Abstract
Increasing attention has been drawn to the sentiment analysis of financial documents. The most popular examples of such documents include analyst reports and economic news, the analysis of which is frequently used to capture the trends in market sentiments. On the other hand, the significance of the role sentiment analysis plays in the financial domain has given rise to the efforts to construct a financial domain-specific sentiment lexicon. Sentiment lexicons lend a hand for solving various text mining tasks, such as unsupervised classification of text data, while alleviating the arduous human labor required for manual labeling. One of the challenges in the construction of an effective sentiment lexicon is that the semantic orientation of a word may change depending on the context in which it appears. For instance, the word "profit" usually conveys positive sentiments; however, when the word is juxtaposed with another word "decrease," the sentiment associated with the phrase "profit decreases" now becomes negative. Hence, the sentiment of a given word may shift as one begins to consider the context surrounding the word. In this paper, we address this issue by incorporating context when building sentiment lexicon from a given corpus. Specifically, we construct a lexicon named Senti-DD for the Sentiment lexicon composed of Direction-Dependent words, which expresses each term a pair of a directional word and a direction-dependent word. Experiment results show that higher classification performance is achieved with Senti-DD, proving the effectiveness of our method for automatically constructing a context-aware sentiment lexicon in the financial domain.
Introduction
Sentiment analysis in the financial sector is notoriously complex because the meaning of a word often depends entirely on its linguistic neighbors. Most existing tools rely on "unigram" lexicons—simple lists where a word like "profit" is statically marked as positive. However, in a financial report, the phrase "profit decreases" is clearly negative, while "costs decrease" is positive. Standard lexicons, such as the widely used Loughran-McDonald (LM) list, often miss these nuances because they fail to account for the interaction between directional words (e.g., up, down, rise, fall) and direction-dependent words (e.g., profit, liability, cost).
While some researchers have attempted to build context-aware lexicons manually, the process is slow, expensive, and prone to human bias. This paper proposes Senti-DD, a framework that automates this construction using a data-driven approach, identifying terms whose sentiment flips based on directional modifiers.
Methodology
The authors propose a two-stage framework consisting of lexicon construction and subsequent sentiment classification.
Senti-DD Construction
The construction phase identifies the relationship between nouns and sentiment directionality:
- Tagging: Sentences in a labeled financial corpus are categorized as either "proportional" (sentiment follows the direction, e.g., "revenue increased") or "inversely proportional" (sentiment is opposite to the direction, e.g., "loss increased").
- Statistical Association: Using Pointwise Mutual Information (PMI), the system measures the strength of association between specific nouns and these two categories. A word like "profit" shows high PMI with the "proportional" tag, while "cost" associates with "inversely proportional."
- Pairing: The system generates a lexicon of pairs:
(Directional Word, Dependent Word). For example,(up, profit)is labeled positive, while(down, profit)is labeled negative.
Sentiment Classification
The classification system uses Senti-DD as a refinement layer. It starts with a baseline score from the Loughran-McDonald list and then adjusts this score when Senti-DD pairs are detected. If the system finds a (down, profit) pair, it adjusts the sentence's sentiment score downward, even if "profit" was originally flagged as a positive unigram.
Experimental Setup
The framework was evaluated using the Financial Phrase Bank, a collection of ~5,000 sentences from English financial news headlines annotated by 16 experts. Testing was conducted across four subsets (DS50, DS66, DS75, and DS100) categorized by the level of annotator agreement. The Loughran-McDonald (LM) Word List served as the primary baseline lexicon, and the Harvard General Inquirer was used as a seed list to identify initial directional words.
Key Results and Insights
The proposed LM+Senti-DD framework consistently outperformed all baselines, including SentiWordNet, TextBlob, VADER, and the standard LM list.
- Accuracy Improvements: On the DS100 dataset (100% expert agreement), the F1-score reached 0.81, compared to 0.59 for the standard LM list.
- Negative Sentiment Detection: The improvement was most dramatic in the negative class. For the DS100 dataset, the F1-score for negative sentiment was 0.71, nearly double the 0.40 achieved by the LM list alone.
- Robustness: The method showed performance gains even on the "DS50" dataset, suggesting the statistical signal of direction-dependency is strong enough to overcome noisy training data.
Qualitatively, the model correctly identified 72 "proportional" words (like ebit, margin, revenue) and 7 "inversely proportional" words (like cost). This confirms that the relationship between a directional verb and a noun is a reliable, automatable proxy for sentiment in finance.
Conclusion
Senti-DD provides a robust, "plug-and-play" solution for context-aware sentiment analysis. By fixing the specific errors that unigram-based lists are most prone to making, it significantly enhances the accuracy of financial text mining. Future work will aim to address current limitations, such as the noun-centric preprocessing and the data imbalance between proportional and inversely proportional samples, while exploring applications in other domains like electronics reviews.
Summary updated