2023 FromSparsetoDenseGPT4Summarizat
- (Adams, Fabbri et al., 2023) ⇒ Griffin Adams, Alexander Fabbri, Faisal Ladhak, Eric Lehman, and Noémie Elhadad. (2023). “From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting.” doi:10.48550/arXiv.2309.04269
Subject Headings: LLM-based Summarization
Notes
- It introduces the Chain of Density (CoD) prompting technique to generate dense GPT-4 summaries without extending their length.
- It employs an iterative method where GPT-4 starts with an entity-sparse summary and then incorporates missing salient entities, maintaining the summary's original length.
- It emphasizes that CoD summaries are more abstractive, show more fusion, and reduce lead bias compared to the summaries produced by a vanilla GPT-4 prompt.
- It provides evaluation results from a human preference study on 100 CNN/DailyMail articles, highlighting that humans favor GPT-4 summaries that are denser than vanilla prompts but almost as dense as human-written summaries.
- It acknowledges a tradeoff between a summary's informativeness and its readability.
- It releases a dataset of 500 annotated CoD summaries and an additional 5,000 unannotated summaries available on HuggingFace.
- It aims to understand the optimal information density in summaries for maximizing informativeness while ensuring coherence and clarity.
- It uses the average number of entities per token in a summary as a metric to gauge its density.
- It identifies a challenge: ensuring that summaries remain legible and accurate as they become more information-dense.
- It aims to find the limit of density in summaries by considering human preferences on sets of increasingly dense GPT-4-produced summaries.
- High-level Algorithm:
- Generate Initial Summary: Prompt GPT-4 to produce a verbose, sparse initial summary with minimal entities.
- Identify Missing Entities: Extract 1-3 concise, relevant, novel entities from the source text not in previous summary.
- Fuse Entities: Prompt GPT-4 to rewrite previous summary fusing in missing entities without increasing length. Employ compression and abstraction techniques to make space.
- Iterate: Repeat Identify Missing Entities and Fuse Entities steps multiple times, incrementally densifying summary by packing in more entities per token through rewriting.
- Output Chain: The final output is a chain of fixed-length summaries with increasing density produced through iterative abstraction, fusion, and compression.
Cited By
Quotes
Abstract
Selecting the "right" amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (this https URL).
Introduction
Automatic summarization has come a long way in the past few years, largely due to a paradigm shift away from supervised fine-tuning on labeled datasets to zero-shot prompting with Large Language Models (LLMs), such as GPT-4 (OpenAI, 2023). Without additional training, careful prompting can enable fine-grained control over summary characteristics, such as length (Goyal et al., 2022), topics (Bhaskar et al., 2023), and style (Pu and Demberg, 2023).
An overlooked aspect is the information density of an summary. In theory, as a compression of another text, a summary should be denser–containing a higher concentration of information–than the source docu- ment. Given the high latency of LLM decoding (Kad- dour et al., 2023), covering more information in fewer
image
1https://huggingface.co/datasets/
griffin/chain_of_density
Figure 1: Chain of Density (CoD) summaries grow increasingly entity dense, starting off closer to vanilla GPT-4 summaries and eventually surpassing that of human written summaries. Human annotations suggest that a density similar to that of human-written summaries is preferable–striking the right balance between clarity (favors less dense) and informativeness (favors more dense).
words is a worthy goal, especially for real-time appli- cations. Yet, how dense is an open question. A sum- mary is uninformative if it contains insufficient detail. If it contains too much information, however, it can be- come difficult to follow without having to increase the overall length. Conveying more information subject to a fixed token budget requires a combination of abstrac- tion, compression, and fusion. There is a limit to how much space can be made for additional information before becoming illegible or even factually incorrect.
