENTHESES}}\\)` +
    `|\\[[^\\r\\n\\[\\]]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}(?:\\[[^\\r\\n\\[\\]]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}\\][^\\r\\n\\[\\]]{0,${MAX_PARENTHETICAL_CONTENT_LENGTH}}){0,${MAX_NESTED_PARENTHESES}}\\]` +
    `|\\$[^\\r\\n$]{0,${MAX_MATH_INLINE_LENGTH}}\\$` +
    `|\`[^\`\\r\\n]{0,${MAX_MATH_INLINE_LENGTH}}\`` +
    ")" +
    "|" +
    // 9. Paragraphs (with length constraints)
    `(?:(?:^|\\r?\\n\\r?\\n)(?:<p>)?(?:(?:[^\\r\\n]{1,${MAX_PARAGRAPH_LENGTH}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:[^\\r\\n]{1,${MAX_PARAGRAPH_LENGTH}}(?=[\\r\\n]|$))|(?:[^\\r\\n]{1,${MAX_PARAGRAPH_LENGTH}}(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))(?:</p>)?(?=\\r?\\n\\r?\\n|$))` +
    "|" +
    // 11. HTML-like tags and their content (including self-closing tags and attributes, with length constraints)
    `(?:<[a-zA-Z][^>]{0,${MAX_HTML_TAG_ATTRIBUTES_LENGTH}}(?:>[\\s\\S]{0,${MAX_HTML_TAG_CONTENT_LENGTH}}?</[a-zA-Z]+>|\\s*/>))` +
    "|" +
    // 12. LaTeX-style math expressions (inline and block, with length constraints)
    `(?:(?:\\$\\$[\\s\\S]{0,${MAX_MATH_BLOCK_LENGTH}}?\\$\\$)|(?:\\$[^\\$\\r\\n]{0,${MAX_MATH_INLINE_LENGTH}}\\$))` +
    "|" +
    // 14. Fallback for any remaining content (with length constraints)
    `(?:(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))|(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?=[\\r\\n]|$))|(?:[^\\r\\n]{1,${MAX_STANDALONE_LINE_LENGTH}}(?=[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?:.{1,${LOOKAHEAD_RANGE}}(?:[.!?…]|\\.{3}|[\\u2026\\u2047-\\u2049]|[\\p{Emoji_Presentation}\\p{Extended_Pictographic}])(?=\\s|$))?))` +
    ")",
    "gmu"
);

// read from the arg[1] file
const testText = fs.readFileSync(process.argv[2], 'utf8');

// Function to format bytes to a human-readable string
function formatBytes(bytes) {
    if (bytes < 1024) return bytes + " bytes";
    else if (bytes < 1048576) return (bytes / 1024).toFixed(2) + " KB";
    else if (bytes < 1073741824) return (bytes / 1048576).toFixed(2) + " MB";
    else return (bytes / 1073741824).toFixed(2) + " GB";
}

// Start measuring time and memory
const startTime = process.hrtime();
const startMemory = process.memoryUsage().heapUsed;

// Apply the regex
const matches = testText.match(chunkRegex);

// End measuring time and memory
const endTime = process.hrtime(startTime);
const endMemory = process.memoryUsage().heapUsed;

// Calculate execution time and memory usage
const executionTime = endTime[0] + endTime[1] / 1e9;
const memoryUsed = endMemory - startMemory;

// Output results
console.log(`Number of chunks: ${matches ? matches.length : 0}`);
console.log(`Execution time: ${executionTime.toFixed(3)} seconds`);
console.log(`Memory used: ${formatBytes(memoryUsed)}`);

// Output the first 5 matches (or fewer if there are less than 5)
console.log('\nFirst 10 chunks:');
if (matches) {
    matches.slice(0, 100).forEach((match, index) => {
        console.log(util.inspect(match, {maxStringLength: 50}));
    });
} else {
    console.log('No chunks found.');
}

// Output regex flags
console.log(`\nRegex flags: ${chunkRegex.flags}`);

// Check for potential issues
if (executionTime > 5) {
    console.warn('\nWarning: Execution time exceeded 5 seconds. The regex might be too complex or the input too large.');
}
if (memoryUsed > 100 * 1024 * 1024) {
    console.warn('\nWarning: Memory usage exceeded 100 MB. Consider processing the input in smaller chunks.');
}

As expressões regulares nesse código levam em conta uma variedade de estruturas de texto, incluindo cabeçalhos, itens de lista, referências de bloco, blocos de código, tabelas, regras horizontais, linhas ou frases separadas, sentenças ou frases com pontuação, texto entre aspas, conteúdo de parênteses, blocos de código, tabelas, regras horizontais, linhas ou frases separadas, conteúdo de tags HTML, expressões matemáticas LaTeX e muito mais. Ele se aproxima da fragmentação do texto por meio de padrões cuidadosamente projetados, embora as expressões regulares em si não entendam o contexto ou a semântica do texto.

A expressão regular no exemplo de código usa "backtracking", o que é essencial para uma segmentação semântica mais significativa. Por exemplo, ela não é interrompida no meio de uma frase. Entretanto, para listas profundamente aninhadas, referências de bloco ou estruturas como parênteses, o retrocesso pode ser difícil. Para otimizar esses casos, as expressões regulares podem ser aprimoradas para lidar melhor com vários níveis de aninhamento e limitar o aninhamento a níveis práticos, como até 3 níveis, para garantir o desempenho e evitar retrocessos catastróficos.

Embora esse código possa não estar muito completo, mas de acordo com essa ideia de otimizar os detalhes, podemos prever que o efeito pode ser aprimorado ainda mais. Os serviços oficiais de nuvem da Jina fornecidos pela interface oficial do particípio para que os desenvolvedores experimentem o uso e são gratuitos.

versão do python

Este conteúdo foi ocultado pelo autor. Digite o código de verificação para visualizar o conteúdo

Captcha:

Preste atenção ao número público do WeChat deste site, responda "CAPTCHA, um tipo de teste de desafio-resposta (computação)", obtenha o código de verificação. Pesquise no WeChat por "Chefe do Círculo de Compartilhamento de IA"ou"Aparência-AI"ou WeChat escaneando o lado direito do código QR pode prestar atenção a esse número público do WeChat do site.