1 Research on the readability of disclosure texts of listed companies

2 Research on the tone and tone of disclosure texts of listed companies

3 Research on similarity of disclosure texts of listed companies

4 Research on the semantic features of listed companies’ disclosure texts

Media reports also contain information on economic policies. Baker et al (2016) used text mining technology to construct an economic policy uncertainty index (economic policy uncertainty, EPU) based on representative media reports in multiple major economies. , this index can continuously and quantitatively describe economic policy uncertainty. Subsequently, Gulen & Ion (2016) used this index to study the impact of economic policy uncertainty on corporate investment and found that macroeconomic policy uncertainty will affect the financial decisions of micro-enterprises and inhibit corporate investment. Bonaime et al (2018) studied the impact of economic policy uncertainty on corporate M&A based on this index and found that an increase in economic policy uncertainty will reduce the value and quantity of M&A transactions

There is also a local bias in financial media reports. Gurun & Bueler (2012) pointed out that the media use fewer negative words when reporting local company news. The reason for this phenomenon is that local companies invest more advertising expenditures. They also find that abnormal local media bias is strongly associated with corporate stock valuations. In addition, there are also company rumors and fake news in the market. Ahern & Sosyura (2015) studied rumors about corporate mergers and acquisitions and found that the media is more inclined to publish rumors about newsworthy companies. Rumors published by the media will compete for investors’ limited attention, leading to overreaction and subsequent reversal of stock prices. . Kogan et al (2019) found that false news will increase the trading volume and price volatility of the stocks involved. When the false news is exposed, the impact of all news on the platform where the false news is located on the stock trading volume and price volatility will be reduced.

1 Research on text sentiment in social networks. In the early days, Antweiler & Frank (2004) used posts in the Yahoo Finance online forum as the research object and found that the attention index measured by the number of posts can effectively predict stock returns and market volatility, and the sentiment divergence of posts is positively related to stock trading volume in the same period. A study based on articles and comments on the stock forum Seeking Alpha found that the tone of articles and comments can predict a company's future stock returns. Cookson & Niessner (2020) studied posts and user information on the US stock forum StockTwits and found that about half of investor disagreements are caused by different investment philosophies, and investor disagreements can effectively predict abnormal trading volumes in the stock market.

2 Research on strategic information disclosure Companies can also strategically release information through social network platforms. For example, Blankespoor et al (2014) found that links to news published by companies on Twitter can reduce the bid-ask spread of the company's stocks and increase trading depth. Research also shows that social networks, as a channel for information dissemination, can optimize investors' ability to obtain information and reduce investors' information search costs. Jung et al (2018) analyzed the information release on Twitter of S&P 1500 companies and found that companies will reduce the number of Twitter posts when there is bad news. For companies with less mature investors and social media audiences, For companies with a wider coverage, this strategic information disclosure behavior is more obvious

Some scholars have conducted research on other texts. For example, De Franco et al (2015) used analyst reports from 2002 to 2009 and found that more readable analyst reports can increase stock trading volume. Hwang & Kim (2017) analyzed the impact of the readability of annual shareholder reports disclosed by closed-end investment companies on company value. The study found that in a relatively opaque information environment, when a company publishes an annual shareholder report with poor readability, When reporting, investors will have negative emotions such as skepticism, causing the company to trade at a discount, thereby reducing the company's value. Greenetal (2019) used employees' ratings of employers on Glassdoor to conduct research and found that employees' evaluations of employers are related to the company's sales growth and profitability, and can predict unexpected earnings one quarter later. Huang (2018) found that consumer product reviews on Amazon (Amazon.com) can predict company stock returns. Chen et al (2019) studied the value impact of financial technology innovation on the financial industry based on U.S. patent text data. Ryans (2019) used the Naive Bayes text classification method and combined the company's future financial restatements and asset impairments to divide inquiry letters into restatement inquiry letters, non-restatement inquiry letters, and impairment inquiries. and non-impairment inquiry letters, and studied the correlation between inquiry letters and financial reporting quality. Ban-dieraetal (2020) used CEO logs as research text, divided CEOs into "leaders" and "managers" through machine learning algorithms, and studied the relationship between CEO behavior and corporate performance.

