More Like this

1. What might books be teaching young children about gender?

   https://doi.org/10.31234/osf.io/ntgfe  

   Lewis, M.; Borkenhagen, M.C.; Converse, E.; Lupyan, G.; Seidenberg, M.S. (January 2022, Psychological science)

   We investigated how gender is represented in children’s books using a novel 200,000 word corpus comprising 247 popular, contemporary books for young children. Using human judgments and word co-occurrence data, we quantified gender biases of words in individual books and in the whole corpus. We find that children’s books contain many words that adults judge as gendered. Semantic analyses based on co-occurrence data yielded word clusters related to gender stereotypes (e.g., feminine: emotions; masculine: tools). Co-occurrence data also indicate that many books instantiate gender stereotypes identified in other research (e.g., girls are better at reading and boys at math). Finally, we used large-scale data to estimate the gender distribution of the audience for individual books, and find that children are more often exposed to gender stereotypes for their own gender. Together the data suggest that children’s books may be an early source of gender associations and stereotypes. 

   more » « less
2. Data mixture inference attack: BPE tokenizers reveal training data compositions

   Hayase, Jonathan; Liu, Alisa; Choi, Yejin; Oh, Sewoong; Smith, Noah A (December 2024, Advances in Neural Information Processing Systems)

   The pretraining data of today's strongest language models remains opaque, even when their parameters are open-sourced. In particular, little is known about the proportions of different domains, languages, or code represented in the data. While a long line of membership inference attacks aim to identify training examples on an instance level, they do not extend easily to global statistics about the corpus. In this work, we tackle a task which we call data mixture inference, which aims to uncover the distributional make-up of the pretraining data. We introduce a novel attack based on a previously overlooked source of information—byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered vocabulary learned by a BPE tokenizer naturally reveals information about the token frequencies in its training data: the first token is the most common byte pair, the second is the most common pair after merging the first token, and so on. Given a tokenizer's merge list along with data samples for each category of interest (eg, different natural languages), we formulate a linear program that solves for the relative proportion of each category in the tokenizer's training set. Importantly, to the extent to which tokenizer training data is representative of the pretraining data, we indirectly learn about the pretraining data. In controlled experiments, we show that our attack can recover mixture ratios with high precision for tokenizers trained on known mixtures of natural languages, programming languages, and data sources. We then apply our approach to off-the-shelf tokenizers released alongside recent LMs. We confirm much publicly disclosed information about these models, and also make several new inferences: GPT-4o is much more multilingual than its predecessors, training on 10x more non-English data than GPT-3.5, Llama 3 and Claude are trained on predominantly code, and many recent models are trained on 7-16% books. We hope our work sheds light on current design practices for pretraining data, and inspires continued research into data mixture inference for LMs. 

   more » « less
3. A Stylometric Application of Large Language Models

   Stropkay, Harrison F; Chen, Jiayi; Latifi, Mohammad J; Rockmore, Daniel N; Manning, Jeremy R (October 2025, arXivorg)

   We show that large language models (LLMs) can be used to distinguish the writings of different authors. Specifically, an individual GPT-2 model, trained from scratch on the works of one author, will predict held-out text from that author more accurately than held-out text from other authors. We suggest that, in this way, a model trained on one author's works embodies the unique writing style of that author. We first demonstrate our approach on books written by eight different (known) authors. We also use this approach to confirm R. P. Thompson's authorship of the well-studied 15th book of the Oz series, originally attributed to F. L. Baum. 

   more » « less
4. corpus and behavioral analysis of the quality of young children’s STEM books.

   Miller, H.E. (January 2020, Proceedings of the 42nd Annual Conference of the Cognitive Science Society)

   Promoting early STEM knowledge helps to prepare children for formal schooling. Shared book reading may promote early STEM knowledge. This research examined the quality of available STEM books in children’s environments and investigated how such books influenced children’s learning in shared book reading contexts. In Study 1, we used both meaning-based human coding and computerized latent semantic analysis to categorize books based on the extent to which they provided support for encoding and demand for active recall. We found similarity in the ratings using the two approaches. Most books fell into categories characterized by low Support and Demand. In Study 2, we found that 4- to 5-year-olds learned more STEM facts when books were high in Support and/or Demand, although few books fell into those categories. This research highlights the importance that textual features of books play in promoting early STEM knowledge during shared book reading. 

   more » « less
5. A Unified Approach to Online Matching with Conflict-Aware Constraints

   Xu, Pan; Shi, Yexuan; Cheng, Hao; Dickerson, John; Sankararaman, Karthik; Tong, Yongxin; Srinivasan, Aravind; Tsepenekas, Leonidas (January 2019, AAAI Conference on Artificial Intelligence (AAAI))

   Online bipartite matching and allocation models are widely used to analyze and design markets such as Internet advertising, online labor, and crowdsourcing. Traditionally, vertices on one side of the market are fixed and known a priori, while vertices on the other side arrive online and are matched by a central agent to the offline side. The issue of possible conflicts among offline agents emerges in various real scenarios when we need to match each online agent with a set of offline agents. For example, in event-based social networks (e.g., Meetup), offline events conflict for some users since they will be unable to attend mutually-distant events at proximate times; in advertising markets, two competing firms may prefer not to be shown to one user simultaneously; and in online recommendation systems (e.g., Amazon Books), books of the same type “conflict” with each other in some sense due to the diversity requirement for each online buyer. The conflict nature inherent among certain offline agents raises significant challenges in both modeling and online algorithm design. In this paper, we propose a unifying model, generalizing the conflict models proposed in (She et al., TKDE 2016) and (Chen et al., TKDE 16). Our model can capture not only a broad class of conflict constraints on the offline side (which is even allowed to be sensitive to each online agent), but also allows a general arrival pattern for the online side (which is allowed to change over the online phase). We propose an efficient linear programming (LP) based online algorithm and prove theoretically that it has nearly-optimal online performance. Additionally, we propose two LP-based heuristics and test them against two natural baselines on both real and synthetic datasets. Our LP-based heuristics experimentally dominate the baseline algorithms, aligning with our theoretical predictions and supporting our unified approach. 

   more » « less