A Span-Extraction Dataset for Chinese Machine Reading Comprehension

10/17/2018 ∙ by Yiming Cui, et al. ∙ Harbin Institute of Technology Anhui USTC iFLYTEK Co 0

Machine Reading Comprehension (MRC) has become enormously popular recently and has attracted a lot of attention. However, the existing reading comprehension datasets are mostly in English. In this paper, we introduce a Span-Extraction dataset for Chinese Machine Reading Comprehension to add language diversities in this area. The dataset is composed by near 20,000 real questions annotated by human on Wikipedia paragraphs. We also annotated a challenge set which contains the questions that need comprehensive understanding and multi-sentence inference throughout the context. With the release of the dataset, we hosted the Second Evaluation Workshop on Chinese Machine Reading Comprehension (CMRC 2018). We hope the release of the dataset could further accelerate the machine reading comprehension research in Chinese language. The data is available through: https://github.com/ymcui/cmrc2018

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

To read and comprehend natural languages is the key to achieve advanced artificial intelligence. Machine Reading Comprehension (MRC) aims to comprehend the context of given articles and answer the questions based on them. Various types of machine reading comprehension datasets have been proposed, such as cloze-style reading comprehension

(Hermann et al., 2015; Hill et al., 2015; Cui et al., 2016), span-extraction reading comprehension (Rajpurkar et al., 2016; Trischler et al., 2016), open-domain reading comprehension (Nguyen et al., 2016; He et al., 2017), reading comprehension with multiple-choice (Richardson et al., 2013; Lai et al., 2017)

etc. Along with the development of the reading comprehension dataset, various neural network approaches have been proposed and make a big advancement in this area

(Kadlec et al., 2016; Cui et al., 2017; Dhingra et al., 2017; Wang and Jiang, 2016; Xiong et al., 2016; Wang et al., 2017; Hu et al., 2018; Wang et al., 2018; Yu et al., 2018).

We also have seen various efforts on the construction of Chinese machine reading comprehension datasets. In cloze-style reading comprehension, Cui et al. (2016) proposed a Chinese cloze-style reading comprehension dataset: People Daily & Children’s Fairy Tale, including People Daily news datasets and Children’s Fairy Tale datasets. To add difficulties to the dataset, along with the automatically generated evaluation sets (development and test), they also release a human-annotated evaluation set. Later, Cui et al. (2018)

propose another dataset, which is gathered from children’s reading material. To add more diversity and for further investigation on transfer learning, they also provide another evaluation dataset which is also annotated by human, but the query is more natural than the cloze type. The dataset was used in the first evaluation workshop on Chinese machine reading comprehension (CMRC 2017). In open-domain reading comprehension,

He et al. (2017) propose a large-scale open-domain Chinese machine reading comprehension dataset, which contains 200k queries annotated from the user query logs on search engine.

Though we have seen that the current machine learning approaches have surpassed the human performance on the SQuAD dataset

Rajpurkar et al. (2016), we wonder if these state-of-the-art models could also give similar performance on the dataset of different languages. To further accelerate the development of the machine reading comprehension research, we propose a span-extraction dataset for Chinese machine reading comprehension.

The main contributions of our work can be concluded as follows.

  • We propose a Chinese span-extraction based reading comprehension dataset to add linguistic diversity in the machine reading comprehension field.

  • To throughly test the ability of the machine reading comprehension systems, we made a set of 500 annotated questions that require various clues in the context to answer, which is relatively difficult than the development and test set.

  • We hosted the second evaluation workshop on Chinese machine reading comprehension, and the top system still have a much bigger gap to the human performance, indicating that there are still much efforts should be made.