In this paper, we seek to identify this limit by solic- iting human preferences on a set of increasingly dense summaries produced by GPT-4. Treating entities, and, in particular, the average number of entities per token, as a proxy for density, we generate an initial, entity- sparse summary. Then, we iteratively identify and fuse 1-3 missing entities from the previous summary with- out increasing the overall length (5x overall). Each summary has a higher ratio of entities to tokens than the previous one. Based on human preference data, we determine that humans prefer summaries that are al- most as dense as human-written summaries and more
image image image image image image image image image image image
image
Figure 2: Chain of Density (CoD) Prompt and example output. At each step, 1-3 additional details (entities) are added to the previous summary without increasing the length. To make room for new entities, existing content is re-written (e.g., compression, fusion). Half the annotators (2/4) prefer the second to last summary, with the others preferring the final one.
dense than those generated by a vanilla GPT-4 prompt. Our primary contributions are to:
Develop a prompt-based iterative method (CoD) for making summaries increasingly entity dense.
Conduct both human and automatic evaluation of increasingly dense summaries on CNN/Dai- lymail articles to better understand the tradeoff between informativeness (favoring more entities) and clarity (favoring fewer entities).
Open source GPT-4 summaries, annotations, and a set of 5,000 unannotated CoD summaries to be used for evaluation or distillation.
Chain of Density Prompting Prompt. Our goal is to generate a set of summaries with GPT-4 with varying levels of information density, while controlling for length, which has proven to be a strong confounder when evaluating summaries (Fabbri et al., 2021; Liu et al., 2023b). To do this, we formu- late a single Chain of Density (CoD) prompt, whereby an initial summary is generated and made increasingly entity dense. Specifically, for a fixed number of turns, a set of unique salient entities from the source text are identified and fused into the previous summary without increasing the length. The first summary is entity-sparse as it focuses on only 1-3 initial entities.
To maintain the same length while increasing the num- ber of entities covered, abstraction, fusion, and com- pression is explicitly encouraged, rather than dropping meaningful content from previous summaries.
Figure 2 displays the prompt along with an example output. Rather than be prescriptive about the types of entities, we simply define a Missing Entity as:
Relevant: to the main story. Specific: descriptive yet concise (5 words or fewer). Novel: not in the previous summary. Faithful: present in the Article. Anywhere: located anywhere in the Article. Data. We randomly sample 100 articles from the CNN/DailyMail summarization (Nallapati et al., 2016) test set for which to generate CoD summaries.
Reference Points. For frame of reference, we compare CoD summary statistics to human-written bullet-point style reference summaries as well as summaries generated by GPT-4 with a vanilla prompt: “Write a VERY short summary of the Article. Do not exceed 70 words.” We set the desired token length to match that of CoD summaries (shown in Table 1). Statistics Direct statistics (tokens, entities, entity density) are ones directly controlled for by CoD, while Indirect
image
Figure 3: CoD-generated summaries grow increasingly abstractive while exhibiting more fusion and less of a lead bias.
statistics are expected byproducts of densification.
CoD Step
Tokens
Entities
Density (E/T)
1
72
6.4
0.089
2
67
8.7
0.129
3
67
9.9
0.148
4
69
10.8
0.158
5
72
12.1
0.167
Human
60
8.8
0.151
Vanilla GPT-4
70
8.5
0.122
Table 1: Explicit statistics for GPT-4 CoD summaries.
- Direct Statistics.
In Table 1, we compute tokens with NLTK (Loper and Bird, 2002), measure unique entities with Spacy2, and compute entity density as the ratio. The CoD prompt largely adheres to a fixed to- ken budget. In fact, the second step leads to an average 5-token (72 to 67) reduction in length as unnecessary words are removed from the initially verbose summary. The entity density rises–starting at 0.089, initially below Human and Vanilla GPT-4 (0.151 and 0.122)–to 0.167 after 5 steps of densification.
- Indirect Statistics.