1 Text readability

2 text sentiment

3Semantic relevance

4Text similarity

summary

1 word cloud

2 bag-of-words model

3 word embeddings

4 topic model

Under the information disclosure regulatory system, companies are required to publicly release electronic documents regularly or irregularly. Documentation of related information. However, these documents are only stored electronically for readers to read on electronic devices, but this does not mean that machines can automatically process them and achieve "machine readability". In the computer field, electronic documents for information disclosure are collectively referred to as rich-format documents. These documents contain text paragraphs, tables, charts and other content modalities. They are usually organized into a hierarchical directory structure and are presented to readers through beautified typesetting and formatting. Judging from the format of the document, the vast majority of information disclosure documents required by the financial market are in PDF format. Therefore, parsing rich-format documents is often the first step in text preprocessing, that is, obtaining the information content inside. There are two aspects that need to be paid attention to in the process of PDF document parsing: On the one hand, the generation of PDF documents is not a reversible process. When we use the editor of Word or Excel to export the document into a PDF document, although the formatting format of the document, etc. Visual presentation is maintained, but structural information within the document is partially or completely lost. On the other hand, the parsed document is the basis for text analysis, and inaccurate PDF parsing may have a serious impact on subsequent text analysis. Therefore, for financial text big data, it is necessary to carefully select document structure parsing tools.

On the one hand, researchers need to use computer programs to locate text information. For example, The MD&A part is the research object of many scholars. Researchers can use regular expressions to locate MD&A in the text of financial reports. The beginning and end of the part, and then extract the content of the part. On the other hand, researchers also need to analyze what is considered noise in the text. Clean and delete content (Jiang et al, 2019), mainly including advertisements, hypertext markup language (HTML), and literal scripting language (JavaScript) and other codes and pictures, etc.

In English text, word segmentation is automatically completed when words are separated by spaces. In addition, words can be further divided through lemmatization and stemming. However, there are no spaces between Chinese characters, and words are the smallest language units that can be used independently. Therefore, researchers need to perform specialized word segmentation processing of Chinese texts. At present, most scholars use the P open source "jieba" Chinese word segmentation module to segment corporate financial reports, annual performance briefings, and stock forum posts. There are three difficulties in Chinese text word segmentation, namely segmentation granularity, identification of ambiguous words and identification of new words. If the segmentation granularity is too small, it will easily destroy the meaning of the words. For example, it is easy to split "machine learning" into "machine" and "learning". For ambiguous words, an appropriate word segmentation mode should be selected. For example, when using "jieba" to segment the word, in order to improve the accuracy of the segmentation, you should select Precise word segmentation mode. For new words (such as company names, product names, and key people's names), users can customize dictionaries to help Word segmentation software identifies new words

Part-of-speech is an important grammatical feature for identifying semantic information, such as nouns, verbs, connectives, etc. Part-of-speech tagging is to mark the part-of-speech of words after segmentation. Through part-of-speech tagging, computers can identify word types, eliminate word ambiguities, and then identify grammatical structures, reducing the difficulty of computer semantic analysis. There are big differences in part-of-speech tagging between Chinese and English. English words are more rigorous in part-of-speech classification. Changes in part-of-speech can be revealed through word-suffix changes. For example, "-ing", "-ness" and "-ment" are all useful for confirming part-of-speech. Specific tips were given. However, Chinese words do not have clear specifications for part-of-speech, and they mainly rely on grammar and semantics to identify part-of-speech, that is, "English emphasizes form combination, and Chinese emphasizes meaning combination."

In order to improve the accuracy of text mining information, stop words in the text also need to be eliminated. Stop words refer to words that are important to the grammatical structure of a sentence but convey little meaning by themselves. They increase the dimensionality of text data and increase the cost of text analysis. In English texts, stop words mainly include articles (the a), conjunctions (and or), and the verb "to be" (Gentzkow et al, 2019). However, in Chinese texts, stop words should be determined according to Chinese language habits. In addition to punctuation marks and special symbols, they also include connectives that express logical relationships (and, however, etc.) and slang. In addition, stop words need to be determined based on the content of the study. For example, when studying the emotion of a text, retaining modal particles and specific punctuation marks will help measure the emotional level of the text.

