What is in a name? Mitigating Name Bias in Text Embeddings via Anonymization

Pannaga Shivaswamy; Sahil Manchanda

REVIEW 3 major objections 4 minor 13 references

What is in a name? Mitigating Name Bias in Text Embeddings via Anonymization

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Text-embedding models rank same-name different-plot pairs as more similar than same-plot different-name pairs; deleting names at inference restores semantic ordering.

desk verdict Real phenomenon, plausible mitigation, but the evidence for the mitigation is thinner than the claims—needs a non-name deletion control. read the letter →

arxiv 2502.02903 v1 pith:Q2IQSOKJ submitted 2025-02-05 cs.CL cs.AIcs.LG

Sahil Manchanda , Pannaga Shivaswamy This is my paper · ORCID

classification cs.CLcs.AIcs.LG

keywords namebiastextembeddingsanonymizationsemanticsimilaritynamedentityperturbationinference-timedebiasingcosine

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that text-embedding models are systematically biased by named entities: swapping person or country names inside a passage moves the embedding even when the passage's meaning is unchanged, and in the authors' tests most models rated a same-name/different-plot story as more similar to a query than a different-name/same-meaning story. The authors measure this by creating twenty reworded copies of each movie or book summary with names replaced, then averaging pairwise cosine similarity across the copies; a bias-free model should score close to 1, while observed averages range roughly from 0.68 to 0.89. They then propose inference-time anonymization, deleting proper-noun person and location names with an LLM prompt before embedding, which needs no retraining. On a ten-triplet semantic similarity task, anonymization lifts AUC-ROC from mostly below 0.5 (worse than chance) to roughly 0.94 to 1.0, and it improves correlation with human summary-quality judgments on a 1600-item evaluation. The practical point is that entity-agnostic text matching in search, clustering, or plagiarism detection can be made more reliable by dropping names before computing embeddings.

What carries the argument

The load-bearing object is the perturbation benchmark (Algorithm 1) paired with the anonymization pipeline. The benchmark formalizes name bias as any deviation in average pairwise cosine similarity among K=20 embeddings of texts that differ only by randomly swapped person and country names; the expected value under a name-blind, meaning-only model is 1. The remedy is inference-time anonymization: an LLM prompt deletes all proper-noun person, city, country, and organization names, leaving pronouns and structure, and the anonymized text is embedded. The identity that cosine similarity between anonymized texts of entity-only variants equals 1 is what makes the method's success testable.

What would settle it

Collect human similarity judgments on pairs of passages that differ only in named entities but where the entity is semantically informative (for example, 'Ganga and Yamuna are mighty rivers' versus 'Yangtze is a mighty river'); if raters judge these pairs as less than fully similar, the measured cosine drop reflects real semantic content, and the paper's bias metric would need to be corrected for entity-borne meaning before anonymization can be credited with removing bias rather than information.

Watch

Extended reading notes

Core claim

The paper's central claim is that name bias, not just gender or racial bias, distorts text embeddings: the identity tokens in a text can dominate the embedding's position in semantic space. The authors establish this with a perturbation benchmark: for each text, replace every person and country name with randomly sampled alternatives, generating K=20 perturbed copies, embed all copies, and average the pairwise cosine similarity. If embeddings tracked meaning, the average would be close to 1; across fourteen open and commercial embedding models on movie and book plot summaries the averages fall between roughly 0.68 and 0.89, and a paired illustration shows same-name opposite-meaning stories ranked more similar than different-name same-meaning stories. They then show that applying an LLM-based anonymization step, which removes proper-noun person, location, and organization names at inference time, makes embeddings of entity-only variants nearly identical (cosine similarity 1) and improves two downstream evaluations: a binary semantic-similarity task where most models went from below-chance AUC to near 1, and a graded machine-summary relevance task where Spearman and Pearson correlations with human scores increased for nearly all models.

Load-bearing premise

The experimental measure treats any drop in cosine similarity after replacing names as bias, assuming a correct model should rate texts identical when only their named entities differ; this assumption is strained when a name itself carries legitimate topical or geographic meaning.

