請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73117
標題: | 以文字探勘方法理解財報中的管理層討論與分析 Understanding the MD&A Section in Financial Reports through Sentence Classification |
作者: | Sung-Ying Fang 方松營 |
指導教授: | 盧信銘(Hsin-Min Lu) |
關鍵字: | 文字探勘,管理層與討論分析,10-K財報,人工標記,BERT,句子分類器, Textual Analysis,MD&A,10-K Filings,Manual Tagging,BERT,Sentence Classifier, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 在會計與財經領域研究文字資訊的文獻中,文字探勘與其內容分析方法向來是實驗重要的基礎與環節之一。然而,這些文獻中常用的內容分析方法多為情緒辭典或基礎的機器學習方法,少有考量自然語言處理近年在深度學習與移轉學習上的突破。
本研究利用自然語言處理預訓練的模型(Bidirectional Encoder Representation from Transformer,簡稱BERT),希望藉此提高財經領域文本文字探勘方法的準確率。為此,本研究對年報中的管理層與討論分析項目(MD&A)進行標籤作為訓練資料,以便句子分類模型能準確分辨是否具前瞻性、句子的文本情緒、以及句子內容包含哪些會計項目。建構句子分類模型後,本研究分析2007年至2016年的MD&A,並探討近年文獻回顧中財經文字探勘領域重要的研究議題,以證明標籤資料與分類器的準確率。最終我們得出,我們的實驗結果多與過去文獻結果相符,尤其與Li(2010) 的研究得出相近的結論,發現MD&A的情緒文本可闡釋部分財經相關的數據資料和各個產業內的環境因素。 Textual analysis has been an emerging area in accounting and finance research. With the growing realization that economical and statistical models may not adequately explain the market with conventional quantitative measures alone, there has been extensive empirical literature attempting to incorporate verbal, non-quantitative measures. However, in fields such as natural language processing, there have been extensive progress in machine learning methods, such as new methods of language representations and the application of transfer learning, which has not been commonly used in academic papers featuring textual analysis within financial context. This paper wishes to contribute to the field by preparing a training dataset suitable for a wider array of research questions, and apply a contemporary machine learning method using bidirectional encoding representations from transformers, or BERT. Our sentence classifier aims to correctly classify sentences with respect to their tone, their accounting category or topic based on the context of the sentence, and whether or not the sentence is a forward-looking statement. By applying our sentence classifier to out-of-sample annual filings, we evaluate our dataset and classification method by revisiting a subset of research questions concluded from our literature review. Our dataset and preliminary descriptive analysis align with the results of many empirical models from other studies, most notably with the analysis made by Li (2010). In conclusion, the tone of MD&A sections is mean-reverting, may proxy for economic determinants, and can be useful in inspecting the macro-environment of an industry. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73117 |
DOI: | 10.6342/NTU201901359 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 3.52 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。