Datawhale AI夏令营 基于术语词典干预的机器翻译挑战赛

题目

题目链接点这里

赛事背景

目前神经机器翻译技术已经取得了很大的突破，但在特定领域或行业中，由于机器翻译难以保证术语的一致性，导致翻译效果还不够理想。对于术语名词、人名地名等机器翻译不准确的结果，可以通过术语词典进行纠正，避免了混淆或歧义，最大限度提高翻译质量。

赛事任务

基于术语词典干预的机器翻译挑战赛选择以英文为源语言，中文为目标语言的机器翻译。本次大赛除英文到中文的双语数据，还提供英中对照的术语词典。参赛队伍需要基于提供的训练数据样本从多语言机器翻译模型的构建与训练，并基于测试集以及术语词典，提供最终的翻译结果，数据包括：

• 训练集：双语数据：中英14万余双语句对

• 开发集：英中1000双语句对

• 测试集：英中1000双语句对

• 术语词典：英中2226条

评审规则

数据说明

所有文件均为UTF-8编码，其中测评官方发放的训练集、开发集、测试集和术语词典皆为文本文件，格式如下所示：

训练集为双语数据，每行为一个句对样本，其格式如图1所示。

术语词典格式如图2所示。

评估指标

对于参赛队伍提交的测试集翻译结果文件，采用自动评价指标BLUE-4进行评价，具体工具使用sacrebleu开源版本。

baseline

baseline 代码

代码如下：

import torchtext
torchtext.disable_torchtext_deprecation_warning() # 忽略警告

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torchtext.data.utils import get_tokenizer
from collections import Counter
import random
from torch.utils.data import Subset, DataLoader
import time

# 定义数据集类
# 修改TranslationDataset类以处理术语
class TranslationDataset(Dataset):
    def __init__(self, filename, terminology):
        self.data = []
        with open(filename, 'r', encoding='utf-8') as f:
            for line in f:
                en, zh = line.strip().split('\t')
                self.data.append((en, zh))
        
        self.terminology = terminology
        
        # 创建词汇表，注意这里需要确保术语词典中的词也被包含在词汇表中
        self.en_tokenizer = get_tokenizer('basic_english')
        self.zh_tokenizer = list  # 使用字符级分词
        
        en_vocab = Counter(self.terminology.keys())  # 确保术语在词汇表中
        zh_vocab = Counter()
        
        for en, zh in self.data:
            en_vocab.update(self.en_tokenizer(en))
            zh_vocab.update(self.zh_tokenizer(zh))
        
        # 添加术语到词汇表
        self.en_vocab = ['<pad>', '<sos>', '<eos>'] + list(self.terminology.keys()) + [word for word, _ in en_vocab.most_common(10000)]
        self.zh_vocab = ['<pad>', '<sos>', '<eos>'] + [word for word, _ in zh_vocab.most_common(10000)]
        
        self.en_word2idx = {
   word: idx for idx, word in enumerate(self.en_vocab)}
        self.zh_word2idx = {
   word: idx for idx, word in enumerate(self.zh_vocab)}


    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        en, zh = self.data[idx]
        en_tensor = torch.tensor([self.en_word2idx.get(word, self.en_word2idx['<sos>']) for word in self.en_tokenizer(en)] + [self.en_word2idx['<eos>']])
        zh_tensor = torch.tensor([self.zh_word2idx.get(word, self.zh_word2idx['<sos>']) for word in self.zh_tokenizer(zh)] + [self.zh_word2idx['<eos>']])
        return en_tensor, zh_tensor

def collate_fn(batch):
    en_batch, zh_batch = [], []
    for en_item, zh_item in batch:
        en_batch.append(en_item)
        zh_batch.append(zh_item)
    
    # 对英文和中文序列分别进行填充
    en_batch = nn.utils.rnn.pad_sequence(en_batch, padding_value=0, batch_first=True)
    zh_batch = nn.utils.rnn.pad_sequence(zh_batch, padding_value=0, batch_first=True)
    
    return en_batch, zh_batch

class Encoder(nn.Module):
    def __init__(self, input_dim, emb_dim, hid_dim, n_layers, dropout):
        super().__init__()
        self.embedding = nn.Embedding(input_dim, emb_dim)
        self.rnn = nn.GRU(emb_dim, hid_dim, n_layers, dropout=dropout, batch_first=True)
        self.dropout = nn.Dropout(dropout)

    def forward(self, src):
        # src shape: [batch_size, src_len]
        embedded = self.dropout(self.embedding(src))
        # embedded shape: [batch_size, src_len, emb_dim]
        outputs, hidden = self.rnn(embedded)
        # outputs shape: [batch_size, src_len, hid_dim]
        # hidden shape: [n_layers, batch_size, hid_dim]
        return outputs, hidden

class Decoder(nn.Module):
    def __init__(self, output_dim, emb_dim, hid_dim, n_layers, dropout):
        super().__init__()
        self.output_dim = output_dim
        self.embedding = nn.Embedding(output_dim, emb_dim)
        self.rnn = nn.GRU(emb_dim, hid_dim, n_layers, dropout=dropout, batch_first=True)
        self.fc_out = nn.Linear(hid_dim, output_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, input, hidden):
        # input shape: [batch_size, 1]
        # hidden shape: [n_layers, batch_size, hid_dim]

        embedded = self.dropout(self.embedding(input))
        # embedded shape: [batch_size, 1, emb_dim]

        output, hidden = self.rnn(embedded, hidden)
        # output shape: [batch_size, 1, hid_dim]
        # hidden shape: [n_layers, batch_size, hid_dim]

        prediction = self.fc_out(output.squeeze(1))
        # prediction shape: [batch_size, output_dim]

        return prediction, hidden

class Seq2Seq(nn.Module):
    def __init__(self, encoder, decoder, device):
        super().__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.device = device

    def forward(self, src, trg, teacher_forcing_ratio=0.5):
        # src shape: [batch_size, src_len]
        # trg shape: [batch_size, trg_len]

        batch_size = src.shape[0]
        trg_len = trg.shape[1]
        trg_vocab_size = self.decoder.output_dim

        outputs = torch.zeros(batch_size, trg_len, trg_vocab_size).to(self.device)

        _, hidden = self.encoder(src)

        input = trg[:, 0].unsqueeze(1)  # Start token

        for t in range(1, trg_len):
            output, hidden = self.decoder(input, hidden)
            outputs[:, t, :] = output
            teacher_force = random.random() < teacher_forcing_ratio
            top1 = output.argmax(1)
            input = trg[:, t].unsqueeze(1) if teacher_force else top1.unsqueeze(1)

        return outputs

# 新增术语词典加载部分
def load_terminology_dictionary(dict_file):
    terminology = {
   }
    with open(dict_file, 'r', encoding='utf-8') as f:
        for line in f:
            en_term, ch_term = line.strip().split('\t')
            terminology[en_term] = ch_term
    return terminology

def train(model, iterator, optimizer, criterion, clip):
    model.train()
    epoch_loss = 0
    for i, (src, trg)