Search

LLM-as-a-Judge offers a promising alternative to human judges across various tasks, yet inherent biases, particularly position bias - a systematic preference for answers based on their position in the prompt - compromise its effectiveness. Our study investigates this issue by developing a framework to systematically study and quantify position bias using metrics such as repetitional consistency, positional consistency, and positional fairness. We conduct experiments with 9 judge models across 22 tasks from the MTBench and DevBench benchmarks and nearly 40 answer-generating models, generating approximately 80,000 evaluation instances. This comprehensive assessment reveals significant variations in bias across judges and tasks. Although GPT-4 often excels in positional consistency and fairness, some more cost-effective models perform comparably or even better in specific tasks, highlighting essential trade-offs between consistency, fairness, and cost. Our results also demonstrate high consistency of judgment across repetitions, confirming that position bias is not due to random variations. This research significantly contributes to the field by introducing new concepts for understanding position bias and providing a multi-dimensional framework for evaluation. These insights guide the selection of optimal judge models, enhance benchmark design, and lay the foundation for future research into effective debiasing strategies, ultimately enhancing the reliability of LLM evaluators., Comment: 70 pages, around 200 figures and subfigures

In the pursuit of accurate and scalable quantitative methods for financial market analysis, the focus has shifted from individual stock models to those capturing interrelations between companies and their stocks. However, current relational stock methods are limited by their reliance on predefined stock relationships and the exclusive consideration of immediate effects. To address these limitations, we present a groundbreaking framework for financial market analysis. This approach, to our knowledge, is the first to jointly model investor expectations and automatically mine latent stock relationships. Comprehensive experiments conducted on China's CSI 300, one of the world's largest markets, demonstrate that our model consistently achieves an annual return exceeding 10%. This performance surpasses existing benchmarks, setting a new state-of-the-art standard in stock return prediction and multiyear trading simulations (i.e., backtesting)., Comment: In Advances in Knowledge Discovery and Data Mining 2023 (PAKDD 2023)

Graph representation learning (also known as network embedding) has been extensively researched with varying levels of granularity, ranging from nodes to graphs. While most prior work in this area focuses on node-level representation, limited research has been conducted on graph-level embedding, particularly for dynamic or temporal networks. However, learning low-dimensional graph-level representations for dynamic networks is critical for various downstream graph retrieval tasks such as temporal graph similarity ranking, temporal graph isomorphism, and anomaly detection. In this paper, we present a novel method for temporal graph-level embedding that addresses this gap. Our approach involves constructing a multilayer graph and using a modified random walk with temporal backtracking to generate temporal contexts for the graph's nodes. We then train a "document-level" language model on these contexts to generate graph-level embeddings. We evaluate our proposed model on five publicly available datasets for the task of temporal graph similarity ranking, and our model outperforms baseline methods. Our experimental results demonstrate the effectiveness of our method in generating graph-level embeddings for dynamic networks., Comment: In Companion Proceedings of the ACM Web Conference 2023

Differential framing of issues can lead to divergent world views on important issues. This is especially true in domains where the information presented can reach a large audience, such as traditional and social media. Scalable and reliable measurement of such differential framing is an important first step in addressing them. In this work, based on the intuition that framing affects the tone and word choices in written language, we propose a framework for modeling the differential framing of issues through masked token prediction via large-scale fine-tuned language models (LMs). Specifically, we explore three key factors for our framework: 1) prompt generation methods for the masked token prediction; 2) methods for normalizing the output of fine-tuned LMs; 3) robustness to the choice of pre-trained LMs used for fine-tuning. Through experiments on a dataset of articles from traditional media outlets covering five diverse and politically polarized topics, we show that our framework can capture differential framing of these topics with high reliability., Comment: In Findings of EMNLP 2022

While cultural backgrounds have been shown to affect linguistic expressions, existing natural language processing (NLP) research on culture modeling is overly coarse-grained and does not examine cultural differences among speakers of the same language. To address this problem and augment NLP models with cultural background features, we collect, annotate, manually validate, and benchmark EnCBP, a finer-grained news-based cultural background prediction dataset in English. Through language modeling (LM) evaluations and manual analyses, we confirm that there are noticeable differences in linguistic expressions among five English-speaking countries and across four states in the US. Additionally, our evaluations on nine syntactic (CoNLL-2003), semantic (PAWS-Wiki, QNLI, STS-B, and RTE), and psycholinguistic tasks (SST-5, SST-2, Emotion, and Go-Emotions) show that, while introducing cultural background information does not benefit the Go-Emotions task due to text domain conflicts, it noticeably improves deep learning (DL) model performance on other tasks. Our findings strongly support the importance of cultural background modeling to a wide variety of NLP tasks and demonstrate the applicability of EnCBP in culture-related research., Comment: In Findings of ACL 2022