Editorial extensions

If this is right

Anonymized embeddings make near-duplicate documents that use different names map to almost the same vector, so duplicate detection, clustering, and plagiarism screening become less sensitive to entity substitutions.
Retrieval systems that index anonymized passages will rank a query higher against a same-meaning, different-name document than the raw embeddings currently do, fixing the reversed ordering the paper documents.
Because anonymization needs no retraining, it can be applied on top of any existing embedding model at inference time, including commercial APIs that cannot be fine-tuned.
Removing names before embedding makes summary-quality evaluation align more closely with human relevance judgments, giving automatic evaluation a better proxy for content quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If name bias is as universal as the measurements suggest, any embedding-based system that ingests news or social media, where the same event is described with different person and place names, will systematically under-link equivalent texts; indexing an anonymized copy alongside the original could recover those links.
The same anonymization trick might be adapted as a training-time augmentation: randomly deleting named entities during contrastive training could force models to rely on structure rather than identity, a testable extension the paper only gestures at in its conclusion.
A practical boundary is that anonymization is only safe when entity identity is not the signal of interest; for queries about a specific person or place, the proposed fix would destroy the relevant semantics, so deployment likely needs task-aware selective anonymization.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear [name],

Quick take: the paper finds a real, underappreciated phenomenon—text embeddings are surprisingly sensitive to the names of people and places in a text—and it proposes a simple inference-time fix: delete the names before embedding. That part is worth knowing. But the evaluation of the fix is much weaker than the claims in the abstract, and the central causal claim about anonymization is not yet established.

What's genuinely new: a systematic measurement of name sensitivity across 15 embedding models using K=20 name perturbations, with heatmaps. The drop in pairwise cosine similarity from near 1 to 0.68–0.78 on CMU movie/book plots is a real signal. The anonymization approach is training-free and easy to deploy, and the paper is honest about its limitations.

Where it gets soft. First, the STS task is built from 10 hand-authored triplets where the negatives share names with the query by construction. When the task is defined by names being irrelevant, deleting names will fix it; the near-perfect AUC after anonymization is partly definitional. Second, the SummEval experiment lacks any control that removes a comparable amount of non-name content. Removing names shortens and genericizes every text, which can inflate correlations with human relevance even if name deletion is irrelevant. The paper even shows voyage-3-lite degrading after anonymization, and some anonymized examples are ungrammatical fragments, which suggests the gains could come from degraded text being more similar to itself, not from any name-specific debiasing. Third, the benchmarking assumption that a bias-free model should treat "Alejandro" and "Jelani" as interchangeable ignores the legitimate topical signal carried by some names, especially country names in plot summaries. Part of the measured drop may be correct semantics, not bias.

The stress-test note is right: without a matched non-name deletion control, the causal claim that anonymization specifically mitigates name bias is unproven. The existence of name sensitivity is supported; the mitigation's specific mechanism is not.

Who is this for? Anyone working on embedding fairness, retrieval, or evaluation pipelines. The paper is a useful red flag. But I would not cite it as evidence for the mitigation without the control experiment.

Recommendation: send it to peer review, but be prepared to require the control, a larger naturally occurring evaluation set, and a more measured abstract.

Referee Report

3 major / 4 minor

Summary. The paper proposes that text-embedding models exhibit a 'name bias': cosine similarity between embeddings can be driven by personal and geographical names rather than by semantics. It supports this with a perturbation benchmark (Sec. 3) in which person and country names are randomly replaced and the average pairwise cosine similarity across perturbed texts is measured; a low value is interpreted as name bias. The paper then proposes inference-time anonymization (Sec. 4): deleting person and location names from text using an LLM prompt before embedding. The method is evaluated on a 10-triplet binary semantic-similarity task (Sec. 5.1) and on the SummEval summary-quality correlation task (Sec. 5.2). The paper reports large AUC improvements after anonymization on the STS task and mixed but mostly positive correlation improvements on SummEval.

Significance. If the causal claim were established, this would be a practically interesting result: the mitigation is training-free, requires no fine-tuning, and can be applied at inference time with a simple prompt. The paper also provides useful reproducible materials: the full 10-triplet dataset in Appendix D, the perturbation name lists in Appendix A, the anonymization prompts in Table 5, and similarity heatmaps in Appendix C. However, as it stands, the evidence mainly demonstrates that embeddings are lexically sensitive to named entities; whether this constitutes a semantic bias, and whether anonymization specifically mitigates that bias rather than merely shortening or degrading the text, is not yet established. The SummEval results are also mixed, with at least one model degrading on both correlation measures and another degrading on Pearson, which weakens the broad claim of 'significant performance gains' in the abstract and conclusion.

