The task of argument mining aims to detect all possible argumentative components and identify their relationships automatically. As a thriving field in natural language processing, there has been a large amount of corpus for academic study and application development in argument mining. However, the research in this area is still constrained by the inherent limitations of existing datasets. Specifically, all the publicly available datasets are relatively small in scale, and few of them provide information from other modalities to facilitate the learning process. Moreover, the statements and expressions in these corpora are usually in a compact form, which means non-adjacent clauses or text segments will always be regarded as multiple individual components, thus restricting the generalization ability of models. To this end, we collect and contribute a novel dataset AntCritic to serve as a helpful complement to this area, which consists of about 10k freeform and visually-rich financial comments and supports both argument component detection and argument relation prediction tasks. Besides, we design a simple but effective model architecture and conduct extensive experiments to provide a reliable benchmark for our proposed dataset and serve as a preliminary solution to the task of argument mining on visually-rich free-form documents.
Figure: An example of visually-rich free-form documents in our proposed AntCritic dataset.
AntCritic contains about 10k Chinese financial comments collected from an open online forum supported by Alipay, which covers the themes of fund introductions, stock market analysis, investment advice, etc. The types of authors span a wide range from ordinary users to financial practitioners, resulting in significant variance in objectivity, professionalism, and descriptiveness. You can download the splited datasets via the following links:
- train dataset: antcritic/train.1.csv
- dev dataset: antcritic/test.1.csv
- test dataset: antcritic/dev.1.csv
or here in case the above links expire;
| Dataset | Domain | #Docs | #Sents | #Claims | #Tokens | Unit | Relation? | Modal | Lang |
|---|---|---|---|---|---|---|---|---|---|
| antcritic | Financial Comments | 9994 | 214585 | 88311 | 11436977 | Segment | Yes | Text&HTML | Chinese |
Example
| field | detail | field type | example |
|---|---|---|---|
| srcs | 原始文章 | string | （1）一波行情，往大了说，都是时代的礼物，比如12年重工业见顶后，内需服务消费，代表品种消费股投资，我常说4点，社交属性、成瘾性（复购率）、产品生命力、金融属性；综合而言，为何酒是最好的品种，没有之一呢，比如社交属性，举个小例子，酒庄上最容易明白的就是关系，你说客户让我来参加一个酒局，你来不来？你肯定得来啊，这是给客户来表诚意来了，我来了，又喝多了，出丑了。我把我的面子全给放下了，这某种情况来说，就是给客户的诚意的抵押物呀。比如从低度酒到高度酒啊，可能是一个我们从熟人社会向生人社会变迁的这么一个过程，要迅速的找到这个四十度、五十度甚至六十度的东西呢，把我们的情绪顶起来。\n往后，19年缩表减税搞好资本市场，就类似美股1980，结束了漂亮50，为啥说是19年而不是20年或21年呢，觉得是疫情再放水，导致了这个周期被延后，严格点说消费周期是19年结束了。往后就制造业起来，光伏、锂电为代表的能源革命，既然也是时代礼物，那么通常三波走势，第一波先来个2-3倍，比如价格从10块干到40，然后回撤50%，然后再来一波3倍以上行情，从20干到70，然后再回落个下，然后上涨到80-100，这么完整一轮行情就结束了，龙头品种10倍。现在锂电处于第二波主升浪的末端，很多标的完成了50%回撤后的3倍以上行情了，所以我给的建议是有格局的选手，认为基本面不断刷新大家认知的，可以坚定持股，哪怕是顶部也是阶段性的或是走势复合型的，不用在意一时波动，喜欢拥抱波动的觉得不妨可以减仓。\n赣锋锂业上修业绩预告，预计上半年净利润13亿元-16亿元，同比增长730.75%-922.46%，此前预计盈利8亿元-12亿元。\n这波跟容百一起，也算给大家账户增值助力不少。\n（2）说说车载摄像头光学设备，负面的觉得摄像头这东西一直觉得没啥利润，也没啥技术含量，你说占了个认证优势吧，一般车企都要认证几家的摄像头，也不是就它一家，再说了，你汽车摄像头再多也比不上手机吧。乐观的觉得，一个汽车摄像头相当于3个手机摄像头，相关企业给自己带来的增量是明显的，另外摄像头不至于新能源车，2500w台车，每车10个，得有2.5个摄像头，跟手机也差不多了。不过甭管乐观还是负面，一个道理总是错不了的，智能万联时代，信息汲取，他需要一个入口，视觉，靠摄像头光学是最重要的来源。长期拥抱光学资产，从上游的芯片到下游的模组企业。\n（3）风电这个风电，能源行业就看运营商对折旧的容忍度，要让运营商相信可以把成本降下来，快速实现平价，但风电成本曲线与光伏不同，与产品规模效应和大型化相关，我们之前缺乏大型海工平台，然后陆地跟近海优势风力资源可开发资源不多。\n无非就是成本上，就是随着超大风机12MW以上以及漂浮式技术的出现，海上风电度电成本快速下降。\n暂时列入观察窗口！看装机能否上去，成本能否下来！\n |
