We examine the ability of large language models (LLMs) to generate salient (interesting) negative statements about real-world entities; an emerging research topic of the last few years. We probe the LLMs using zero- and -shot unconstrained probes, and compare with traditional methods for negation generation, i.e., pattern-based textual extractions and knowledge-graph-based inferences, as well as crowdsourced gold statements. We measure the correctness and salience of the generated lists about subjects from diferent domains. Our evaluation shows that guided probes do in fact improve the quality of generated negatives, compared to the zero-shot variant. Nevertheless, using both prompts, LLMs still struggle with the notion of factuality of negatives, frequently generating many ambiguous statements, or statements with negative keywords but a positive meaning.
KG
Human didn’t make his high school team
˂˂˂˂
˂do˂esn˂’t ˂ha˂ve ˂so˂cial media
isn’t a basketball coach
didn’t play as a power forward
didn’t invent basketball
didn’t only play basketball (positive)
never played for a team outside the u.s.
didn’t play for the bulls exclusively (positive)
didn’t play for chicago until 84 (positive)
didn’t win a championship for the lakers
wasn’t the youngest player in the nba
didn’t win an oscar
didn’t buy stakes in the chicago bulls
never coached the chicago bulls
lists of statements (biographic summaries) about subjects, where the statements are truly
negative, but also salient, unexpected, or normally mistaken as true positives. To compile
these lists, diferent data sources and methodologies have been explored. In [
We summarize our contributions as follows.
• We design constraint-free prompts for LLM-based negation generation, where we only instantiate the input subject. • We examine LLMs’ understanding of salient factual negation, finding that, even though they struggle with the notion of true negativity (-18% in correctness compared to SOTA model), on truly negative statements, the guided few-shot ChatGPT variant ranks first among models in salience. • We study both encyclopedic and commonsense domains, finding that it is more challenging for LLMs to generate longer lists of salient commonsense negatives. For instance, the zero-shot ChatGPT variant shows a decrease of 22% in correctness@5 (compared to @1) for commonsense subjects. No decrease is observed for encyclopedic subjects. • We compare the LLM-generated negative statements to existing SOTA methods, from text [5] and knowledge graphs [3]. • We measure the quality of the negative statements over two aspects, the correctness (true negativity) and salience (interestingness).
The data generated can be downloaded at: https://www.mpi-inf.mpg.de/fileadmin/inf/d5/ research/negation_in_KBs/data.csv.
Given a subject, we probe the LLM to generate a list of salient negative statements about it. Zero-shot Probe. In this probe, we test the performance of the LLM without providing any samples in our instructions.
Write a list of [ n] salient factual negated statement about [ SUBJECT].
The goal is to inspect the model’s interpretation of the notion of salient negation without any prior examples nor definitions.
Guided Few-shot Probe. In this probe, we guide the model with both definitions and examples (for in-context learning).
A salient factual negated statement about an entity means that the statement doesn’t hold in reality. Moreover, the negated statement is either surprising, unexpected, or useful to the reader. For example: [ EXAMPLE1] [...]
Given this definition and examples, write a list of [ n] salient factual negated statement about [ SUBJECT].
In the following sample, we show a 4-shot probe with 2 salient and 2 nonsalient samples about diferent types of subjects, and request 3 salient negative statements about lebanon (LLM=ChatGPT).
A salient factual negated statement about an entity means that the statement doesn’t hold in reality. Moreover, the negated statement is either surprising, unexpected, or useful to the reader. For example: penguins can’t fly.
tom cruise never won an oscar.
penguins can’t run for presidency.
tom cruise never won the nba best player award.
Given this definition and examples, write a list of 3 salient factual negated statement about lebanon.
Answer: 1. is not a desert country. 2. is not an oil-rich country. 3. is not a landlocked country.
Human
Human
Human
Data. We consider 50 subjects, 25 encyclopedic entities such as elon musk, and 25 commonsense concepts, such as jogging (Full list in Appendix A). Our intuition behind these choices is diversity: (i) in types, e.g., activities, occupations, people; and (ii) in popularity, e.g., tom cruise (a famous hollywood actor) and peri gilpin (a less known tv actor).
