Spot the BlindSpots: Systematic Identification and Quantification of Fine-Grained LLM Biases in Contact Center Call Summarization

Kawin Mayilvaghanan, Siddhant Gupta, Ayush KumarEMNLP 2025

TL;DR

Large Language Models (LLMs) are great at summarizing contact center calls, but they can introduce subtle "operational biases" that distort the original interaction's meaning. This paper introduces "BlindSpot," a framework to systematically identify and quantify these biases, revealing that they are widespread across LLMs and can be reduced with targeted prompting.

What Problem Does This Paper Address?

Contact centers rely heavily on call summaries for everything from regulatory compliance to business intelligence. LLMs have revolutionized this process, generating summaries at scale. However, while standard metrics like factual correctness and fluency are important, they don't capture a crucial aspect: operational bias. This refers to systematic distortions in summaries that misrepresent the original conversation's context, sentiment, or key details. For instance, a summary might overemphasize negative customer sentiment or omit crucial resolution steps, leading to flawed agent evaluations, skewed business insights, and ultimately, poor customer satisfaction. Existing research has largely focused on social or positional biases, leaving operational biases in call summarization largely unexplored.

Methods

The core of this paper is the BlindSpot framework, designed to systematically identify and quantify operational biases. It works in three stages:

  1. Transcript Pipeline (Reference Distribution):
    • Transcripts are segmented into manageable chunks.
  2. Each turn is annotated with labels for 15 operational bias dimensions (e.g., sentiment, topic, agent action, disfluency) using a hybrid approach:
  3. LLM-Annotated: For semantic dimensions, an LLM (GPT-40) acts as a "labeler." * Computed: For structural dimensions (like turn length or position), labels are derived algorithmically from metadata.
  4. Derived: Some dimensions (like emotion shift) are inferred from others.
  5. These turn-level labels are aggregated to create a "ground truth" distribution for each bias dimension.
  6. Summary Pipeline (Summary Distribution):
    • The LLM under evaluation generates a summary of the transcript.
  7. The summary is broken down into atomic "propositions" (minimal semantic units).
  8. Each proposition is then labeled using the same LLM labeler and methodology as the transcript turns.
  9. These proposition-level labels are aggregated to form the summary's distribution for each bias dimension.
  10. Bias Quantification and Interpretation:
    • Fidelity Gap (Total Variation Distance - TVD): Measures the overall distributional distortion between the transcript and summary distributions. A score of 0 means identical distributions; higher scores indicate greater divergence.
  11. Coverage %: Measures the percentage of labels present in the transcript that are also present in the summary. This directly quantifies omissions.

The authors also validated their LLM labeler against human annotations, achieving 93.68% accuracy.

Datasets

The study used a dataset of 2,500 real, anonymized contact center transcripts from 12 different domains (e.g., FinTech, Healthcare). These transcripts were substantial, with an average of 317 turns and over 5,000 tokens. No new benchmark dataset was introduced; instead, the focus was on evaluating existing LLMs on this real-world data.

Key Insights and Ideas

  • Operational Biases are Systemic: The paper convincingly shows that biases are not isolated incidents but are present across a wide range of LLMs, regardless of their size or family.
  • LLM-Judge is Insufficient: High scores from "LLM-as-a-Judge" evaluations don't guarantee low bias. Models perceived as high-quality can still exhibit severe operational biases, highlighting the need for specialized bias evaluation.
  • Compression Amplifies Bias: As summaries become more compressed (shorter), bias tends to increase, and coverage of important details decreases.
  • Targeted Prompting Works: The BlindSpot framework's detailed analysis directly informed the creation of a system prompt that measurably reduced bias across multiple LLMs. This demonstrates the practical utility of the framework.
  • Scaling Effect in Mitigation: Larger, more capable LLMs benefited more from the bias mitigation prompt, suggesting they are better equipped to understand and implement complex instructions.

Key Results

  • Widespread Bias: All 20 evaluated LLMs exhibited operational biases across the 15 dimensions.
  • Most Challenging Dimensions: Temporal Sequence (high TVD, indicating misordering of events) and Entity Type (low coverage, meaning many key entities are omitted) were particularly challenging for LLMs.
  • Most Robust Dimensions: Speaker and Position dimensions showed relatively low bias and high coverage, suggesting LLMs are good at attributing speech and identifying general location but struggle with finer details.
  • Bias Mitigation Impact: Targeted prompting led to increased Coverage % (up to +5.87%) and reduced Fidelity Gap (TVD) across nine models, demonstrating the effectiveness of the approach.
  • Fidelity-Conciseness Trade-off: Improving summary fidelity often led to longer, less compressed summaries.

Limitations and Future Work

  • Multilingual Contexts: The study was limited to English-language transcripts, and its applicability to multilingual contexts remains untested.
  • LLM Labeler Bias: While scalable, using LLMs as labelers can propagate existing model biases, especially for subjective dimensions.
  • Harmfulness and User Impact: The current metrics quantify distributional misalignments but don't directly measure the harmfulness, user impact, or real-world consequences of these biases.
  • Future Work: The authors suggest exploring the impact of these biases on downstream applications and developing more robust and domain-aware summarization systems. They also highlight the need for further research into bias mitigation strategies.

Summary updated

BibTeX
@inproceedings{mayilvaghanan2025spot,
  author = {Kawin Mayilvaghanan and Siddhant Gupta and Ayush Kumar},
  title = {{Spot the BlindSpots: Systematic Identification and Quantification of Fine-Grained LLM Biases in Contact Center Call Summarization}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}

Loading related content...

Figure