major comments (3)

[Sec. 5.1, Table 6 and Appendix D] The STS benchmark is constructed so that the negative paragraph shares the same person and location names as the query while the positive paragraph shares meaning but not names. Under any token-sensitive embedding model, the low original AUC is largely expected from this construction; the anonymized condition then removes exactly the tokens that the 'Identical Names' column shows are sufficient for correct ranking. This before/after comparison therefore does not isolate name bias from generic lexical overlap, and the dataset is small (10 hand-authored triplets) with no per-triplet variability reported. Please add a non-name lexical-overlap control (e.g., deleting or replacing an equal number of non-name content words) and report per-triplet AUCs or a bootstrap interval over the 10 triplets, rather than means over two LLM anonymization runs.
[Sec. 5.2, Table 8] The SummEval evaluation compares original vs. anonymized text only; there is no control condition that removes an equivalent amount of non-name content. LLM-based anonymization deletes proper-noun tokens and thereby shortens and genericizes both machine and human summaries, which can inflate correlations with human relevance scores even if the active ingredient is text shortening rather than name removal. The results are also not uniformly supportive of the paper's claim: voyage-3-lite degrades on both Spearman and Pearson after anonymization, and gemini degrades on Pearson. Please add a matched deletion control (removing the same number of non-name tokens) or a placeholder-replacement control on this task, and report significance or confidence intervals for the before/after differences.
[Sec. 3.1 and Tables 3, 4] The benchmarking metric treats a cosine similarity of 1.0 after name replacement as the bias-free ideal. This conflates two very different effects: replacing a person name may test token-level identity sensitivity, but replacing a country name (e.g., Belgium with France in Table 2) can legitimately alter the geographical or topical semantics of a movie or book plot. The main Tables 3 and 4 mix both types of replacement, and the filtering described in the table footnotes (excluding city/town/village/nationality keywords) does not remove country names, which are themselves topical. The person-only results in Appendix B are a better design, but they are not integrated into the main claim. Please report the person-only and country-only perturbation results side by side, or use non-semantic placeholders, so that the measured deviation from 1.0 can be attributed to name identity rather than to legitimate geographic content.

minor comments (4)

[Sec. 6] The phrase 'deanonymized text embeddings' is misleading; the comparison is between original and anonymized embeddings, and the original embeddings are not 'deanonymized'. Please use 'original' or 'non-anonymized'.
[Table 7 / Appendix D] The anonymized examples contain ungrammatical fragments such as 'and are two mighty rivers' and 'is a mighty river'. Because the proposed method produces such degraded text, the paper should explicitly discuss how this output quality affects embedding variability and whether downstream users should expect fluency loss as a cost of anonymization.
[References] Several API-based models are cited without version identifiers or retrieval dates (OpenAI Embeddings, Voyage AI, Gemini), which limits reproducibility as these services change over time. Please add model versions and access dates.
[General readability] There are repeated spacing and grammar issues, such as 'V oyage AI' in the references and the sentence in the introduction that reads 'the issue is not confined to certain cultures, cross-culture, but is universal'; a careful proofread is needed.

Circularity Check

1 steps flagged · score 2.0 of 10

One self-definitional bias metric; downstream mitigation evidence is independent, so overall circularity is low.

self definitional [Section 3.1 (Benchmarking Methodology) and Section 3.4 (Analyzing Bias), Tables 3-4]
"A higher average similarity indicates that the perturbed texts are closer to each other in the semantic space, suggesting less deviation. Conversely, a lower average similarity score suggests ... It suggests that the embedding model exhibits a bias towards names in the text. ... In Table 3 and 4 we observe a significant deviation in the average cosine similarity which should be close to one if the cosine similarity captured the real semantic similarity rather than information in names present in the text."

