转变器之前的文字生成 N-gram语言模型

from collections import Counter, defaultdict


def train_ngram(corpus_tokens, n=3):
    ngrams = Counter()
    contexts = Counter()
    for sentence in corpus_tokens:
        padded = ["<s>"] * (n - 1) + sentence + ["</s>"]
        for i in range(len(padded) - n + 1):
            ctx = tuple(padded[i:i + n - 1])
            word = padded[i + n - 1]
            ngrams[ctx + (word,)] += 1
            contexts[ctx] += 1
    return ngrams, contexts


def raw_probability(ngrams, contexts, context, word):
    ctx = tuple(context)
    if contexts.get(ctx, 0) == 0:
        return 0.0
    return ngrams.get(ctx + (word,), 0) / contexts[ctx]

def laplace_probability(ngrams, contexts, vocab_size, context, word):
    ctx = tuple(context)
    numerator = ngrams.get(ctx + (word,), 0) + 1
    denominator = contexts.get(ctx, 0) + vocab_size
    return numerator / denominator

def kneser_ney_bigram_model(corpus_tokens, discount=0.75):
    unigrams = Counter()
    bigrams = Counter()
    unigram_contexts = defaultdict(set)

    for sentence in corpus_tokens:
        padded = ["<s>"] + sentence + ["</s>"]
        for i, w in enumerate(padded):
            unigrams[w] += 1
            if i > 0:
                prev = padded[i - 1]
                bigrams[(prev, w)] += 1
                unigram_contexts[w].add(prev)

    total_unique_bigrams = sum(len(ctx_set) for ctx_set in unigram_contexts.values())
    continuation_prob = {
        w: len(ctx_set) / total_unique_bigrams for w, ctx_set in unigram_contexts.items()
    }

    context_totals = Counter()
    for (prev, w), count in bigrams.items():
        context_totals[prev] += count

    unique_follow = defaultdict(set)
    for (prev, w) in bigrams:
        unique_follow[prev].add(w)

    def prob(prev, w):
        count = bigrams.get((prev, w), 0)
        denom = context_totals.get(prev, 0)
        if denom == 0:
            return continuation_prob.get(w, 1e-9)
        first_term = max(count - discount, 0) / denom
        lambda_prev = discount * len(unique_follow[prev]) / denom
        return first_term + lambda_prev * continuation_prob.get(w, 1e-9)

    return prob

import random


def generate(prob_fn, vocab, prefix, max_len=30, seed=0):
    rng = random.Random(seed)
    tokens = list(prefix)
    for _ in range(max_len):
        candidates = [(w, prob_fn(tokens[-1], w)) for w in vocab]
        total = sum(p for _, p in candidates)
        r = rng.random() * total
        acc = 0.0
        for w, p in candidates:
            acc += p
            if r <= acc:
                tokens.append(w)
                break
        if tokens[-1] == "</s>":
            break
    return tokens

import math


def perplexity(prob_fn, sentences):
    total_log_prob = 0.0
    total_tokens = 0
    for sentence in sentences:
        padded = ["<s>"] + sentence + ["</s>"]
        for i in range(1, len(padded)):
            p = prob_fn(padded[i - 1], padded[i])
            total_log_prob += math.log(max(p, 1e-12))
            total_tokens += 1
    return math.exp(-total_log_prob / total_tokens)

---
name: lm-baseline
description: Build a reproducible n-gram language model baseline before training a neural LM.
phase: 5
lesson: 16
---

Given a corpus and target use (next-word prediction, rescoring, perplexity baseline), output:

1. N-gram order. Trigram for general English, 4-gram if corpus is large, 5-gram for speech rescoring.
2. Smoothing. Modified Kneser-Ney is the default; Laplace only for teaching.
3. Library. `kenlm` for production, `nltk.lm` for teaching, roll your own only to learn.
4. Evaluation. Held-out perplexity with consistent tokenization between train and test sets.

Refuse to report perplexity computed with different tokenization between systems being compared — perplexity numbers are comparable only under identical tokenization. Flag OOV rate in test set; KN handles OOV poorly unless you reserve a special <UNK> token during training.