Nodes in networks may have one or more functions that determine their role in the system. As opposed to local proximity, which captures the local context of nodes, the role identity captures the functional "role" that nodes play in a network, such as being the center of a group, or the bridge between two groups. This means that nodes far apart in a network can have similar structural role identities. Several recent works have explored methods for embedding the roles of nodes in networks. However, these methods all rely on either approximating or indirect modeling of structural equivalence. In this paper, we present a novel and flexible framework using stress majorization, to transform the high-dimensional role identities in networks directly (without approximation or indirect modeling) to a low-dimensional embedding space. Our method is also flexible, in that it does not rely on specific structural similarity definitions. We evaluated our method on the tasks of node classification, clustering, and visualization, using three real-world and five synthetic networks. Our experiments show that our framework achieves superior results than existing methods in learning node role representations., Comment: In CIKM 2021

Recent years have seen a rise in the development of representational learning methods for graph data. Most of these methods, however, focus on node-level representation learning at various scales (e.g., microscopic, mesoscopic, and macroscopic node embedding). In comparison, methods for representation learning on whole graphs are currently relatively sparse. In this paper, we propose a novel unsupervised whole graph embedding method. Our method uses spectral graph wavelets to capture topological similarities on each k-hop sub-graph between nodes and uses them to learn embeddings for the whole graph. We evaluate our method against 12 well-known baselines on 4 real-world datasets and show that our method achieves the best performance across all experiments, outperforming the current state-of-the-art by a considerable margin., Comment: In CIKM 2021

A key problem in multi-task learning (MTL) research is how to select high-quality auxiliary tasks automatically. This paper presents GradTS, an automatic auxiliary task selection method based on gradient calculation in Transformer-based models. Compared to AUTOSEM, a strong baseline method, GradTS improves the performance of MT-DNN with a bert-base-cased backend model, from 0.33% to 17.93% on 8 natural language understanding (NLU) tasks in the GLUE benchmarks. GradTS is also time-saving since (1) its gradient calculations are based on single-task experiments and (2) the gradients are re-used without additional experiments when the candidate task set changes. On the 8 GLUE classification tasks, for example, GradTS costs on average 21.32% less time than AUTOSEM with comparable GPU consumption. Further, we show the robustness of GradTS across various task settings and model selections, e.g. mixed objectives among candidate tasks. The efficiency and efficacy of GradTS in these case studies illustrate its general applicability in MTL research without requiring manual task filtering or costly parameter tuning., Comment: In EMNLP 2021

This paper studies the relative importance of attention heads in Transformer-based models to aid their interpretability in cross-lingual and multi-lingual tasks. Prior research has found that only a few attention heads are important in each mono-lingual Natural Language Processing (NLP) task and pruning the remaining heads leads to comparable or improved performance of the model. However, the impact of pruning attention heads is not yet clear in cross-lingual and multi-lingual tasks. Through extensive experiments, we show that (1) pruning a number of attention heads in a multi-lingual Transformer-based model has, in general, positive effects on its performance in cross-lingual and multi-lingual tasks and (2) the attention heads to be pruned can be ranked using gradients and identified with a few trial experiments. Our experiments focus on sequence labeling tasks, with potential applicability on other cross-lingual and multi-lingual tasks. For comprehensiveness, we examine two pre-trained multi-lingual models, namely multi-lingual BERT (mBERT) and XLM-R, on three tasks across 9 languages each. We also discuss the validity of our findings and their extensibility to truly resource-scarce languages and other task settings., Comment: In ACL 2021

Networks found in the real-world are numerous and varied. A common type of network is the heterogeneous network, where the nodes (and edges) can be of different types. Accordingly, there have been efforts at learning representations of these heterogeneous networks in low-dimensional space. However, most of the existing heterogeneous network embedding methods suffer from the following two drawbacks: (1) The target space is usually Euclidean. Conversely, many recent works have shown that complex networks may have hyperbolic latent anatomy, which is non-Euclidean. (2) These methods usually rely on meta-paths, which require domain-specific prior knowledge for meta-path selection. Additionally, different down-streaming tasks on the same network might require different meta-paths in order to generate task-specific embeddings. In this paper, we propose a novel self-guided random walk method that does not require meta-path for embedding heterogeneous networks into hyperbolic space. We conduct thorough experiments for the tasks of network reconstruction and link prediction on two public datasets, showing that our model outperforms a variety of well-known baselines across all tasks., Comment: In proceedings of the 35th AAAI Conference on Artificial Intelligence

This paper describes a system submitted by team BigGreen to LCP 2021 for predicting the lexical complexity of English words in a given context. We assemble a feature engineering-based model with a deep neural network model founded on BERT. While BERT itself performs competitively, our feature engineering-based model helps in extreme cases, eg. separating instances of easy and neutral difficulty. Our handcrafted features comprise a breadth of lexical, semantic, syntactic, and novel phonological measures. Visualizations of BERT attention maps offer insight into potential features that Transformers models may learn when fine-tuned for lexical complexity prediction. Our ensembled predictions score reasonably well for the single word subtask, and we demonstrate how they can be harnessed to perform well on the multi word expression subtask too., Comment: In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021). Colocated with ACL-IJCNLP 2021

