EDGAR-CORPUS: Billions of Tokens Make The World Go Round

Abstract

We release edgar-corpus, a novel corpus comprising annual reports from all the publicly traded companies in the us spanning a period of more than 25 years. To the best of our knowledge, edgar-corpus is the largest financial nlp corpus available to date. All the reports are downloaded, split into their corresponding items (sections), and provided in a clean, easy-to-use json format. We use edgar-corpus to train and release edgar-w2v, which are word2vec embeddings for the financial domain. We employ these embeddings in a battery of financial nlp tasks and showcase their superiority over generic glove embeddings and other existing financial word embeddings. We also open-source edgarcrawler, a toolkit that facilitates downloading and extracting future annual reports.

KPI-EDGAR: A Novel Dataset and Accompanying Metric for Relation Extraction from Financial Documents

“…In this section, we shortly review the key points of KPI-BERT, a model tailored for key performance indicator extraction, introduced and illustrated in much greater detail in [2] Thereafter, one span-level approach is briefly touched upon, namely the SpERT model introduced by [16]. We then provide two further baselines building on EDGAR-W2V [5] and GloVe [43], leveraging a similar setup like [2]. These four models will be the baselines we provide to other researchers to benchmark their model against on KPI-EDGAR.…”

“…Due to EDGAR's popularity, many researchers ( [3], [4], [6]) have developed methods to extract data from EDGAR and used it in their own research. [5] even released a comprehensive corpus comprising annual reports from all the publicly traded companies in the US spanning a period of more than 25 years and accompanied it with a word2vec [37] model titled EDGAR-W2V, which we will also use as a baseline in our experiments. Furthermore, [38], [39], and [40] have applied machine learning methods to EDGAR data and have provided useful results that can be used in real-world financial applications.…”

“…As defined in the methodology section IV, the introduced joint NER and RE model consists of three building blocks: a sentence encoder, NER decoder, and RE decoder. For generating further baselines, the sentence encoder of the described model is replaced by EDGAR-W2V [5] and GloVe [43]. The rest of the architecture is left unchanged.…”

“…V. EXPERIMENTS Herein, we briefly evaluate the unmodified KPI-BERT [2] (see subsection IV-A), the same structure but with word embeddings from EDGAR-W2V [5] and GloVe [43] instead of embeddings from BERT [27] (see subsection IV-C for these two), and a competing span-based approach (see subsection IV-B), as introduced by [16] and named SpERT, on our novel dataset. The results should be seen as a baseline on which further research can be benchmarked.…”

“…• We introduce a novel dataset, named KPI-EDGAR, for joint named entity recognition and relation extraction in the financial text domain based on actual corporate reports submitted to the EDGAR database. In contrast to other resources scraped from EDGAR (see [3], [4], [5], and [6]), we provide manual annotations along with the corpus to allow for joint named entity recognition and relation extraction.…”