Methods. To compile lists of negative statements about these subjects, we consider: • Text Extractions: The pattern-based method [5] relies on a handful of manually crafted patterns, in the form of why-questions, to extract interesting negative statements from rich query logs, e.g., “why doesn’t amazon..” with the completion “accept paypal”. We instantiate the query-log API with Google and Bing, merge the results, and rank by frequency. • KG Inferences: The peer-based negation inference methodology [3] relies on a given KG to identify highly related entities to the input entity (called peers). Positive statements about these peers are used to infer candidate negatives, which are finally ranked using Positive Meaning 0 0 0.11 0.07 0.09 0.10 0.14 lebanon isn’t devoid of historical sites
Human Sample Statement rabbits can’t vomit the beatles didn’t tour avocado isn’t bad
statistical metrics, such as relative frequency, e.g., “unlike similar physicists, such as max planck and albert einstein, stephen hawking never won the nobel prize in physics”. We instantiate the KGs to Wikidata [14] and Ascent [15], for encyclopedic/commonsense subjects, respectively. • ChatGPT 0-shot: The zero-shot probe introduced in Section 2 is submitted to Chat
GPT [13] (May 2023 version). • ChatGPT k-shot: The few-shot probe in Section 2, with =3 (salient:nonsalient 3:0), is submitted to ChatGPT. • Alpaca 0-shot: The zero-shot probe introduced in Section 2 is submitted to Alpaca-13B, a model fine-tuned from LLaMA on instruction-following demonstrations by Stanford [ 16]. • Alpaca k-shot: The few-shot probe from Section 2, with =3 (salient:nonsalient 3:0), is submitted to Alpaca-13B.
• Human 2: We ask MTurkers to write lists of salient negative statements about a given subject. We show them examples of what a salient negative statement looks like. We collect, for each subject, two lists of statements from two workers. The performance is later measured as the average of the two.
• Correctness: The true negativity (is it actually false?) and factuality of a statement (is it a judgeable statement?), e.g., not an opinion. We allow the labels: correct, incorrect, ambiguous, or positive meaning. Samples are shown in Table 3. • Salience: The unexpectedness, informativeness, or interestingness of a statement. We allow: salient (1), somehow salient (0.5), and nonsalient (0).
Results are annotated on their salience by 2 domain-experts 3, with inter-annotator agreement = 60%. Correctness, the more straight forward metric of the two, was annotated by 1 of the domain-experts.
2We are aware of the risk that workers might use LLMs to generate these statements. In the absence of reliable detection tools on this newly emerging problem, we rely on our personal judgement as well as string matchings to discard untrustworthy answers. In particular, any response that matches the exact wording of one of the responses of the LLM baselines, or any near-duplicates in human-generations, were rejected.
3Experts on the topic of salient negative knowledge at web-scale. sal:nonsal 3:0 0:3 3:3 10:10 True Factuality and Negativity of Statements. Results for correctness are shown in Table 2, and investigated further in Table 3. The KG inferences model ranks first on correctness overall. This is due to the factuality of KG statements. KG triples, especially encyclopedic ones, are expressed using precise and well-defined relations, such as award received. Moreover, they have been curated using manual and automated techniques, and hence, their truthfulness is easy to verify. Moreover, both variants of ChatGPT’s probes perform significantly better than variants of Alpaca on correctness in both domains, with an out-performance of up to 36% in correctness@1. We also notice that, for both Alpaca and ChatGPT, their few-shot probes perform better than the zero-shot probes, with an improvement of 16% for Alpaca and 5% for ChatGPT. Finally, we find that many of the generated statements by humans and LLMs were actually statements with negative keywords but a positive meaning, such as lebanon isn’t devoid of historical sites, with up to 14% of generated statements for the former and 11% for the latter.
Salience of Truly Negative Statements. Results for salience are shown in Table 2. This metric is only computed over (previously annotated) correct statements. The best performances are shared between the KG inferences model and ChatGPT’s few-shot variant. Though not performing comparably well overall, the text extractions model ranks first on salience of encyclopedic subjects @3 and 5. This is especially apparent for prominent entities, which are frequently queried using famous search engines. Again, ChatGPT’s variants significantly outperforms Alpaca’s on the notion of salience, with up to 23% improvement in salience@1, maintaining the same level of quality for both types of subjects. Sample results from all models are shown in Table 1 and Appendix E. An experiment on the quality of generated negatives over two popularity levels, namely prominent and long tail subjects, is in Appendix B. Efect of Value on LLM’s Few-shot Probe. We examine the LLM using diferent numbers of samples, for in-context learning. We consider a subset of 5 entities (3 encyclopedic and 2 commonsense), and assess the performance of the few-shot ChatGPT using diferent values of , with diferent salient:nonsalient ratios. Results are in Table 4. Adding a small but equal number of salient and nonsalient samples (3:3) improves the correctness by 8%, compared to only adding salient samples (3:0), however, at the expense of their salience, which drops by by 14%. Adding only nonsalient samples (0:3) compromises both metrics. Finally, adding a larger but equal number of salient and nonsalient samples (10:10) does not result in any improvements.
