GeniL: A Multilingual Dataset on Generalizing Language

Summary

This submission introduces a corpus for testing (stereotype) generalization in multiple languages and builds classifers to detect generalizing language, demonstrating clear improvement over template- and cooccurrence-based approaches.

Rating

8

Confidence

4

Ethics flag

1

Reasons to accept

1. A clear, well-motivated and well-grounded definition of the generalizing language task, covering different usage. 2. A novel dataset operationalizing this task definition, using native speakers as annotators (not LLMs!), across 9 languages, with interesting insight along the way, e.g. regarding how prevalent generalizing language is. The dataset will be released with multiple individual annotation to encourage further analysis. 3. Multilingual generalizing language detectors, providing a reasonably high performance baseline on most languages for this task.

Reasons to reject

There is clearly some issue with either the data or the annotation in some languages that cause a large disparity in performance and would warrant further analysis. The last paragraph of Section 5 would deserve more detail and clarity. Apart from that, this is a clear paper with a well defined goal, a valuable contribution, and well-done experiments providing interesting and non-trivial insight; I see really no reason to reject.

Questions to authors

Assuming guidelines were applied similarly in each language, how do you explain the large disparity between languages in the analysis and results (e.g. Fig. 3 or Tab. 3). I would recommend using box-and-whiskers plots in Figure 2 rather than mean and standard deviation, which are not that informative for strongly skewed distributions. I was confused by "respective societies" in Sec. 4.2 -- can you clarify or rephrase? Also later in that Section: "that that", "a stereotypical associations" [Read reviews and rebuttal -- thanks]

Summary

In this paper, the authors present a new dataset and task for "Generalizing Language," dubbed GeniL. GeniL consists of 50k sentences annotated for _generalization_. For each sentence, generalization is determined given an association pair of (_identity_, and _attribute_). The association pairs were extracted from "Multilingual SeeGULL" and the sentences containing those pairs were queried from multilingual Common Crawl. Human annotators were then asked to tag each sentence for generalization using three tags: **P**romoting **G**eneralization (PG), **M**entioning **G**eneralization (MG), and **N**ot **G**eneralizing (NG). Both PG and MG indicate generalization. GeniL is a multilingual dataset spanning **nine** languages from different language families. While the primary contribution of this paper is the dataset, the authors also introduced a generalization classification task and benchmarked it. The main task is to classify generalization regardless of its type (PG or MG). The major finding of this paper is the importance of the sentential context of the association pair in determining generalizations, let alone the detection of stereotypes. The authors demonstrated this by contrasting the annotator labels with previously existing stereotype labels. The authors showed that the existence of the association in the sentence rarely entails a generalization, and hence, current stereotype classification models may inflate such classes by introducing false positives.

Rating

9

Confidence

4

Ethics flag

1

Reasons to accept

- The paper is very well written and easy to follow. - Contributions and novelty can seen clearly from the text. - The different views of the data showcase its value. - The benchmarking methodologies seem to be sound and convincing.

Reasons to reject

There are no major weaknesses or reject-warranting reasons. However, I list some points that may be concerning. - Some of the discrepancies in the inter-annotator agreement for a language like Arabic may be directly correlated with the diversity of the cultures and regions it is spoken in, similar to English in that sense. It would be useful also to include the primary place of residence/childhood home for the annotators and try to balance annotators based on that. - The translation model used to translate the data isn't mentioned in the paper. - In the discussion, the authors said " ... task itself is harder in these languages for both humans and classifiers." How can we know it is hard for the classifiers if the evaluation data has already low IAA? In other words, we can deduce that it is hard for the classifiers if our evaluation data is consistent. (Maybe this is what the authors meant)

Questions to authors

- It may be computationally consuming, but I believe part of a more comprehensive benchmark is to fine-tune the models for every language separately and evaluate them the same way as the English one was evaluated. - Have the authors looked at the syntactic structures of the datasets? There could be useful insights regarding the correlation between certain structures and generalizations. The same goes for the correlation with the [committed] belief of the author and whether it indicates the generalization to be PG or MG.

Summary

This paper focuses on the challenge of detecting stereotypes in LLM-generated language by studying the expression of generalizations. It proposes a multilingual dataset GeniL that is annotated for generalizations. The experimental results show that the mere co-occurrence of an identity term and attribute being a generalization is generally unlikely and varies by language, identity group, and attribute. The authors also train classifiers on the proposed dataset to improve the performance of generalization detection.

Rating

6

Confidence

3

Ethics flag

1

Reasons to accept

This paper proposes a multilingual dataset for generalization, which may be useful for future research. The experiments and some analyses are interesting, although further experiments are necessary to support the findings.

Reasons to reject

(1) The technological impact of this paper is limited. (2) I have some concerns about the GeniL dataset. The annotations are subjective. For example, the labels of the examples presented in the appendix are controversial, and the annotator agreement is weak for some languages. These inconsistency could undermine the reliability of analyses in the paper. It would be better if more experiments could be conducted to support the claims. (3) Relevant to the previous point, the purpose of the GeniL dataset may appear unclear, given that the same association pairs occur in various contexts. Why not analyze and label the generalizations within the existing SGM dataset? If extending to a broader corpus such as mC4, why not extract pairs of identity and attributes from scratch, instead of querying sentences by using the existing (potentially biased) pairs from SGM? (4) The method to create GeniL dataset using the existing stereotype dataset might introduce bias into some conclusions. It would be better to see some discussions on this aspect. (5) The authors claim to propose the new task of detecting generalization to help identify stereotypes. However, there is limited evidence showing that detecting generalization does help.

Questions to authors

(1) On Page 6, how is the factor of "10" derived in "any co-occurrence based approaches to estimate the extent of generalization will be over-estimating by a factor of 10 or more."? (2) It is interesting in Figure 3 that stereotypical associations are much more frequent in non-generalizing contexts. What could be the possible reasons?

Thanks

Thank you for the replies -- looking forward to reading the final version. I am slightly puzzled by your comment on the boxplots: In that case, the box should go from zero to the 3rd quartile, with whisker extending above, but not below, which would arguably give a more accurate idea of the distribution than error bars extending largely below zero.

Thanks for your response!

Thanks for the response and clarification! It clears up several of my concerns, so I have increased my scores. However, I still have some reservations about the consistency of the experiments and the collected dataset.