This paper describes our approach to the Toxic Spans Detection problem (SemEval-2021 Task 5). We propose BERToxic, a system that fine-tunes a pre-trained BERT model to locate toxic text spans in a given text and utilizes additional post-processing steps to refine the boundaries. The post-processing steps involve (1) labeling character offsets between consecutive toxic tokens as toxic and (2) assigning a toxic label to words that have at least one token labeled as toxic. Through experiments, we show that these two post-processing steps improve the performance of our model by 4.16% on the test set. We also studied the effects of data augmentation and ensemble modeling strategies on our system. Our system significantly outperformed the provided baseline and achieved an F1-score of 0.683, placing Lone Pine in the 17th place out of 91 teams in the competition. Our code is made available at https://github.com/Yakoob-Khan/Toxic-Spans-Detection, Comment: 7 pages, 3 figures. Accepted at SemEval-2021 Workshop, ACL-IJCNLP 2021

Metaphors are ubiquitous in human language. The metaphor detection task (MD) aims at detecting and interpreting metaphors from written language, which is crucial in natural language understanding (NLU) research. In this paper, we introduce a pre-trained Transformer-based model into MD. Our model outperforms the previous state-of-the-art models by large margins in our evaluations, with relative improvements on the F-1 score from 5.33% to 28.39%. Second, we extend MD to a classification task about the metaphoricity of an entire piece of text to make MD applicable in more general NLU scenes. Finally, we clean up the improper or outdated annotations in one of the MD benchmark datasets and re-benchmark it with our Transformer-based model. This approach could be applied to other existing MD datasets as well, since the metaphoricity annotations in these benchmark datasets may be outdated. Future research efforts are also necessary to build an up-to-date and well-annotated dataset consisting of longer and more complex texts.

The field of NLP has seen unprecedented achievements in recent years. Most notably, with the advent of large-scale pre-trained Transformer-based language models, such as BERT, there has been a noticeable improvement in text representation. It is, however, unclear whether these improvements translate to noisy user-generated text, such as tweets. In this paper, we present an experimental survey of a wide range of well-known text representation techniques for the task of text clustering on noisy Twitter data. Our results indicate that the more advanced models do not necessarily work best on tweets and that more exploration in this area is needed., Comment: In proceedings of the 6th Workshop on Noisy User-generated Text (W-NUT) at EMNLP 2020

Data augmentation is proven to be effective in many NLU tasks, especially for those suffering from data scarcity. In this paper, we present a powerful and easy to deploy text augmentation framework, Data Boost, which augments data through reinforcement learning guided conditional generation. We evaluate Data Boost on three diverse text classification tasks under five different classifier architectures. The result shows that Data Boost can boost the performance of classifiers especially in low-resource data scenarios. For instance, Data Boost improves F1 for the three tasks by 8.7% on average when given only 10% of the whole data for training. We also compare Data Boost with six prior text augmentation methods. Through human evaluations (N=178), we confirm that Data Boost augmentation has comparable quality as the original data with respect to readability and class consistency., Comment: In proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP 2020). Online

Authorship identification tasks, which rely heavily on linguistic styles, have always been an important part of Natural Language Understanding (NLU) research. While other tasks based on linguistic style understanding benefit from deep learning methods, these methods have not behaved as well as traditional machine learning methods in many authorship-based tasks. With these tasks becoming more and more challenging, however, traditional machine learning methods based on handcrafted feature sets are already approaching their performance limits. Thus, in order to inspire future applications of deep learning methods in authorship-based tasks in ways that benefit the extraction of stylistic features, we survey authorship-based tasks and other tasks related to writing style understanding. We first describe our survey results on the current state of research in both sets of tasks and summarize existing achievements and problems in authorship-related tasks. We then describe outstanding methods in style-related tasks in general and analyze how they are used in combination in the top-performing models. We are optimistic about the applicability of these models to authorship-based tasks and hope our survey will help advance research in this field.

Emojis are a succinct form of language which can express concrete meanings, emotions, and intentions. Emojis also carry signals that can be used to better understand communicative intent. They have become a ubiquitous part of our daily lives, making them an important part of understanding user-generated content. The emoji prediction task aims at predicting the proper set of emojis associated with a piece of text. Through emoji prediction, models can learn rich representations of the communicative intent of the written text. While existing research on the emoji prediction task focus on a small subset of emoji types closely related to certain emotions, this setting oversimplifies the task and wastes the expressive power of emojis. In this paper, we extend the existing setting of the emoji prediction task to include a richer set of emojis and to allow multi-label classification on the task. We propose novel models for multi-class and multi-label emoji prediction based on Transformer networks. We also construct multiple emoji prediction datasets from Twitter using heuristics. The BERT models achieve state-of-the-art performances on all our datasets under all the settings, with relative improvements of 27.21% to 236.36% in accuracy, 2.01% to 88.28% in top-5 accuracy and 65.19% to 346.79% in F-1 score, compared to the prior state-of-the-art. Our results demonstrate the efficacy of deep Transformer-based models on the emoji prediction task. We also release our datasets at https://github.com/hikari-NYU/Emoji_Prediction_Datasets_MMS for future researchers., Comment: In Proceedings of the 9th KDD Workshop on Issues of Sentiment Discovery and Opinion Mining (WISDOM 20). San Diego, CA, USA, 7 pages