The paper's operational definition of 'name bias' is the drop in average cosine similarity between K name-perturbed copies of the same text. The later 'demonstration' that models exhibit name bias then reads directly off this same defined quantity: any model that is not perfectly invariant to token substitution scores below 1. No independent, externally validated ground truth is introduced to calibrate what fraction of the drop is bias versus legitimate semantic sensitivity (e.g., country names carrying topical information in plot summaries). Thus the headline evidence in Tables 3-4 is a restatement of the metric's threshold rather than a derived result.

full rationale

Most of the paper's central mitigation claim is not circular. No parameters are fitted to the target results, no self-citation chain is load-bearing, and the two downstream evaluations rest on external or controlled evidence: SummEval uses human relevance judgments, and the 'Identical Names' column in Table 6 provides a within-task control showing that models can rank correctly when names are held constant. The anonymization gains in Tables 6 and 8 are empirical outcomes, not mathematically forced by the experimental setup. The one self-definitional element is the bias benchmark itself: the paper defines name bias as deviation from cosine similarity 1 under name perturbation and then reports that deviation as evidence of the bias. This is a legitimate operationalization, but it makes the absolute existence claim partly tautological. The absence of a matched non-name-deletion control in Section 5.2 is a validity concern rather than a circularity, and the paper itself acknowledges in Limitations that similarity is only a proxy for thematic semantics. Overall, the derivation chain is largely self-contained, so the circularity score is low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the assumption that names are semantically irrelevant to theme in the tested texts, that cosine similarity is a valid semantic proxy, and that LLM anonymization preserves meaning. No free parameters are fitted to make the results work, but several experimental choices (K, name lists, dataset filters, prompts, aggregation rule) are unshared and influence the measured effect sizes.

free parameters (5)

Number of perturbations K = 20
K=20 is chosen in Sec 3.1 to compute average pairwise cosine similarity; the value affects the stability of the bias estimate and is not justified or varied.
Dataset selection filters = plots < 250 words; contain both person and country names; exclude city/town/nationality keywords
Sec 3.3 defines the filtered CMU subsets but does not report how many samples remain or how sensitive the results are to these filters.
Name replacement list = hand-curated list in Table 9 (hundreds of names)
Results depend on which names are used; name frequencies and cultural associations may influence embedding distances.
Anonymization prompts = two hand-written prompts in Table 5
The quality of anonymization and downstream gains depend on prompt wording and on the proprietary LLM (gemini or claude) used; no systematic comparison is provided.
Summary prediction rule = max cosine similarity to any human summary
Adopted from MTEB for SummEval in Sec 5.2; a different aggregation (e.g., mean) would change the correlations.

assumptions (4)

domain assumption Cosine similarity between text embeddings is a valid proxy for semantic similarity.
Used to define the bias metric in Sec 3.1 and to score downstream tasks in Sec 5; the Limitations section acknowledges that cosine similarity may overlook deeper semantics.
ad hoc to paper Names are semantically irrelevant to the theme in the benchmark texts, so a bias-free model should give cosine similarity 1.0 after name replacement.
This assumption underlies the interpretation of Tables 3 and 4 in Sec 3.4; it is reasonable for the constructed examples but questionable for movie and book plots where names can carry geographical or cultural signal.
domain assumption LLM-based anonymization removes exactly the named entities and preserves the core theme of the text.
Sec 4 and 5 rely on the outputs of gemini and claude to produce anonymized texts; the paper provides prompts but no quality audit of the anonymized outputs.
domain assumption SummEval human relevance scores are a valid ground truth for summary quality.
Adopted from Muennighoff et al. (2022) in Sec 5.2 without reassessment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of What is in a name? Mitigating Name Bias in Text Embeddings via Anonymization." pith.science (2026). https://pith.science/paper/Q2IQSOKJ

@misc{pith2026250202903,
  author       = {Pith},
  title        = {Pith review of: What is in a name? Mitigating Name Bias in Text Embeddings via Anonymization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q2IQSOKJ}},
  note         = {Machine review of arXiv:2502.02903}
}

abstract

Text-embedding models often exhibit biases arising from the data on which they are trained. In this paper, we examine a hitherto unexplored bias in text-embeddings: bias arising from the presence of $\textit{names}$ such as persons, locations, organizations etc. in the text. Our study shows how the presence of $\textit{name-bias}$ in text-embedding models can potentially lead to erroneous conclusions in assessment of thematic similarity.Text-embeddings can mistakenly indicate similarity between texts based on names in the text, even when their actual semantic content has no similarity or indicate dissimilarity simply because of the names in the text even when the texts match semantically. We first demonstrate the presence of name bias in different text-embedding models and then propose $\textit{text-anonymization}$ during inference which involves removing references to names, while preserving the core theme of the text. The efficacy of the anonymization approach is demonstrated on two downstream NLP tasks, achieving significant performance gains. Our simple and training-optimization-free approach offers a practical and easily implementable solution to mitigate name bias.

