Summarization › Limitations and Challenges

The most pressing limitation of abstractive summarization is faithfulness. Large language models frequently generate summaries containing facts not present in, or contradicted by, the source document. A study of BART-generated summaries found that roughly 25% contained at least one factual error, ranging from incorrect numbers to wrong entity relationships.

This data quality issue follows directly from the generation mechanism. During decoding, the model balances many competing signals: the source content, the fluency of the text being generated, and the prior probability of token sequences learned from pretraining. When these signals conflict, the model does not always prioritize faithfulness. It may smooth over an awkward phrase from the source with a fluent but incorrect alternative. A model that has seen thousands of news articles stating that inflation "rose to 4.2%" will have a strong prior for that sentence pattern, and may reproduce a plausible-looking percentage even when the source document says 3.8%.

Hallucination in summarization takes several distinct forms. Intrinsic hallucination introduces information that contradicts the source: changing a date, inverting a causal relationship, or swapping two entities. Extrinsic hallucination introduces information that is not contradicted by the source but is also not present in it: adding a detail that seems plausible given the topic but was never stated. Both types are harmful, but extrinsic hallucination is harder to detect because the hallucinated content may be factually true (just not in this specific document) and therefore not obviously wrong to a reader.

Faithfulness problems are harder to detect than grammatical errors because a hallucinated fact can be indistinguishable from a correct fact to a reader who does not independently verify it. Automatic faithfulness metrics like FactCC and QAGS attempt to measure this by checking whether claims in the summary can be inferred from the source, but they are imperfect and expensive to compute. For high-stakes applications (medical, legal, financial), post-generation fact verification against the source is often necessary. One practical approach is to use an LLM to check each claim in the generated summary against the source document, producing a faithfulness score alongside the summary itself.

Summarization models exhibit length biases in both directions. Very short documents are sometimes padded with repetition to meet minimum length constraints. Very long documents are often truncated before being summarized, because most Transformer models have context windows of 512 to 4,096 tokens. A research paper may be 10,000 words; summarizing only the first 1,000 words can produce misleading summaries that miss the conclusions entirely. This truncation problem is particularly insidious because the model produces a confident-looking summary without any indication that it never read the most important part of the document.

Long-context models (as we discussed in Chapter 15) partially address this, but even with longer windows, the model's ability to attend to distant parts of the input tends to degrade. The "lost in the middle" phenomenon describes a consistent finding in long-context LLM evaluations: information in the middle of a long document is recalled less reliably than information near the beginning or end. For summarization, this means that key findings buried in the middle sections of a long document may be systematically underrepresented.

Hierarchical summarization approaches, which first summarize sections and then summarize the section summaries, are a practical workaround that avoids the direct length constraint. But they introduce their own failure modes: information lost in the first stage of summarization cannot be recovered in the second stage. A critical caveat mentioned in the discussion section of a paper will not appear in a section summary that focuses on the main results, and therefore will not appear in the final summary either.

ROUGE scores are widely reported but poorly correlated with human judgments of quality in many settings. Studies have found correlations between ROUGE-2 and human quality judgments as low as 0.2 to 0.3 depending on the evaluation dimension being measured. Models fine-tuned to maximize ROUGE can learn to produce verbose repetitions of common words that inflate recall without improving quality. This optimization pressure creates a metric gaming problem: the winning model under ROUGE may not be the model that produces the most useful summaries for readers.

BERTScore and newer learned metrics (SummEval, BLANC) show better correlation with human judgments but are more computationally expensive and themselves require validation against human data. The SummEval benchmark (Fabbri et al., 2021) systematically compared many automatic metrics against expert human judgments and found that no single metric consistently outperformed others across all quality dimensions. Coherence, in particular, is almost impossible to capture automatically because it requires understanding narrative flow across sentences, something that token-overlap metrics cannot access.

The field currently lacks a single trusted automatic metric that generalizes across domains and tasks. This creates a practical problem: when you compare two summarization systems, ROUGE might prefer one while BERTScore prefers the other and human judges prefer a third. Understanding which metric to trust requires domain-specific calibration against human judgments, which most practitioners do not have the resources to perform.

A model fine-tuned on CNN/DailyMail news articles may perform poorly on scientific papers, legal documents, or medical records. News articles follow predictable structures, use general vocabulary, and have available training data in the millions of examples. Technical documents do not follow predictable structures, use specialized vocabulary that may not appear in general pretraining corpora, and often have no publicly available summarization training data.

The domain mismatch manifests in several ways. The model may produce summaries at the wrong level of technical detail (too simplified for expert readers, or too technical for the intended audience). It may hallucinate domain-specific facts by drawing on its general pretraining knowledge rather than the source document. It may miss domain-specific conventions: a medical summary that omits statistical significance values is incomplete by clinical standards, even if those values seem like minor technical details to a general reader.

This is especially problematic for organizations that want to summarize proprietary internal documents that differ substantially from public training data. Meeting transcripts, internal policy documents, and technical specifications have very different structure and vocabulary from news or academic literature. Fine-tuning on domain-specific data helps significantly but requires annotated examples, which are expensive to create. For many specialized domains, the cost of creating enough labeled data for fine-tuning makes pre-trained general models the only practical option, despite their limitations.

Current models offer limited control over the level of abstraction in the output. A user might want a one-sentence headline, a three-sentence executive summary, or a structured bullet-point breakdown of the same document. While length parameters partially control this, the style and structure of the output are harder to specify. Instruction-tuned models like GPT-4 and Claude show significantly better adherence to specific summarization instructions, but this capability is not universal and depends heavily on how the instruction is phrased and whether the model was trained on similar instruction patterns.

The mismatch between user intent and model output is often subtle. A user who asks for "a brief summary focusing on the key risks" may receive a summary that is brief and covers risks, but weights financial risks more heavily than operational or reputational risks based on the model's training distribution rather than the user's actual priorities. Making these preferences explicit in natural language requires users to anticipate and articulate every dimension of their intent, which is cognitively demanding and error-prone. Structured output formats (bullet points with specific headings) help constrain the output format but do not fully solve the content selection problem.

Every summary involves information loss, and the acceptable loss depends entirely on the use case. A 10:1 compression ratio (reducing a 1,000-word article to 100 words) necessarily discards 90% of the content. For a casual reader seeking a quick overview, this may be appropriate. For a researcher who needs complete understanding of the methodology, it is almost certainly too aggressive.

Current models lack principled mechanisms for deciding which 90% to discard. They use proxies (word frequency, position, attention weights) that correlate with importance on average but fail in specific cases. The supporting detail that appears only once in a document and never attracts high attention might be precisely the piece of information the reader needs. A summary system that consistently discards unique details in favor of repeated assertions will produce summaries that reflect popular content rather than critical content.

This problem is intrinsic to the task formulation. If the user's information need is not specified, the model must guess at it. The more aligned the model's guess is with typical readers of typical documents in its training distribution, the better it will perform on average. The more atypical the user's need or the document's topic, the more likely the model is to discard important information. Understanding this limitation helps practitioners set appropriate expectations: summarization systems work best when the user's information need closely matches the system's implicit assumptions about what is important.