[Document]
静电感应是物体内的电荷因受外界电荷的影响而重新分布。这个现象由英国科学家约翰·坎通和瑞典科学家分别在1753年和1762年发现。静电发电机,例如威姆斯赫斯特电机、范德格拉夫起电机和起电盘,都使用这个原理。正常的物质都带有等量的正电荷和负电荷,因此总的来说是不带电的。如果把带电的物体靠近不带电的导体,例如一片金属,则导体上的电荷将会重新分布。例如,如果把带正电的物体靠近一块金属(参见右面的图),则金属上的负电荷将会被吸引过去,而正电荷则会被排斥。这样便导致金属的靠近外界电荷的部分带有负电荷,而远离外界电荷的部分则带有正电荷。
……
在绝缘体中,电子被原子束缚着,不能在物体中自由移动;但是在原子内可以移动一点点。如果把带正电的物体靠近绝缘体,则每一个原子中的电子都会被吸引而稍微移动一点,而原子核则会被排斥,而往相反的方向移动一点。这种现象称为极化。由于这时物体中的负电荷离外面的带电物体较近,而正电荷则距离较远,将导致吸引力比排斥力大一点点。这个现象是微观的,但因为有那么多的原子,加起来效果就很明显了,足以使较轻的物体(如小纸片)被吸引。
[Question]
静电感应是什么时候发现的?
[Answer 1]
1753年和1762年
[Answer 2]
分别在1753年和1762年发现
[Answer 3]
分别在1753年和1762年发现
Figure 1: A sample from the CMRC 2018 development set.

2 The Proposed Dataset

The Figure 1 show an example of the proposed dataset.

2.1 Task Definition

The reading comprehension can be described as a triple , where represents Document, represents Question and the represents Answer. Specifically, for span-extraction reading comprehension task, the question is annotated by the human which is much more natural than the cloze-style reading comprehension dataset (Hill et al., 2015; Cui et al., 2016). The answer should be directly extracted from the document . According to the most of the works on SQuAD dataset, the task can be simplified by predicting the start and end point in the document to get the answer span (Wang and Jiang, 2016).

2.2 Data Pre-processing

We downloaded Chinese portion of Wikipedia webpage dump on Jan 22, 2018222https://dumps.wikimedia.org/zhwiki/latest/ and use open-source toolkit Wikipedia Extractor333http://medialab.di.unipi.it/wiki/Wikipedia_Extractor for pre-processing the files into plain text. We also convert the traditional Chinese characters into simplified Chinese for normalization purpose using opencc444https://github.com/BYVoid/OpenCC toolkit.

2.3 Human Annotation

The questions in the proposed dataset are totally annotated by human which is different from previous works that relies on the automatic data generation (Hermann et al., 2015; Hill et al., 2015; Cui et al., 2016). Before annotating, the document is divided into several paragraphs and each paragraph is limited to have no more than 500 Chinese words, where the word is counted by using LTP (Che et al., 2010). Then, the annotator was instructed to first evaluate the appropriateness of the document, because some of the documents are extremely difficult for the public to understand. Following rules are applied when discarding the document.

  • The paragraph contains many non-Chinese characters, say over 30

  • The paragraph contains many professional words that hard to understand.

  • The paragraph contains many special characters (may induced by the pre-processing process).

  • The paragraph is written in classical Chinese.

After identifying the paragraph is appropriate for annotation, the annotator will read the paragraph and ask the questions based on it and annotated a primary answer. During the question annotation, the following rules are applied.

  • No more than 5 questions for each paragraph.

  • The answer MUST be a span in the paragraph to meet the task definition.

  • Encourage the question diversity in various types, such as who/when/where/why/how etc.

  • Avoid directly using the description in the paragraph. Use paraphrase or syntax transformation to add difficulties for answering.

  • Long answers ( characters) will be discarded.

For the evaluation sets (development, test, challenge), there are three answers available for better evaluation. Besides the primary answer that was annotated by the question proposer, we also invite two additional annotator to write the second and third answer for the question. During this phase, the annotator could not see the primary answer to ensure the answer was not copied from others and encourage the diversities in the answer.