Abstractiveness should increase with each CoD step because summaries are iteratively re-written to make space for each additional entity. We measure abstractiveness with extractive density: the average squared length of extractive fragments (Grusky et al., 2018). Similarly, the level of concept Fusion should increase monotonically as entities are added to a fixed-length summary. We proxy fusion as average number of source sentences aligned to each summary sentence. For alignment, we use the relative ROUGE gain method (Zhou et al., 2018), which aligns source sentences to a target sentence until the relative ROUGE gain of an additional sentence is no longer positive. We also expect the Content Distribution – the position in the Article from which summary content is sourced–to shift. Specifically, we expect that CoD summaries initially exhibit a strong Lead Bias yet gradually start to pull in entities from the middle and end of the article. To measure this, we use our alignments from fusion and measure the average sentence rank of all aligned source sentences. Figure 3 confirms these hypotheses: abstractiveness increases with the number of re-writing steps (lower extractive density on the left), the rate of fusion rises (middle figure), and the summaries start to incorporate content from the middle and end of the article (right figure). Interestingly, all CoD summaries are more abstractive than both human written and baseline summaries.
image
2https://spacy.io.
- Results
To better understand the tradeoffs present with CoD summaries, we conduct a preference-based human study and a rating-based evaluation with GPT-4.
image
image
CoD % Share of First Place Votes
Step Individual Annotators Aggregate
1
3.0
2.0
13.0
17.4
8.3
2
25.0
28.0
43.0
31.4
30.8
3
22.0
28.0
21.0
24.4
23.0
4
29.0
25.0
13.0
26.7
22.5
5
21.0
17.0
10.0
16.3
15.5
Table 2: Breakdown of first-place votes for CoD summaries by step. Based on aggregate preferences, the modal CoD step is 2, median is 3, and expected is 3.06.
Human Preferences. We conduct a human evaluation to assess the impact of densification on human assessments of overall quality. Specifically, the first four authors of the paper were presented with randomly shuffled CoD summaries, along with the articles, for the same 100 articles (5 steps * 100 = 500 total summaries). Based on the definition of a “good summary" from Stiennon et al. (2020) (Table 6 from their paper), each annotator indicated their top preferred summary. Table 2 reports the breakdown of first place votes by CoD step across annotators–as well as aggregated across annotators. First, we report a low Fleiss’ kappa (Fleiss, 1971) of 0.112, which points to the subtle differences between summaries and the subjective nature of the task. Recent work has
CoD
Step
Entity Density
Informative
Quality
Coherence
Attributable
Overall
GPT-4 Eval Average
1
0.089
4.34
4.75
4.96
4.96
4.41
4.69
2
0.129
4.62
4.79
4.92
5.00
4.58
4.78
3
0.148
4.67
4.76
4.84
5.00
4.57
4.77
4
0.158
4.74
4.69
4.75
5.00
4.61
4.76
5
0.167
4.73 4.65 4.61 4.97 4.58 4.71
- Table 3: GPT-4 Likert-scale (1-5) assessments of Chain of Density (CoD) Summaries by step.
Figure 4: An example of a human-preferred densification step (left) and one which is not preferred. For the left, the bottom summary is preferred because the addition of “Liverpool” and the goal-scorers is relevant. The second summary makes room with sensible compressions, such as synthesizing “a potential route back into the game” into “a comeback”. For the right, the addition of more details on “TVMonde” does not make up for the presence of an awkward fusion of entities (“cyberattack”, and “Yves Bigot”), which was a direct result of having to tighten the previous summary.
similarly noted low instance-level agreement when judging GPT-based summaries (Goyal et al., 2022).
Yet, at the system level, some trends start to emerge. For 3 of the 4 annotators, CoD step 1 received the largest share of first-place votes across the 100 examples (28, 43, and 31.4%, respectively). Yet, in aggregate, 61% of first-placed summaries (23.0+22.5+15.5) involved ≥ 3 densification steps. The median preferred CoD step is in the middle (3), and the expected step is 3.06.
Based on the average density of Step 3 summaries, we can roughly infer a preferred entity density of
∼ 0.15 across the CoD candidates. From Table 1, we can see that this density aligns with human-written summaries (0.151), yet is noticeable higher than sum- maries produced with a vanilla GPT-4 prompt (0.122).