Figures

Figures reproduced from arXiv: 2502.02903 by the authors.

**Figure 2.** Cosine Similarity Heatmap with Gemini model for example in Sec. [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗

**Figure 3.** Cosine Similarity Heatmap with all-mpnet-base-v2 model for example in Sec. C 18 [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Cosine Similarity Heatmap with text-embedding-3-large(Open AI) model for example in Sec. C 19 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages

[1]

The city’s hustle and bustle couldn’t dim the spark between them

Query: Nikolai and Deborah met on a rainy Tuesday in New York. The city’s hustle and bustle couldn’t dim the spark between them. Deborah, with her radiant smile and infectious laughter, had captured Nikolai’s heart from the moment he saw her. Nikolai, a charming and witty gentleman, returned her affection with equal fervor. • Positive: Kashvi and Oluwafem...

work page
[2]

He was eager to have a glass of cold drink

Query: Alejandro quickly ran to the store to buy a cold drink. He was eager to have a glass of cold drink. • Positive: Quickly, Hiroki dashed to the local market to procure some cold drinks. He was yearning for a chilled glass of cold drink. • Negative: Alejandro has stopped buying cold drinks from market. He only drinks cold drinks made at home

work page
[3]

Their bond was unbreakable, a love that transcended all ob- stacles

Query: Mayatoshi and Alex had a deep, pas- sionate love for each other. Their bond was unbreakable, a love that transcended all ob- stacles. They shared dreams, hopes, and aspi- rations, and their love was the foundation of their happiness. • Positive: Priyanka and Yuan were deeply in love. Their affection for each other was profound and unwavering. They ...

work page
[4]

Amazon’s main business is online shopping and Apple is a phone maker giant • Positive: Alibaba and Xiaomi are two Chinese corporations

Query: Amazon and Apple are two Ameri- can corporations. Amazon’s main business is online shopping and Apple is a phone maker giant • Positive: Alibaba and Xiaomi are two Chinese corporations. Alibaba’s main business is online shopping and Xiaomi is a producer of phones • Negative: Amazon is a river in South America. Apples are not grown in the Amazon basin

work page
[5]

They are lifelines for millions of people in the region

Query: Ganga and Yamuna are two mighty rivers. They are lifelines for millions of people in the region. • Positive: Yangtze is a mighty river. It is a long river and is the lifeline for millions of people in the region. • Negative: Ganga and Yamuna are two sisters. They had their schooling in the region and schooling provided a lifeline for them

work page
[6]

Recently, Alice lent Bob a signifi- cant sum of money

Query: Alice and Bob often helped each other financially. Recently, Alice lent Bob a signifi- cant sum of money. Bob promised to return it soon. • Positive: Yuri and Haruto frequently helped each other out, including with money. Lately, Yuri had loaned Haruto a substantial amount of money, which Haruto assured her he’d repay promptly. • Negative: Alice an...

work page
[7]

John is determined to prove Mike’s innocence and secure his acquittal

Query: John, a renowned lawyer, is defend- ing his client, Mike, who is accused of a seri- ous crime. John is determined to prove Mike’s innocence and secure his acquittal. • Positive: Armaan, a man falsely accused of a heinous crime, is relying on his skilled lawyer, Udit, to exonerate him. Udit is committed to presenting a strong defense and clearing Ar...

work page
[8]

Alexander, a seasoned physician, meticulously analyzed patient Sarah’s intri- cate heart condition

Query: Dr. Alexander, a seasoned physician, meticulously analyzed patient Sarah’s intri- cate heart condition. He prescribed a tailored regimen of medications and rigorous lifestyle modifications to significantly improve her car- diac health. • Positive: The esteemed doctor, Dr. Yerusha, conducted a thorough assess- ment of patient Reyan’s complex symp- t...

work page

Show all 13 references

[9]

Smith, a dedicated teacher, guided his students, including the bright young minds of Miller and Pristina, towards academic excellence

Query: Mr. Smith, a dedicated teacher, guided his students, including the bright young minds of Miller and Pristina, towards academic excellence. • Positive: Mr. Yang, a committed edu- cator, mentored his students, including the talented Shruti and Ren, to achieve academic suc...
[10]

CHAR_A” / “CHARACTER_B

Query: Martinez gently examined the injured bird. He gave it food. • Positive: Yohan tenderly inspected the wounded bird and gave it a meal to eat. • Negative: The skilled hunter Martinez tracked the injured bird. He captured it for food. E Example of Semantic Similarity post-...
[2013]

In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 352–361

Learning latent personas of film characters. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 352–361. David Bamman and Noah A Smith. 2013. New align- ment methods for discriminative book summarization. a...