[Document]
《黄色脸孔》是柯南·道尔所著的福尔摩斯探案的56个短篇故事之一,收录于《福尔摩斯回忆录》。孟罗先生素来与妻子恩爱,但自从最近邻居新入伙后,孟罗太太则变得很奇怪,曾经凌晨时份外出,又藉丈夫不在家时偷偷走到邻居家中。于是孟罗先生向福尔摩斯求助,福尔摩斯听毕孟罗先生的故事后,认为孟罗太太被来自美国的前夫勒索,所以不敢向孟罗先生说出真相,所以吩咐孟罗先生,如果太太再次走到邻居家时,即时联络他,他会第一时间赶到。孟罗太太又走到邻居家,福尔摩斯陪同孟罗先生冲入,却发现邻居家中的人是孟罗太太与前夫生的女儿,因为孟罗太太的前夫是黑人,她怕孟罗先生嫌弃混血儿,所以不敢说出真相。
[Question]
孟罗太太为什么在邻居新入伙后变得很奇怪?
[Answer 1]
邻居家中的人是孟罗太太与前夫生的女儿,因为孟罗太太的前夫是黑人,她怕孟罗先生嫌弃混血儿
[Answer 2]
邻居家中的人是孟罗太太与前夫生的女儿,因为孟罗太太的前夫是黑人,她怕孟罗先生嫌弃混血儿,所以不敢说出真相。
[Answer 3]
邻居家中的人是孟罗太太与前夫生的女儿,因为孟罗太太的前夫是黑人,她怕孟罗先生嫌弃混血儿,所以不敢说出真相。
Figure 2: A snippet of the CMRC 2018 challenge set.

2.4 Challenge Set

In order to examine how well can reading comprehension models deal with the questions that need comprehensive reasoning over various clues in the context, we additionally annotated a small challenge set for this purpose. The annotation was also done by three annotators in a similar way that for development and test set annotation. The question should meet the following standard to be qualified into this set.

  • The answer cannot be inferred by a single sentence in the context, if the answer is only a single word or short phrase.

  • If the answer belongs to a type of named entity, it can not be the only one in the context, or the machine could easily pick it out according to its named entity type. For example, if there is only one person name appears in the context, then it cannot be used for annotating questions. There should be at least two person names that could mislead the machine for answering.

The Figure 2 show an example of the challenge dataset.

2.5 Statistics

The general statistics of the pre-processed data is given in Table 1. The query type distribution of the development set is given in Figure 3.

     Train      Dev      Test   Challenge
Question # 10,321 3,351 4,895 504
Answer # per query 1 3 3 3
Max doc tokens 962 961 980 916
Max question tokens 89 56 50 47
Max answer tokens 100 85 92 77
Average doc tokens 452 469 472 464
Average question tokens 15 15 15 18
Average answer tokens 17 9 9 19
Table 1: Statistics of the proposed CMRC 2018 dataset.
Figure 3: Question type distribution of the CMRC 2018 development set.
Development Test Challenge
EM F1 EM F1 EM F1
Estimated Human Performance 91.083 97.348 92.400 97.914 90.382 95.248
Z-Reader (single model) 79.776 92.696 74.178 88.145 13.889 37.422
MCA-Reader (ensemble) 66.698 85.538 71.175 88.090 15.476 37.104
RCEN (ensemble) 76.328 91.370 68.662 85.753 15.278 34.479
MCA-Reader (single model) 63.902 82.618 68.335 85.707 13.690 33.964
OmegaOne (ensemble) 66.977 84.955 66.272 82.788 12.103 30.859
RCEN (single model) 73.253 89.750 64.576 83.136 10.516 30.994
GM-Reader (ensemble) 58.931 80.069 64.045 83.046 15.675 37.315
OmegaOne (single model) 64.430 82.699 64.188 81.539 10.119 29.716
GM-Reader (single model) 56.322 77.412 60.470 80.035 13.690 33.990
R-NET (single model) 45.418 69.825 50.112 73.353 9.921 29.324
SXU-Reader (ensemble) 40.292 66.451 46.210 70.482 N/A N/A
SXU-Reader (single model) 37.310 66.121 44.270 70.673 6.548 28.116
T-Reader (single model) 39.422 62.414 44.883 66.859 7.341 22.317
Unnamed Sys by usst (single model) 34.490 59.539 37.916 63.502 5.159 18.687
Unnamed Sys by whu (single model) 18.577 42.560 22.288 46.774 2.183 21.587
Unnamed Sys by LittleBai (single model) 7.021 31.657 10.848 37.231 0.397 9.498
Unnamed Sys by jspi (single model) 13.793 39.720 0.449 34.224 2.579 20.048
Table 2: Overall leaderboard on the CMRC 2018 dataset. The ranking is obtained by the average score of the test EM and F1 in descending order. The results marked as ‘N/A’ are due to the bundle deletion by the participants.