The steps and methods of text big data mining are introduced in detail. Describes the text corpus acquisition, preprocessing process, document representation and document feature extraction

Secondly, it introduces the domestic and foreign financial literature This big data mines the main text information sources, and sorts out the domestic financial text big data based on different text information sources. External research progress in order to grasp the current research directions and key areas of text big data in the field of finance

The research content and information sources of text big data in the field of finance can be further refined and enriched. For example, in terms of financial media reports, researchers can not only analyze the quantity and sentiment of media reports, but also judge the types of reported events, thereby identifying media related to corporate mergers and acquisitions, IPOs, financial fraud, personal news of executives, etc. Reports, studying the impact of different event types on companies and the stock market. In terms of stock forums, researchers can build corporate networks based on investors' attention to companies and study the competitive relationships between different companies. In terms of data sources, more text big data can be developed. For example, WeChat public accounts, government work reports, State Council policy documents, court rulings, recruitment websites, performance revision announcements issued by enterprises, social responsibility assurance opinions, internal control evaluation reports, etc.

At present, in text analysis research in the field of finance, the "bag of words" method that cannot reflect contextual meaning is still widely used. However, there are still many new analysis methods and tools in the field of natural language processing (NLP). They have not yet received enough attention in text analysis in the field of finance, but they have great potential. For example: (1) Named entity recognition (NER). NER is an important basic tool in the field of NLP. It can identify named entities in the text to be processed, thereby extracting time, place, name, institution, currency, percentage and date. Commonly used NER tools include StanfordNER. (2) Relationship extraction. Supervised machine learning methods are usually used to extract the correspondence between entity pairs from sentences containing entity pairs and analyze their co-occurrence. (3) Text summarization, which is the process of compressing text content using computer algorithms. The length of the summary depends on the compression rate. Cardinaels et al (2018) study pointed out that summaries based on algorithms are less positive than management disclosure summaries, and algorithm summaries can enable investors to make more conservative estimates of corporate stock prices.

Deep learning methods have developed rapidly in the field of NLP. Deep learning models mainly include: convolutional neural networks (CNN) model, recurrent neural network (re-current neural networks, RNN) model and its variant long short term memory network (LSTM) model , Generative Adversarial Network (GAN), reinforcement learning, and models such as BERT and XL-Net that are currently popular in the field of NLP. Introducing deep learning into the text field will expand richer research content and improve the accuracy of text information extraction. For example, Caoetal (2018) used a variant of the LSTM model to detect inconsistency errors in corporate disclosure texts. There are major differences between deep learning and traditional machine learning methods in terms of feature representation and number of model parameters, as shown in Table 1. Heatonetal (2016) pointed out that deep learning methods have the following advantages in research in the financial field: first, the model considers as much data information related to the prediction problem as possible; second, it can capture the non-linear relationship between input data and improve The degree of fitting within the sample; third, it can effectively avoid the over-fitting problem of shallow structures. In addition, when the amount of training set data increases to a certain level, the accuracy of deep learning information extraction is significantly higher than that of traditional machine learning methods.

The dictionary method is based on a preset dictionary to calculate the word frequency of different types of words in the text, and combines it with appropriate weighting methods to extract text information. However, in terms of sentiment classification of Chinese texts, the application of lexicon method is still in the exploratory stage. Most scholars choose existing English emotion dictionaries and lexicon as references to construct Chinese text emotion dictionaries, which leads to the problem that the constructed dictionaries lack the Chinese context. In addition, text information from different sources and categories has different characteristics in terms of language use. For example, there are many professional terms in corporate annual reports, and there are many slangs and emoticons used in social network media (Yao Jiajian et al., 2019). Therefore, a targeted Chinese text emotion dictionary should be constructed for different Chinese text contents, and the dictionary content should be continuously verified and updated in future research.

Regarding the annual reports of listed companies, Qiu Xinying et al. (2016) studied the annual reports of listed companies. Examining the relationship between report readability and analysts’ information interpretation, it was found that there is no significant relationship between annual report readability and analyst forecast quality, indicating that Chinese analysts fail to effectively play the role of professional interpretation of information. Zeng Qingsheng et al. (2018) studied the relationship between the tone of annual reports and insider trading behavior after annual report disclosure, and found that corporate managers had a phenomenon of "duplicity" when preparing annual reports, but a positive tone of annual reports was accompanied by higher managers' Stock selling size. Wang Xiongyuan et al. (2018) conducted a study based on the risk-related information in the annual report MD&A and found that the similarity in the content of the risk paragraphs in the two previous annual reports can reduce audit fees. Regarding other texts disclosed by the company, Lin Le and Xie Deren (2016) pointed out that investors can identify the tone of management in the annual performance briefings of listed companies. Yanetal (2019) found that uncertainty or negative tone in IPO prospectuses is significantly related to the volatility of IPO initial stock returns and subsequent stock returns, and reduces the long-term stock returns.