In this paper, we perform a systematic evaluation of LLMs’ ability to generate salient negative statements. We assess them against existing method and crowdsourced statements. We find that LLMs’ few-shot probes show promising results in salience@1. Moreover, we find that ChatGPT outperforms Alpaca on this task, in both correctness and salience. One of the remaining limitations, however, is the ability of LLMs to recognize truly negative factual statements, as opposed to ambiguous, or seemingly negative statements with positive meaning. We hope that this study, as well as the following observations, give insights to future researchers on this topic. Prompt Engineering. There is a wide consensus that LLMs are very powerful when you ask them for information in the right manner. In our task, we notice that the wording, especially of the zero-shot probe, changes the results dramatically. For instance, using the expressions negative statements, negated statements, and negation statements returns completely diferent responses. For instance, the probe with the word negated (alone without salient factual) returns obviously true statements with negative keywords added to them, e.g., “stephen hawking was not a physicist”. The probe with the word negative does not return any results, but an apology from the AI about not being able to give bad statements about individuals. On this and other tasks, designing intuitive prompts and studying the ability of LLMs to understand them is the most important part of the process [17].
The Notion of Salient Negation. Assessing the truthfulness of statements is one thing, but assessing the salience of negatives is more challenging. Salience is a subjective metric. For instance, for a Basketball fan, the fact that jordan did not star in the film space jam 2 (the first was built around him), is a big deal. For others, the salience is not obvious. In addition of the expertise of the reader, their nature is also important. In other words, are these negations generated for a human-reader, or to equip machines with better negative knowledge? For instance, what might not appear salient to a human, can be important to improve the reasoning skills of a chat bot. In this study, we assume that the reader is a human, who usually has a higher standard for what is interesting than a machine. Generally, designing experiments should take into consideration downstream applications and information about the end-user. Maintenance. Ideally, models must always keep track of real-world changes which afect the truthfulness of statements, coverage of emerging entities, etc. This is relatively easy in the collaborative knowledge graphs, which are updated on a daily basis. For LLMs, the process of re-training is much more expensive. e.g., in May 2023, ChatGPT still generates the statement brendan fraser has never won an oscar, which is no longer true, due to his win in 2023 (the training of the model has been completed in September 2021). [3] H. Arnaout, S. Razniewski, G. Weikum, Enriching knowledge bases with interesting negative statements, in: AKBC, 2020. [4] T. Safavi, J. Zhu, D. Koutra, NegatER: Unsupervised Discovery of Negatives in Commonsense Knowledge Bases, in: EMNLP, 2021. [5] J. Romero, S. Razniewski, K. Pal, J. Z. Pan, A. Sakhadeo, G. Weikum, Commonsense properties from query logs and question answering forums, in: CIKM, 2019. [6] G. Karagiannis, I. Trummer, S. Jo, S. Khandelwal, X. Wang, C. Yu, Mining an "antiknowledge base" from Wikipedia updates with applications to fact checking and beyond, PVLDB (2019). [7] N. Lee, B. Z. Li, S. Wang, W.-t. Yih, H. Ma, M. Khabsa, Language models as fact checkers?, in: ACL, FEVER workshop, 2020. [8] F. Petroni, T. Rocktäschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, A. Miller, Language models as knowledge bases?, in: EMNLP, 2019. [9] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional transformers for language understanding, in: NAACL, 2019. [10] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, Language Models are
[11] N. Kassner, H. Schütze, Negated and misprimed probes for pretrained language models:
Birds can talk, but cannot fly, in: ACL, 2020. [12] J. Chen, W. Shi, Z. Fu, S. Cheng, L. Li, Y. Xiao, Say what you mean! large language models speak too positively about negative commonsense knowledge, arXiv (2023). [13] OpenAI, Introducing chatgpt, https://openai.com/blog/chatgpt, 2022. [14] D. Vrandečić, M. Krötzsch, Wikidata: a free collaborative knowledge base, CACM (2014). [15] T. Nguyen, S. Razniewski, J. Romero, G. Weikum, Refined commonsense knowledge from large-scale web contents, TKDE (2022). [16] R. Taori, I. Gulrajani, T. Zhang, Y. Dubois, X. Li, C. Guestrin, P. Liang, T. B. Hashimoto,
03/13/alpaca.html, 2023. [17] J. Jang, S. Ye, M. Seo, Can large language models truly understand prompts? a case study with negated prompts, in: Proceedings of The 1st Transfer Learning for Natural Language
We consider 50 subjects of diferent domains, namely commonsense and and of diferent popularity, namely prominent and long tail (see Table 5).
