Dataset Card for financial_phrasebank

Dataset Summary

Polar sentiment dataset of sentences from financial news. The dataset consists of 4840 sentences from English language financial news categorised by sentiment. The dataset is divided by agreement rate of 5-8 annotators.

Supported Tasks and Leaderboards

Sentiment Classification

Languages

English

Dataset Structure

Data Instances

{ "sentence": "Pharmaceuticals group Orion Corp reported a fall in its third-quarter earnings that were hit by larger expenditures on R&D and marketing .",
  "label": "negative"
}

Data Fields

Each example contains:

• sentence: a tokenized line from the dataset
• label: a label corresponding to the class as a string: 'positive', 'negative' or 'neutral'

Data Splits

There's no train/validation/test split.

However the dataset is available in four possible configurations depending on the percentage of agreement of annotators:

• sentences_50agree; Number of instances with >=50% annotator agreement: 4846
• sentences_66agree: Number of instances with >=66% annotator agreement: 4217
• sentences_75agree: Number of instances with >=75% annotator agreement: 3453
• sentences_allagree: Number of instances with 100% annotator agreement: 2264

Usage

from datasets import load_dataset

# Load the highest-agreement configuration
ds = load_dataset("takala/financial_phrasebank", "sentences_allagree")

print(ds)
print(ds["train"][0])

Other configurations (e.g. sentences_75agree, sentences_66agree) can be loaded by changing the second argument.

Quick baseline (Transformers)

from datasets import load_dataset
from transformers import pipeline

ds = load_dataset("takala/financial_phrasebank", "sentences_allagree")["train"]
clf = pipeline("text-classification", model="distilbert-base-uncased-finetuned-sst-2-english")

print(ds[0]["sentence"])
print(clf(ds[0]["sentence"]))

Dataset Creation

Curation Rationale

The key arguments for the low utilization of statistical techniques in financial sentiment analysis have been the difficulty of implementation for practical applications and the lack of high quality training data for building such models. Especially in the case of finance and economic texts, annotated collections are a scarce resource and many are reserved for proprietary use only. To resolve the missing training data problem, we present a collection of ∼ 5000 sentences to establish human-annotated standards for benchmarking alternative modeling techniques.

The objective of the phrase level annotation task was to classify each example sentence into a positive, negative or neutral category by considering only the information explicitly available in the given sentence. Since the study is focused only on financial and economic domains, the annotators were asked to consider the sentences from the view point of an investor only; i.e. whether the news may have positive, negative or neutral influence on the stock price. As a result, sentences which have a sentiment that is not relevant from an economic or financial perspective are considered neutral.

Source Data

Initial Data Collection and Normalization

The corpus used in this paper is made out of English news on all listed companies in OMX Helsinki. The news has been downloaded from the LexisNexis database using an automated web scraper. Out of this news database, a random subset of 10,000 articles was selected to obtain good coverage across small and large companies, companies in different industries, as well as different news sources. Following the approach taken by Maks and Vossen (2010), we excluded all sentences which did not contain any of the lexicon entities. This reduced the overall sample to 53,400 sentences, where each has at least one or more recognized lexicon entity. The sentences were then classified according to the types of entity sequences detected. Finally, a random sample of ∼5000 sentences was chosen to represent the overall news database.

Who are the source language producers?

The source data was written by various financial journalists.

Annotations

Annotation process

This release of the financial phrase bank covers a collection of 4840 sentences. The selected collection of phrases was annotated by 16 people with adequate background knowledge on financial markets.

Given the large number of overlapping annotations (5 to 8 annotations per sentence), there are several ways to define a majority vote based gold standard. To provide an objective comparison, we have formed 4 alternative reference datasets based on the strength of majority agreement:

Who are the annotators?

Three of the annotators were researchers and the remaining 13 annotators were master's students at Aalto University School of Business with majors primarily in finance, accounting, and economics.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

All annotators were from the same institution and so interannotator agreement should be understood with this taken into account.

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-sa/3.0/.

If you are interested in commercial use of the data, please contact the following authors for an appropriate license:

• Pekka Malo
• Ankur Sinha

Citation Information

@article{Malo2014GoodDO,
  title={Good debt or bad debt: Detecting semantic orientations in economic texts},
  author={P. Malo and A. Sinha and P. Korhonen and J. Wallenius and P. Takala},
  journal={Journal of the Association for Information Science and Technology},
  year={2014},
  volume={65}
}

Contributions

Thanks to @frankier for adding this dataset.