标签和语法解析

The/DET cats/NOUN were/AUX running/VERB at/ADP 3pm/NOUN ./PUNCT

running is ROOT
cats is nsubj of running
were is aux of running
at is prep of running
3pm is pobj of at

from collections import Counter, defaultdict


def train_mft(train_examples):
    word_tag_counts = defaultdict(Counter)
    all_tags = Counter()
    for tokens, tags in train_examples:
        for token, tag in zip(tokens, tags):
            word_tag_counts[token.lower()][tag] += 1
            all_tags[tag] += 1
    word_best = {w: c.most_common(1)[0][0] for w, c in word_tag_counts.items()}
    default_tag = all_tags.most_common(1)[0][0]
    return word_best, default_tag


def predict_mft(tokens, word_best, default_tag):
    return [word_best.get(t.lower(), default_tag) for t in tokens]

import math


def train_hmm(train_examples, alpha=0.01):
    transitions = defaultdict(Counter)
    emissions = defaultdict(Counter)
    tags = set()
    vocab = set()

    for tokens, ts in train_examples:
        prev = "<BOS>"
        for token, tag in zip(tokens, ts):
            transitions[prev][tag] += 1
            emissions[tag][token.lower()] += 1
            tags.add(tag)
            vocab.add(token.lower())
            prev = tag
        transitions[prev]["<EOS>"] += 1

    return transitions, emissions, tags, vocab


def log_prob(table, given, key, smooth_denom, alpha):
    return math.log((table[given].get(key, 0) + alpha) / smooth_denom)


def viterbi(tokens, transitions, emissions, tags, vocab, alpha=0.01):
    tags_list = list(tags)
    n = len(tokens)
    V = [[0.0] * len(tags_list) for _ in range(n)]
    back = [[0] * len(tags_list) for _ in range(n)]

    for j, tag in enumerate(tags_list):
        em_denom = sum(emissions[tag].values()) + alpha * (len(vocab) + 1)
        tr_denom = sum(transitions["<BOS>"].values()) + alpha * (len(tags_list) + 1)
        tr = log_prob(transitions, "<BOS>", tag, tr_denom, alpha)
        em = log_prob(emissions, tag, tokens[0].lower(), em_denom, alpha)
        V[0][j] = tr + em
        back[0][j] = 0

    for i in range(1, n):
        for j, tag in enumerate(tags_list):
            em_denom = sum(emissions[tag].values()) + alpha * (len(vocab) + 1)
            em = log_prob(emissions, tag, tokens[i].lower(), em_denom, alpha)
            best_prev = 0
            best_score = -1e30
            for k, prev_tag in enumerate(tags_list):
                tr_denom = sum(transitions[prev_tag].values()) + alpha * (len(tags_list) + 1)
                tr = log_prob(transitions, prev_tag, tag, tr_denom, alpha)
                score = V[i - 1][k] + tr + em
                if score > best_score:
                    best_score = score
                    best_prev = k
            V[i][j] = best_score
            back[i][j] = best_prev

    last_best = max(range(len(tags_list)), key=lambda j: V[n - 1][j])
    path = [last_best]
    for i in range(n - 1, 0, -1):
        path.append(back[i][path[-1]])
    return [tags_list[j] for j in reversed(path)]

import spacy

nlp = spacy.load("en_core_web_sm")
doc = nlp("The cats were running at 3pm.")
for token in doc:
    print(f"{token.text:10s} tag={token.tag_:5s} pos={token.pos_:6s} dep={token.dep_:10s} head={token.head.text}")

The        tag=DT    pos=DET    dep=det        head=cats
cats       tag=NNS   pos=NOUN   dep=nsubj      head=running
were       tag=VBD   pos=AUX    dep=aux        head=running
running    tag=VBG   pos=VERB   dep=ROOT       head=running
at         tag=IN    pos=ADP    dep=prep       head=running
3pm        tag=NN    pos=NOUN   dep=pobj       head=at
.          tag=.     pos=PUNCT  dep=punct      head=running

---
name: grammar-pipeline
description: Design a classical POS + dependency pipeline for a downstream NLP task.
version: 1.0.0
phase: 5
lesson: 07
tags: [nlp, pos, parsing]
---

Given a downstream task (information extraction, rewrite validation, query decomposition, lemmatization), you output:

1. Tagset to use. Penn Treebank for English-only legacy pipelines, Universal Dependencies for multilingual or cross-lingual.
2. Library. spaCy for most production, stanza for academic-grade multilingual, trankit for highest UD accuracy. Name the specific model ID.
3. Integration pattern. Show the 3-5 lines that call the library and consume the needed attributes (`.pos_`, `.dep_`, `.head`).
4. Failure mode to test. Noun-verb ambiguity (`saw`, `book`, `can`) and PP-attachment ambiguity are the classical traps. Sample 20 outputs and eyeball.

Refuse to recommend rolling your own parser. Building parsers from scratch is a research project, not an application task. Flag any pipeline that consumes POS tags without handling lowercase/uppercase variants as fragile.