Summarization in the Financial and Regulatory Domain

Introduction

Inderjeet Mani defined the goal of automatic summarization (also “summarisation” in British English) as “to take an information source, extract content from it, and present the most important content to the user in a condensed form and in a manner sensitive to the user's or application's need” (Mani, 2001). Therefore, the business value of it lies in its potential for enhancing the productivity of human information consumption (Modaresi et al., 2017): the output of the task of summarizing an input text document comprising English prose is a shorter new document or shorter version of the original document that conveys most of the most important information contained in the original document, yet takes less time to read than the original full document.

Figure 1.

Single-document summarization (left) versus multi-document summarization (right).

Traditionally, we can distinguish between single document summarization, which takes as input a single document (source document) that needs to be summarized, and multi-document summarization, which takes as input a set of documents covering the same topic or topic area (Figure 1). In both cases, a single document, the summary (target document) is to be created. We can further distinguish between extractive summarization, which computes summaries by selecting text spans (phrases, sentences, passages) from the original document or documents, and abstractive summarization, which extracts pieces of information in a pre-processing step, and then constructs a synthetic new document, which is a summary that communicates said extracted facts, or it may even introduce new language not found in the source document(s) (Figure 2, right). Mathematically speaking, extractive summarization can be seen as a sequence of projections. Extractive summarization have the advantage of circumventing the problem of how to generate grammatical sentences as it merely selects from existing sentences; it has the disadvantages that a sequence of selected sentences may not make for smooth reading, as it is hard to combine them so as to maintain cohesion (broadly, to be linked together well at the micro-level) and coherence (roughly, to form a meaningful and logical text at the macro-level). The history of automatic summarization goes back to the German researcher Hans Peter Luhn, who worked on automatic summarization at IBM, where he created the method for extractive single-document summarization now named after him (Luhn, 1958).¹

Figure 2.

Extractive (left) versus abstractive (right) Summarization

We can also distinguish between various kinds of methods. Heuristic methods like the Luhn method (outlined below) typically use a human-conceived scoring function to select relevant text spans that ought to be included in the summary, while machine learning methods derive evidence that leads to the rejection or acceptance for inclusion from data. This can be done in one of two ways: in supervised learning, the most relevant sentences or phrases have been marked up in a set of documents, and an induction algorithm learns which properties of the input text are statistically correlated with high relevance during training time. During runtime, it can then classify pieces of text as relevant or not relevant, or rank (order) pieces of text from most to least relevant. In unsupervised learning, clustering algorithms group documents or text fragments using similarity measures. Pieces similar to others are important if repeated, and at the same time redundancy in the output summary is to be avoided.