| sents | 分隔句子(依据标点+html标签), 从0开始标号； | json string | "{""0"": ""（1）"", ""1"": ""一波行情，往大了说，都是时代的礼物，比如12年重工业见顶后，内需服务消费，代表品种消费股投资，我常说4点，社交属性、成瘾性（复购率）、产品生命力、金融属性；"", ""2"": ""综合而言，为何酒是最好的品种，没有之一呢，比如社交属性，举个小例子，酒庄上最容易明白的就是关系，你说客户让我来参加一个酒局，你来不来？"", ""3"": ""你肯定得来啊，这是给客户来表诚意来了，我来了，又喝多了，出丑了。"", ""4"": ""我把我的面子全给放下了，这某种情况来说，就是给客户的诚意的抵押物呀。"", ""5"": ""比如从低度酒到高度酒啊，可能是一个我们从熟人社会向生人社会变迁的这么一个过程，要迅速的找到这个四十度、五十度甚至六十度的东西呢，把我们的情绪顶起来。"", ""6"": ""往后，19年缩表减税搞好资本市场，就类似美股1980，结束了漂亮50，为啥说是19年而不是20年或21年呢，觉得是疫情再放水，导致了这个周期被延后，严格点说消费周期是19年结束了。"", ""7"": ""往后就制造业起来，光伏、锂电为代表的能源革命，既然也是时代礼物，那么通常三波走势，第一波先来个2-3倍，比如价格从10块干到40，然后回撤50%，然后再来一波3倍以上行情，从20干到70，然后再回落个下，然后上涨到80-100，这么完整一轮行情就结束了，龙头品种10倍。"", ""8"": ""现在锂电处于第二波主升浪的末端，很多标的完成了50%回撤后的3倍以上行情了，所以我给的建议是有格局的选手，认为基本面不断刷新大家认知的，可以坚定持股，哪怕是顶部也是阶段性的或是走势复合型的，不用在意一时波动，喜欢拥抱波动的觉得不妨可以减仓。"", ""9"": ""赣锋锂业上修业绩预告，预计上半年净利润13亿元-16亿元，同比增长730.75%-922.46%，此前预计盈利8亿元-12亿元。"", ""10"": ""这波跟容百一起，也算给大家账户增值助力不少。"", ""11"": ""（2）"", ""12"": ""说说车载摄像头光学设备，负面的觉得摄像头这东西一直觉得没啥利润，也没啥技术含量，你说占了个认证优势吧，一般车企都要认证几家的摄像头，也不是就它一家，再说了，你汽车摄像头再多也比不上手机吧。"", ""13"": ""乐观的觉得，一个汽车摄像头相当于3个手机摄像头，相关企业给自己带来的增量是明显的，另外摄像头不至于新能源车，2500w台车，每车10个，得有2.5个摄像头，跟手机也差不多了。"", ""14"": ""不过甭管乐观还是负面，一个道理总是错不了的，智能万联时代，信息汲取，他需要一个入口，视觉，靠摄像头光学是最重要的来源。"", ""15"": ""长期拥抱光学资产，从上游的芯片到下游的模组企业。"", ""16"": ""（3）风电"", ""17"": ""这个风电，能源行业就看运营商对折旧的容忍度，要让运营商相信可以把成本降下来，快速实现平价，但风电成本曲线与光伏不同，与产品规模效应和大型化相关，我们之前缺乏大型海工平台，然后陆地跟近海优势风力资源可开发资源不多。"", ""18"": ""无非就是成本上，就是随着超大风机12MW以上以及漂浮式技术的出现，海上风电度电成本快速下降。"", ""19"": ""暂时列入观察窗口！"", ""20"": ""看装机能否上去，成本能否下来！""}" |
| tags | a. "font-size"：字体大小，分为三类：本文最常出现的大小（0），比常见大小更大（1）， 更小（2）; b. "color"：前景颜色，只要有前景颜色即为1，否则为0； c. "background-color"：背景颜色，只要有背景颜色，即为1，否则为0； d. "sns-small-title"：是否是小标题； e. "sns-blob-tl"：是否是副标题； f. "strong"：是否加粗； g. "supertalk"：是否是话题标识符（#），是即为1，否则为0; h. "blockquote"：是否是引用语； i. "po"：段落序号； j. "pi"：段落内编号； k. "h4"：是否是四级标题； |
list of json string | "[{""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 1, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 2, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 2, ""pi"": 2}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 2, ""pi"": 3}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 2, ""pi"": 4}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 2, ""pi"": 5}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 3, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 3, ""pi"": 2}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 3, ""pi"": 3}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 4, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 5, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 6, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 7, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 7, ""pi"": 2}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 7, ""pi"": 3}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 7, ""pi"": 4}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 8, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 9, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 10, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 11, ""pi"": 1}, {""font-size"": -1, ""color"": -1, ""background-color"": -1, ""strong"": 0, ""sns-small-title"": 0, ""sns-blob-tl"": 0, ""supertalk"": 0, ""blockquote"": 0, ""h4"": 0, ""po"": 11, ""pi"": 2}]" |
| trgs | 标注结果, results字段： a. MajorClaim - 主论点; b. Claim_{i} - 第i个子论点; c.Premise_<i>_<j> - 第i个子论点的第j个子论据, 0<=i<=8, 0<=j<=4; 一个主论点，最多8个子论点，每个子论点最多4个论据； d. relations字段: 子论点和主论点的关系： 默认值是-1，表示不存在该关系，1是支持，0是反驳，2是有关，3是无关； 子论点和论据的关系：默认值是-1，表示不存在该关系，1是支持，0是反驳； |