Prominent stephen hawking, michael jordan, lebanon, michelle obama, microsoft, china, amazon, albert einstein, the beatles, elon musk, angela merkel, taxi driver, taj mahal, white house, eat pray love, tom cruise, brendan fraser, the godfather, my cousin vinny, mercedes-benz group, gmc, linkedin Long tail peri gilpin, caramel, ubisoft Commonsense elephant, soup, lawyer, acne, mother, gorilla, pancake, newspaper, jaguar, avocado, garlic, chef, salad, rabbit, jogging, cuflink, strudel, librarian, armchair tabbouleh, breadfruit, kitchenette, hockey stick, basketball court, coffee table
We recompute the quality of negatives (@5) over two levels of subject-popularity, namely prominent and long tail. Figure 1 indicates a significant decrease in both salience and correctness for long tail subjects, for the text-based method; dropping to only 1% on salience. Using query logs as the corpus, users query prominent/trendy subjects much more frequently than long tail ones. We find the human-written statements for both popularity-levels comparable, with a slight advantage for prominent subjects. Similarly, the KG inferences model shows comparable results with a slight advantage of prominent subjects in correctness, and of long tail subjects in salience. Finally, we find an unexpected improvement, for all LLM variants, of long tail subjects over prominent ones, in both metrics. One interpretation could be the large amount of noisy web sources (main data source for training LLMs), about famous entities. For example, tabbouleh (long tail) is a specific instance of salad (prominent). While negatives about the former are more clear-cut, e.g., tabbouleh isn’t made with rice but bulgur, negatives about the latter seem more unfocused, e.g., salad isn’t always a healthy choice.
As shown in Table 3, many of the LLM-generated and crowdsourced statements are in fact positive. Some of the recurring expressions which convey a positive meaning using negative keywords:
CorTreexcttnEexstsractions
ChatGPT 00-shot
KG Infe0.r02e.202n.42c6es00.3.334 0.4H0u.4m050.4a.48n9 0.580.62 0.69C0h.7a6tGPT k-shot
Alp0a.2ca k-shot 0.4 Alpaca 00-s.6hot 0.8 1
Salience Correctness
Amazon did not exclusively focus on selling its own products.
Example: Strudel is not tasty without sugar.
Expression: not just (9) Example: Acne is not just a teenage problem.
Expression: not only (20) Example: Librarians do not only work in public libraries.
Example: Cofee tables are not limited to indoor use.
Example: GMC does not solely operate in the United States. Expression: not all (10)
Example: Not all librarians are women.
In this probe =6 (3:3); LLM=ChatGPT.
A salient factual negated statement about an entity means that the statement doesn’t hold in reality. Moreover, the negated statement is either surprising, unexpected, or useful to the reader. For example: penguins can’t fly. istanbul isn’t the capital of turkey.
tom cruise never won an oscar.
penguins can’t run for presidency. istanbul isn’t the capital of france.
tom cruise never won the nba best player award.
Given this definition and examples, write a list of 3 salient factual negated statement about microsoft.
Answer: 1. is not primarily a dating platform. 2. does not charge users a fee to create an account.
3. does not allow users to post anonymous content.
The following tables show the top results about linkedin, chef, and angela merkel, respectively: Text KG ChatGPT 0-shot ChatGPT k-shot Alpaca 0-shot Alpaca k-shot Human
Top Negative Statements (linkedin) isn’t working isn’t loading isn’t headquartered in san francisco isn’t a software company isn’t designed for sharing personal content doesn’t permit users to buy followers isn’t used for online dating doesn’t allow users to post pictures of their pets doesn’t have a user-friendly interface doesn’t provide any value to its users isn’t a social media platform doesn’t own the content posted on its platform doesn’t have a billion members wasn’t founded by mark zuckerberg Text KG ChatGPT 0-shot ChatGPT k-shot Alpaca 0-shot Alpaca k-shot Human
Top Negative Statements (chef) doesn’t wear hat doesn’t eat their own food doesn’t take orders doesn’t bring drinks didn’t use any garlic didn’t win any cooking competitions doesn’t just cook food not all have formal culinary training don’t need to have an understanding of nutrition don’t need to have good knife skills don’t need to be certified don’t usually work with raw ingredients doesn’t wash the dishes doesn’t always wear the chef’s hat Text ChatGPT 0-shot ChatGPT k-shot Alpaca 0-shot Alpaca k-shot
Top Negative Statements (angela merkel) didn’t listen to donald trump doesn’t deserve to be honoured by germany isn’t on twitter isn’t a lawyer isn’t a native german speaker didn’t originally pursue a career in politics has never been married is not a member of the SPD isn’t a member of the CDU isn’t a scientist isn’t the first female chancellor of germany isn’t from east germany