Upload 6 files

.gitattributes

@@ -52,3 +52,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.jpg filter=lfs diff=lfs merge=lfs -text
 *.jpeg filter=lfs diff=lfs merge=lfs -text
 *.webp filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README-EN.md

@@ -0,0 +1,64 @@
+# LCSTS
+
+### Introduction
+LCSTS is a Large-scale Chinese Short Text Summarization dataset constructed from the Chinese microblogging website SinaWeibo for the summary generation, which is collected by Harbin Institute of Technology. This corpus consists of over 2 million real Chinese short texts with short summaries given by the writer of each text, as well as 10,666 short summaries marked manually.
+
+### Paper
+[LCSTS: A Large Scale Chinese Short Text Summarization Dataset](https://www.aclweb.org/anthology/D15-1229.pdf). EMNLP 2015.
+
+### Data Size
+Training set: 2,400,591; Validation set: 8,685; Test set: 725.
+
+### Data Format
+Each instance is composed of a human-labeled summary quality score (human_label, an integer), input text (text, a string) and a output summary (summary, an integer).
+
+### Example
+```
+{
+	"human_label": 5, 
+	"summary": "林志颖公司疑涉虚假营销无厂房无研发", 
+	"text": "日前，方舟子发文直指林志颖旗下爱碧丽推销假保健品，引起哗然。调查发现，爱碧丽没有自己的生产加工厂。其胶原蛋白饮品无核心研发，全部代工生产。号称有“逆生长”功效的爱碧丽“梦幻奇迹限量组”售价>高达1080元，实际成本仅为每瓶4元！"
+}
+```
+- "human_label" (`int`): the human-labeled summary quality score(Only the validation set and the test set have this label, and the data set only includes 3, 4, and 5 points data, not including 1, 2 points data.).
+- "text" (`str`): input text.
+- "summary"（`str`): a output summary.
+
+### Evaluation Code
+The prediction result needs to be consistent with the format of the evaluation code.
+
+Dependency packages: rouge==1.0.0, jieba=0.42.1
+```shell
+python eval.py prediction_file test_private_file
+```
+
+The evaluation metrics are rouge-1, rouge-2, rouge-l, and the output is in dictionary format.
+```she
+return {
+    "rouge-1-f": _, 
+    "rouge-1-p": _, 
+    "rouge-1-r": _, 
+    "rouge-2-f": _, 
+    "rouge-2-p": _, 
+    "rouge-2-r": _, 
+    "rouge-l-f": _, 
+    "rouge-l-p": _, 
+    "rouge-l-r": _}
+```
+
+### Author List
+Baotian Hu, Qingcai Chen, Fangze Zhu
+
+### Institutions
+Harbin Institute of Technology
+
+### Citation
+```
+@inproceedings{hu2015lcsts,
+  title={LCSTS: A Large Scale Chinese Short Text Summarization Dataset},
+  author={Hu, Baotian and Chen, Qingcai and Zhu, Fangze},
+  booktitle={Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing},
+  pages={1967--1972},
+  year={2015}
+}
+```