Automatic Metrics. As an evaluator, GPT-4 has been shown to adequately correlate to human judgments (Fu et al., 2023; Liu et al., 2023a), even potentially outperforming crowd-sourced workers on some annotation tasks (Gilardi et al., 2023). As a complement to our human evaluation (below), we prompt GPT-4 to rate CoD summaries (1-5) along 5 dimensions: Informative, Quality, Coherence, At- tributable, and Overall. The definitions of informative, Quality, and Attributable come from Aharoni et al. (2023), while Coherence comes from Fabbri et al. (2021)3. Overall aims to capture the qualities jointly. Please see Appendix A for the prompts used image
3 Quality and Coherence are article-independent metrics.
to solicit scores for each dimension. Table 3 suggests that densification is correlated with informativeness, yet there is a limit, with the score peaking at Step 4 (4.74). Article-free dimensions: Quality and Coher- ence, decline sooner (after 2 and 1 steps, respectively). All summaries are deemed Attributable to the source article. The Overall scores skew toward denser and more informative summaries, with Step 4 having the highest score. On average across dimensions, the first and last CoD steps are least favored, while the mid- dle three are close (4.78, 4.77, and 4.76, respectively). In Appendix A, we report highest summary- level correlations of the Overall metric to human judgments (0.31 Pearson correlation), yet note low cor- relations overall–a phenomenon observed by Deutsch et al. (2022) when summaries are of similar quality.
Qualitative Analysis. There exists a clear trade-off between coherence / readability of summaries and in- formativeness. To illustrate, in Figure 4, we present two CoD steps: one for which the summary is im- proved with more detail, and one for which the sum- mary is harmed. On average, intermediate CoD sum- maries best achieved this balance, yet we leave it to fu- ture work to precisely define and quantify this tradeoff. Related Work GPT Summarization. Goyal et al. (2022) bench- marked GPT-3 on news article summarization and found that humans preferred GPT-3 summaries over previous supervised baselines, which was
not reflective of existing reference-based and reference-free metrics. Zhang et al. (2023) find that zeroshot GPT-3 summaries perform on par with humans by soliciting high-quality summaries from freelance writers. Entity-Based Summarization. Narayan et al. (2021) proposed generating entity chains as a planning step for supervised fine-tuning of summarization models, in contrast to keywords (Li et al., 2020; Dou et al., 2021) or purely extractive units (Dou et al., 2021; Adams et al., 2023a). Entities have also been incorporated for summarization as a form of control (Liu and Chen, 2021; He et al., 2022; Maddela et al., 2022), to improve faithfulness (Nan et al., 2021; Adams et al., 2022), and as a unit for evaluation (Cao et al., 2022; Adams et al., 2023b).
6. Conclusionv
We study the impact of summary densification on human preferences of overall quality. We find that a degree of densification is preferred, yet, when summaries contain too many entities per token, it is very difficult maintain readability and coherence. We open-source annotated test set as well as a larger un-annotated training set for further research into the topic of fixed-length, variable density summarization.
7. Limitations
We only analyze CoD for a single domain, news summarization. Annotations did not show high summary-level agreement yet did start to show system-level trends, which is in line with previous work on LLM-based evaluation (Goyal et al., 2022). Finally, GPT-4 is a closed source model so we cannot share model weights. We do, however, publish all evaluation data, annotations, as well as 5, 000 un-annotated CoD to be used for downstream uses cases, e.g., density distillation into an open-sourced model such as LLAMA-2 (Touvron et al., 2023).
References
;
| Author | volume | Date Value | title | type | journal | titleUrl | doi | note | year | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2023 FromSparsetoDenseGPT4Summarizat | Faisal Ladhak Griffin Adams Alexander Fabbri Eric Lehman Noémie Elhadad | 10.48550/arXiv.2309.04269 | 2023 |
[[title::From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting|]]