wliafe commited on
Commit
c68a5ba
·
verified ·
1 Parent(s): 4bb9954

Document the GloVe word vector dataset

Browse files

Add usage, schema, split, preprocessing, and citation documentation while preserving the existing YAML front matter (SHA-256: b599bfb34b3b82e9fb63c02b85b3a8f339052783b19f2edd38997dd1a470e124).

Files changed (1) hide show
  1. README.md +108 -0
README.md CHANGED
@@ -96,3 +96,111 @@ configs:
96
  - split: 200d
97
  path: twitter27B/200d-*
98
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
96
  - split: 200d
97
  path: twitter27B/200d-*
98
  ---
99
+
100
+
101
+ # GloVe Pre-trained Word Vectors
102
+
103
+ 该仓库将 GloVe(Global Vectors for Word Representation)预训练词向量整理为 Hugging Face Dataset。每个配置对应一套官方发布语料,每个 split 名称表示词向量维度。
104
+
105
+ ## 配置
106
+
107
+ | Config | 语料 | Split / 维度 | 词表大小 |
108
+ | --- | --- | --- | ---: |
109
+ | `6B` | Wikipedia 2014 + Gigaword 5 | `50d`, `100d`, `200d`, `300d` | 400,000 |
110
+ | `42B` | Common Crawl 42B tokens | `300d` | 1,917,494 |
111
+ | `840B` | Common Crawl 840B tokens | `300d` | 2,196,017 |
112
+ | `twitter27B` | Twitter 27B tokens | `25d`, `50d`, `100d`, `200d` | 1,193,514 |
113
+
114
+ 全部配置的下载大小合计约为 7.9 GB。只需加载所需配置和维度,无需下载其他向量。
115
+
116
+ ## 字段说明
117
+
118
+ 每行表示一个 token 及其词向量:
119
+
120
+ - `word`:原始词表中的 token。
121
+ - `vector`:对应的 `float32` 向量;长度由 split 名称决定。
122
+
123
+ 例如,`6B` 配置的 `100d` split 中,每个 `vector` 包含 100 个浮点数。
124
+
125
+ ## 加载数据
126
+
127
+ 加载 `6B` 的全部维度:
128
+
129
+ ```python
130
+ from datasets import load_dataset
131
+
132
+
133
+ dataset = load_dataset("wliafe/glove", "6B")
134
+ print(dataset)
135
+ print(dataset["50d"][0])
136
+ ```
137
+
138
+ 只加载一个维度:
139
+
140
+ ```python
141
+ from datasets import load_dataset
142
+
143
+
144
+ vectors = load_dataset(
145
+ "wliafe/glove",
146
+ "twitter27B",
147
+ split="100d",
148
+ )
149
+
150
+ print(vectors.features)
151
+ print(vectors[0]["word"])
152
+ print(len(vectors[0]["vector"]))
153
+ ```
154
+
155
+ 其他配置示例:
156
+
157
+ ```python
158
+ from datasets import load_dataset
159
+
160
+
161
+ glove_42b = load_dataset("wliafe/glove", "42B", split="300d")
162
+ glove_840b = load_dataset("wliafe/glove", "840B", split="300d")
163
+ ```
164
+
165
+ ## 查询词向量
166
+
167
+ `Dataset.filter()` 可以直接查找少量 token,但它会扫描整个 split:
168
+
169
+ ```python
170
+ from datasets import load_dataset
171
+
172
+
173
+ vectors = load_dataset("wliafe/glove", "6B", split="50d")
174
+ matches = vectors.filter(lambda row: row["word"] == "king")
175
+
176
+ if len(matches) == 0:
177
+ raise KeyError("king 不在词表中")
178
+
179
+ king_vector = matches[0]["vector"]
180
+ print(len(king_vector))
181
+ ```
182
+
183
+ 频繁查询时,建议一次性建立 token 到行号或向量的索引,并根据内存容量选择所需配置。大型配置不适合无条件转换为完整的 Python 字典。
184
+
185
+ ## 使用说明
186
+
187
+ - token 的大小写、标点和分词形式沿用原始 GloVe 文件。
188
+ - 不同配置的词表互不保证一致。
189
+ - split 名称是向量维度,不是训练集或测试集划分。
190
+ - 向量以 `float32` 保存。
191
+
192
+ ## 引用
193
+
194
+ 如果该数据集对你的研究有帮助,请引用 GloVe:
195
+
196
+ ```bibtex
197
+ @inproceedings{pennington2014glove,
198
+ title={GloVe: Global Vectors for Word Representation},
199
+ author={Pennington, Jeffrey and Socher, Richard and Manning, Christopher D.},
200
+ booktitle={Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
201
+ pages={1532--1543},
202
+ year={2014}
203
+ }
204
+ ```
205
+
206
+ 原始向量、语料说明和使用条款请以 [GloVe 官方项目页面](https://nlp.stanford.edu/projects/glove/) 为准。