README-ZH.md

@@ -0,0 +1,64 @@
+# LCSTS
+
+### 简介
+LCSTS数据集是一个大规模、高质量中文短文本摘要数据集，由哈尔滨工业大学收集，用于新闻摘要生成任务。该数据集取自于新浪微博的大规模中文短文本摘要数据集，包含了200万真实的中文短文本数据和每个文本作者给出的摘要，另外还有手动标记的10666份文本摘要。
+
+### 论文
+[LCSTS: A Large Scale Chinese Short Text Summarization Dataset](https://www.aclweb.org/anthology/D15-1229.pdf). EMNLP 2015.
+
+### 数据规模
+训练集：2,400,591个摘要；验证集：8,685个摘要；测试集：725个摘要。
+
+### 数据格式描述
+每个实例包含人工标注的摘要质量打分(human_label，以整数形式存储)，输入文本(text，以字符串形式存储)和输出的摘要(summary，以字符串形式存储)。
+
+### 数据样例
+```
+{
+	"human_label": 5, 
+	"summary": "林志颖公司疑涉虚假营销无厂房无研发", 
+	"text": "日前，方舟子发文直指林志颖旗下爱碧丽推销假保健品，引起哗然。调查发现，爱碧丽没有自己的生产加工厂。其胶原蛋白饮品无核心研发，全部代工生产。号称有“逆生长”功效的爱碧丽“梦幻奇迹限量组”售价>高达1080元，实际成本仅为每瓶4元！"
+}
+```
+- "human_label" (`int`)：人工标注的摘要质量打分（只有验证集和测试集有该标注，且数据集中仅包括3、4、5分数据，而不包括1、2分的数据）。
+- "text" (`str`)：输入文本。
+- "summary"（`str`）：期待输出的摘要。
+
+### 评测代码
+预测结果需要和评测代码保持一样的格式。
+正确提交文件名：LCSTS.jsonl
+依赖：rouge==1.0.0, jieba=0.42.1
+```shell
+python eval.py prediction_file test_private_file
+```
+
+评测指标为rouge-1, rouge-2, rouge-l，输出结果为字典格式：
+```she
+return {
+    "rouge-1-f": _, 
+    "rouge-1-p": _, 
+    "rouge-1-r": _, 
+    "rouge-2-f": _, 
+    "rouge-2-p": _, 
+    "rouge-2-r": _, 
+    "rouge-l-f": _, 
+    "rouge-l-p": _, 
+    "rouge-l-r": _}
+```
+
+### 作者列表
+户保田，陈清财，祝方泽
+
+### 制作单位
+哈尔滨工业大学
+
+### 论文引用
+```
+@inproceedings{hu2015lcsts,
+  title={LCSTS: A Large Scale Chinese Short Text Summarization Dataset},
+  author={Hu, Baotian and Chen, Qingcai and Zhu, Fangze},
+  booktitle={Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing},
+  pages={1967--1972},
+  year={2015}
+}
+```

eval.py

@@ -0,0 +1,72 @@
+import json
+import argparse
+import sys
+from rouge import Rouge
+import numpy as np
+import jieba
+
+def rouge_score(data):
+    """
+    compute rouge score
+    Args:
+        data (list of dict including reference and candidate):
+    Returns:
+            res (dict of list of scores): rouge score
+    """
+    rouge_name = ["rouge-1", "rouge-2", "rouge-l"]
+    item_name = ["f", "p", "r"]
+
+    res = {}
+    for name1 in rouge_name:
+        for name2 in item_name:
+            res["%s-%s"%(name1, name2)] = []
+    for tmp_data in data:
+        origin_candidate = tmp_data['candidate']
+        origin_reference = tmp_data['reference']
+        assert isinstance(origin_candidate, str)
+        if not isinstance(origin_reference, list):
+            origin_reference = [origin_reference]
+
+        tmp_res = []
+        for r in origin_reference:
+            tmp_res.append(Rouge().get_scores(refs=r, hyps=origin_candidate)[0])
+
+        for name1 in rouge_name:
+            for name2 in item_name:
+                res["%s-%s"%(name1, name2)].append(max([tr[name1][name2] for tr in tmp_res]))
+
+    for name1 in rouge_name:
+        for name2 in item_name:
+            res["%s-%s"%(name1, name2)] = np.mean(res["%s-%s"%(name1, name2)])
+    return res
+
+def load_file(filename):
+    data = []
+    with open(filename, "r") as f:
+        for line in f.readlines():
+            data.append(json.loads(line))
+        f.close()
+    return data
+
+def proline(line):
+    return " ".join([w for w in jieba.cut("".join(line.strip().split()))])
+
+
+def compute(golden_file, pred_file, return_dict=True):
+    golden_data = load_file(golden_file)
+    pred_data = load_file(pred_file)
+
+    if len(golden_data) != len(pred_data):
+        raise RuntimeError("Wrong Predictions")
+
+    eval_data = [{"reference": proline(g["summary"]), "candidate": proline(p["summary"])} for g, p in zip(golden_data, pred_data)]
+    return rouge_score(eval_data)
+
+def main():
+    argv = sys.argv
+    print("预测结果：{}, 测试集: {}".format(argv[1], argv[2]))
+    print(compute(argv[2], argv[1]))
+
+
+if __name__ == '__main__':
+    main()

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d48f7b570423997eb995288f265359c2a608a5aa89b9c68f23ab9ac4074fb56
+size 903291188