2013 arXiv
[2019]

you gotta be a doctor, lin

Putting evaluation in context: Contextual em- beddings improve machine translation evaluation. In Proceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 2799– 2808. Chandler May, Alex Wang, Shikha Bordia, Samuel R Bowman, and Rachel Ru...

2019 arXiv
[2024]

Computational Linguistics, pages 1–79

Bias and fairness in large language models: A survey. Computational Linguistics, pages 1–79. Yue Guo, Yi Yang, and Ahmed Abbasi. 2022. Auto- debias: Debiasing masked language models with automated biased prompts. In Proceedings of the 60th Annual Meeting of the Association for...

2022 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

The city’s hustle and bustle couldn’t dim the spark between them

Query: Nikolai and Deborah met on a rainy Tuesday in New York. The city’s hustle and bustle couldn’t dim the spark between them. Deborah, with her radiant smile and infectious laughter, had captured Nikolai’s heart from the moment he saw her. Nikolai, a charming and witty gentleman, returned her affection with equal fervor. • Positive: Kashvi and Oluwafem...

work page

[2] [2]

He was eager to have a glass of cold drink

Query: Alejandro quickly ran to the store to buy a cold drink. He was eager to have a glass of cold drink. • Positive: Quickly, Hiroki dashed to the local market to procure some cold drinks. He was yearning for a chilled glass of cold drink. • Negative: Alejandro has stopped buying cold drinks from market. He only drinks cold drinks made at home

work page

[3] [3]

Their bond was unbreakable, a love that transcended all ob- stacles

Query: Mayatoshi and Alex had a deep, pas- sionate love for each other. Their bond was unbreakable, a love that transcended all ob- stacles. They shared dreams, hopes, and aspi- rations, and their love was the foundation of their happiness. • Positive: Priyanka and Yuan were deeply in love. Their affection for each other was profound and unwavering. They ...

work page

[4] [4]

Amazon’s main business is online shopping and Apple is a phone maker giant • Positive: Alibaba and Xiaomi are two Chinese corporations

Query: Amazon and Apple are two Ameri- can corporations. Amazon’s main business is online shopping and Apple is a phone maker giant • Positive: Alibaba and Xiaomi are two Chinese corporations. Alibaba’s main business is online shopping and Xiaomi is a producer of phones • Negative: Amazon is a river in South America. Apples are not grown in the Amazon basin

work page

[5] [5]

They are lifelines for millions of people in the region

Query: Ganga and Yamuna are two mighty rivers. They are lifelines for millions of people in the region. • Positive: Yangtze is a mighty river. It is a long river and is the lifeline for millions of people in the region. • Negative: Ganga and Yamuna are two sisters. They had their schooling in the region and schooling provided a lifeline for them

work page

[6] [6]

Recently, Alice lent Bob a signifi- cant sum of money

Query: Alice and Bob often helped each other financially. Recently, Alice lent Bob a signifi- cant sum of money. Bob promised to return it soon. • Positive: Yuri and Haruto frequently helped each other out, including with money. Lately, Yuri had loaned Haruto a substantial amount of money, which Haruto assured her he’d repay promptly. • Negative: Alice an...

work page

[7] [7]

John is determined to prove Mike’s innocence and secure his acquittal

Query: John, a renowned lawyer, is defend- ing his client, Mike, who is accused of a seri- ous crime. John is determined to prove Mike’s innocence and secure his acquittal. • Positive: Armaan, a man falsely accused of a heinous crime, is relying on his skilled lawyer, Udit, to exonerate him. Udit is committed to presenting a strong defense and clearing Ar...

work page

[8] [8]

Alexander, a seasoned physician, meticulously analyzed patient Sarah’s intri- cate heart condition

Query: Dr. Alexander, a seasoned physician, meticulously analyzed patient Sarah’s intri- cate heart condition. He prescribed a tailored regimen of medications and rigorous lifestyle modifications to significantly improve her car- diac health. • Positive: The esteemed doctor, Dr. Yerusha, conducted a thorough assess- ment of patient Reyan’s complex symp- t...

work page