The China Securities Regulatory Commission stipulates that listed companies must be listed in "Shanghai Securities News", "China Securities News", "Securities Times", "Financial Times", "Economic Daily", "China Reform News", "China Daily", "Securities Market Weekly" and "seven newspapers and one publication". Publish important corporate information. In addition, my country also has online news media such as Baidu News, Sina Finance, and Hexun.com, which provide a wealth of media reporting information for studying the Chinese stock market. You Jiaxing and Wu Jing (2012) conducted a study using financial newspapers as research texts and found that the higher or lower the media sentiment is, the more serious the asset pricing bias will be. Wang Changyun and Wu Jiawei (2015) conducted research on financial media reports and found that a decrease in the negative tone of the media will increase the IPO underpricing rate, the proportion of IPO over-raised funds, and the proportion of underwriters' fees. Wang Jingyi and Huang Yiping (2018) also studied the impact of online media sentiment on the online loan market. In addition, media reports can be divided into market-oriented media reports and policy-oriented media reports. Piotroskietal (2017) pointed out that the conglomeration reform of Chinese media has made policy-oriented media reports more focused on political goals, and market-oriented media reports more focused on commercial goals. Youtal (2018) studied from the perspective of information supervision and found that compared with policy-oriented media reports, market-oriented media reports can provide more information about enterprises, and only market-oriented media reports can have a significant impact on corporate governance. In addition, in terms of measuring macroeconomic policy uncertainty, Huang & Luke (2020) used multiple newspapers in mainland China to construct a new and more frequent China EPU index. The study found that the new China EPU index can predict China's stock prices, employment and output.

He Xianjie et al. (2016) pointed out based on research on Sina Weibo blog posts that the higher the level of corporate governance, the more likely companies are to open Weibo and publish more company information. Stock forums such as Oriental Fortune Stock Bar and Snowball.com provide opportunities to study investor attention and investor sentiment in my country's stock market. Based on information posted on the Oriental Fortune Stock Forum, Huangetal (2016) found that Chinese investors also have a "local bias" phenomenon. This bias affects stocks in underdeveloped areas, large companies, non-CSI 300 index, low trading volume and names indicating the location of the company. Especially obvious in . Sun Shuna and Sun Qian (2018) found that investor attention based on the self-selected stock information of "Snowball Network" users will increase stock prices and stock trading volumes in the short term, but the impact will gradually decay over time. In addition, Jiang, Liu & Yang (2019) showed that communication among investors in stock forums will also have an impact on stock returns.

In terms of research on search indexes, some scholars have constructed investor stock focus indicators based on the Internet search index and studied the relationship between investor attention and asset pricing. For example, Yu Qingjin and Zhang Bing (2012) used the Baidu Index as an indicator of investor attention and found that investor attention has positive price pressure on current stock returns, but this pressure will reverse in the short term. There are also scholars who use the search index of specific keywords to conduct research. For example, Zeng Jianguang (2015) constructed an investor network security risk perception index based on the Baidu search index of "Yue Bao stolen". The study found that the stronger the investor's perception of Internet security risks, the higher the risk compensation required, and the mobile Internet Investors' risk perception is stronger than that of computer investors.

In the context of my country's financial reform and financial innovation, P2P online lending (peer to peer leading) has set off a new craze, and some scholars have conducted research on the factors affecting the success rate of P2P online lending. Chen Xiao et al. (2018) found that readable loan descriptions can convey positive information to investors and improve the success rate of borrowing. Peng Hongfeng and Lin Chuan (2018) analyzed the impact of the proportion of specific words in loan descriptions on the loan success rate. The study found that the proportion of positive tone words and financial words was positively related to the loan success rate, and the proportion of negative tone words, the proportion of strong tone words and the proportion of weak words were positively related to the loan success rate. The proportion of tone words is negatively related to the loan success rate. In addition, many scholars have conducted research on loan interest rates, financing efficiency and other aspects based on P2P online lending texts.