[Paper Review] Generative AI for Business Strategy: Using Foundation Models to Create Business Strategy Tools

Son The Nguyen, Theja Tulabandhula|arXiv (Cornell University)|Aug 27, 2023

Topic ModelingComputer Science3 citations

TL;DR

This paper proposes a framework that leverages foundation models—specifically GPT-4, transformer-based NER, and entailment-based zero-shot classifiers—to automatically generate dynamic, time-varying signed business networks from unstructured news text. The method enables data-driven, quantitative insights into firm relationships (collaboration, competition, or neutrality), with real-world validation showing accurate detection of strategic shifts like Apple’s privacy policy impact on Facebook and Google.

ABSTRACT

Generative models (foundation models) such as LLMs (large language models) are having a large impact on multiple fields. In this work, we propose the use of such models for business decision making. In particular, we combine unstructured textual data sources (e.g., news data) with multiple foundation models (namely, GPT4, transformer-based Named Entity Recognition (NER) models and Entailment-based Zero-shot Classifiers (ZSC)) to derive IT (information technology) artifacts in the form of a (sequence of) signed business networks. We posit that such artifacts can inform business stakeholders about the state of the market and their own positioning as well as provide quantitative insights into improving their future outlook.

Motivation & Objective

• To develop an automated, low-expertise method for constructing time-varying signed business networks from unstructured news data.
• To enable business stakeholders to quantitatively assess market dynamics, firm positioning, and strategic impacts using NLP-powered network abstractions.
• To overcome limitations of static, rule-based, or supervised relation extraction by using zero-shot and instruction-tuned foundation models.
• To provide interpretable, GPT-4-generated explanations for the signs and weights of relationships in the derived networks.
• To demonstrate the framework’s ability to detect real-world strategic shifts, such as Apple’s policy changes affecting competitors.

Proposed method

• Use a fine-tuned transformer-based NER model to extract organization entities from news headlines and text.
• Apply an entailment-based zero-shot classifier (ZSC) to infer relationship polarity (positive, negative, neutral) between entity pairs without labeled training data.
• Construct a sequence of signed business networks where nodes are firms and weighted edges represent relationship sentiment, updated over time.
• Leverage GPT-4 to generate natural language explanations for the inferred edge signs and weights, enhancing interpretability.
• Integrate multiple foundation models in a pipeline: NER → ZSC → network construction → GPT-4 explanation generation.
• Validate the framework using real news data, including high-impact events such as Apple’s App Store privacy changes.

Experimental results

Research questions

• RQ1 Can foundation models be effectively combined to generate dynamic, signed business networks from unstructured news text with minimal human intervention?
• RQ2 How accurately can zero-shot classifiers detect positive, negative, or neutral relationships between firms in news content?
• RQ3 Can GPT-4 provide reliable, human-interpretable explanations for the inferred relationships in the network?
• RQ4 To what extent can the framework detect and reflect real-world strategic shifts, such as the impact of Apple’s privacy policy on competitors?
• RQ5 How does this approach compare to traditional, static, or supervised relation extraction methods in capturing market dynamics?

Key findings

• The ZSC model correctly identified the negative relationship between Apple and Facebook in the headline 'Apple’s Stunning $10 Billion Blow to Facebook' with a confidence score of 0.95.
• The framework detected that Apple’s privacy changes negatively impacted Snap and Facebook but benefited Google, with ZSC scores of 0.97 and 0.96 respectively.
• GPT-4 provided consistent and contextually grounded explanations for edge signs, such as interpreting 'malign' as a negative intent toward TikTok.
• The system captured the shift in relationships following Apple’s policy change, as visualized in network snapshots showing increased negative ties toward Facebook and Google.
• The approach outperformed traditional n-gram models that failed to classify 'compete' as a negative relationship due to feature weighting issues.
• The method demonstrated feasibility in generating interpretable, time-varying signed networks without requiring labeled training data or domain-specific fine-tuning.