3 Evaluation Metrics

In this paper, we adopt two evaluation metrics following

Rajpurkar et al. (2016). However, Chinese language is fairly different from English, we adapt the original metrics in the following ways. Note that , the common punctuations, white spaces are ignored.

3.1 Exact Match

Measure the exact match between the prediction and ground truths, i.e. 1 for exact match, otherwise the score is 0.

3.2 F1-Score

Measure the character-level fuzzy match between the prediction and ground truths. Instead of treating the predictions and ground truths as bags of words, we calculate the length of the longest common sequence (LCS) between them and compute the F1-score accordingly. We take the maximum F1 over all of the ground truth answers for a given question. Note that, non-Chinese words will not be segmented.

3.3 Human Performance

We also report the human performance in order to measure the difficulty of the proposed dataset. As we have illustrated in the previous section, there are three answers for each question in development, test, and challenge set. Unlike Rajpurkar et al. (2016), we use a cross-validation method to calculate the performance. We regard the first answer as human prediction and treat rest of the answers as ground truths. In this way, we can get three human prediction performance by iteratively regarding the first, second, and third answer as human prediction. Finally, we calculate the average of three results as the final human performance on this dataset.

4 Results

4.1 CMRC 2018 Evaluation Results

The final leaderboard of the CMRC 2018 evaluation is shown in Table 2. As we can see that most of the systems could obtain over 80 in test F1. While compared to F1 metric, the EM metric is substantially lower compared to the SQuAD dataset (usually within 10 points). This suggests that how to determine the exact boundary in Chinese machine reading comprehension plays a key role to improve the system performance.

4.2 Results on the Challenge Set

Not surprisingly, as shown in the last column of Table 2, though the top-ranked systems obtain decent scores on the development and test set, they are failed to give satisfactory results on the challenge set. However, as we can see that the estimated human performance on the development, test, and challenge set are relatively the same, where the challenge set gives slightly lower scores. We also observed that though Z-Reader obtains best scores on the test set, it failed to give consistent performance on the EM metric of the challenge set. This suggests that the current reading comprehension models are not capable of handling difficult questions that need comprehensive reasoning among several clues in the context.

5 Conclusion

In this work, we propose a span-extraction dataset for Chinese machine reading comprehension, namely CMRC 2018, which is also used in the second evaluation workshop on Chinese machine reading comprehension. The dataset is annotated by human with near 20,000 questions as well as a challenging set which is composed by the questions that need reasoning over multiple clues. The evaluation results show that the machine could give excellent scores on the development and test set with only near 10 points below the estimated human performance in F1-score. However, when it comes to the challenge set, the scores are declining drastically while the human performance remains almost the same with the non-challenge set, indicating that there are still potential challenges in designing more sophisticated models to improve the performance. We hope the release of this dataset could bring language diversity in machine reading comprehension task, and accelerate further investigation on solving the questions that need comprehensive reasoning over multiple clues.

Open Challenge

We would like to invite more researchers doing experiments on CMRC 2018 datasets and evaluate on the hidden test and challenge set to further test the generalization of the models. You can follow the instructions on our CodaLab worksheet to submit your model via https://worksheets.codalab.org/worksheets/0x92a80d2fab4b4f79a2b4064f7ddca9ce/.

Acknowledgements

We would like to thank our resource team for annotating and verifying evaluation data. Also, we thank the Seventeenth China National Conference on Computational Linguistics (CCL 2018)555http://www.cips-cl.org/static/CCL2018/index.html and Changsha University of Science and Technology for providing venue for evaluation workshop. This work was supported by the National 863 Leading Technology Research Project via grant 2015AA015409.

References