json string | "{""results"": {""MajorClaim"": [7], ""Claim_1"": [2], ""Claim_2"": [8], ""Claim_3"": [12, 15], ""Claim_4"": [16, 19, 20], ""Claim_5"": [], ""Claim_6"": [], ""Claim_7"": [], ""Claim_8"": [], ""Premise_1_1"": [3, 4], ""Premise_1_2"": [5], ""Premise_1_3"": [], ""Premise_1_4"": [], ""Premise_2_1"": [6], ""Premise_2_2"": [], ""Premise_2_3"": [9, 10], ""Premise_2_4"": [], ""Premise_3_1"": [], ""Premise_3_2"": [13, 14], ""Premise_3_3"": [], ""Premise_3_4"": [], ""Premise_4_1"": [17], ""Premise_4_2"": [18], ""Premise_4_3"": [], ""Premise_4_4"": [], ""Premise_5_1"": [], ""Premise_5_2"": [], ""Premise_5_3"": [], ""Premise_5_4"": [], ""Premise_6_1"": [], ""Premise_6_2"": [], ""Premise_6_3"": [], ""Premise_6_4"": [], ""Premise_7_1"": [], ""Premise_7_2"": [], ""Premise_7_3"": [], ""Premise_7_4"": [], ""Premise_8_1"": [], ""Premise_8_2"": [], ""Premise_8_3"": [], ""Premise_8_4"": []}, ""relations"": {""Claim_1"": 3, ""Claim_2"": 1, ""Claim_3"": 0, ""Claim_4"": 1, ""Claim_5"": -1, ""Claim_6"": -1, ""Claim_7"": -1, ""Claim_8"": -1, ""Premise_1_1"": 1, ""Premise_1_2"": 1, ""Premise_1_3"": -1, ""Premise_1_4"": -1, ""Premise_2_1"": 1, ""Premise_2_2"": 1, ""Premise_2_3"": 1, ""Premise_2_4"": -1, ""Premise_3_1"": -1, ""Premise_3_2"": 1, ""Premise_3_3"": -1, ""Premise_3_4"": -1, ""Premise_4_1"": 1, ""Premise_4_2"": 1, ""Premise_4_3"": -1, ""Premise_4_4"": -1, ""Premise_5_1"": -1, ""Premise_5_2"": -1, ""Premise_5_3"": -1, ""Premise_5_4"": -1, ""Premise_6_1"": -1, ""Premise_6_2"": -1, ""Premise_6_3"": -1, ""Premise_6_4"": -1, ""Premise_7_1"": -1, ""Premise_7_2"": -1, ""Premise_7_3"": -1, ""Premise_7_4"": -1, ""Premise_8_1"": -1, ""Premise_8_2"": -1, ""Premise_8_3"": -1, ""Premise_8_4"": -1}, ""url"": ""https://alphaq.alipay.com/index_manage.htm#/mark?taskId=545757&subTaskId=10148412928&isPreview=1&tntInstId=caeaf603""}" |
Figure: Overall architecture diagram of segment-level model.
install packages: pip install -r requirements.txt
pretrained models:
- char: pretrained_model/paraphrase-xlm-r-multilingual-v1
- word: pretrained_model/FinBERT_L-12_H-768_A-12_pytorch
the checkpoints that fine-tuned on AntCritic dataset:
- char: checkpoints/char/models-9.pt
- word: checkpoints/word/models-12.pt
the model(Figure 4 in paper) trained on AntCritic dataset:
- checkpoints/GRU/models-7.pt
you can download pretrained models here and checkpoints here;
-
preprocess data to get the *_1.hdf5 format for the FirstStageModel or you can download here:
python preprocess/new_preprocess.py -
train FirstStageModel: change _c.model.name in config/config.py to 'FirstStageModel' (you may also need to specify your "saved_path")
python first_main.py --config use_word -
preprocess data to get the *_2.hdf5 format for the SecondStageModel (if you don't want to fine-tune char_model and word_model, simply use checkpoints/char/models-9.pt & checkpoints/word/models-12.pt) or you can download here:
python preprocess/second_stage.py -
train SecondStageModel: change _c.model.name in config/config.py to 'SecondStageModel' (you may also need to specify your "saved_path")
python second_main.py --config use_gru
If you find our work useful, please consider citing:
@article{Zhao2022AntCriticAM,
title={AntCritic: Argument Mining for Free-Form and Visually-Rich Financial Comments},
author={Yang Zhao and Wenqiang Xu and Xuan Lin and Jingjing Huo and Hong Chen and Zhou Zhao},
journal={ArXiv},
year={2022},
volume={abs/2208.09612}
}