Peng Qian

Mycal Tucker, Tiwalayo Eisape, Peng Qian et al. · mit

Recent causal probing literature reveals when language models and syntactic probes use similar representations. Such techniques may yield "false negative" causality results: models may use representations of syntax, but probes may have learned to use redundant encodings of the same syntactic information. We demonstrate that models do encode syntactic information redundantly and introduce a new probe design that guides probes to consider all syntactic information present in embeddings. Using these probes, we find evidence for the use of syntax in models where prior methods did not, allowing us to boost model performance by injecting syntactic information into representations.

Zhengjie Huang, Yunyang Huang, Peng Qian et al.

Bitcoin is one of the decentralized cryptocurrencies powered by a peer-to-peer blockchain network. Parties who trade in the bitcoin network are not required to disclose any personal information. Such property of anonymity, however, precipitates potential malicious transactions to a certain extent. Indeed, various illegal activities such as money laundering, dark network trading, and gambling in the bitcoin network are nothing new now. While a proliferation of work has been developed to identify malicious bitcoin transactions, the behavior analysis and classification of bitcoin addresses are largely overlooked by existing tools. In this paper, we propose BAClassifier, a tool that can automatically classify bitcoin addresses based on their behaviors. Technically, we come up with the following three key designs. First, we consider casting the transactions of the bitcoin address into an address graph structure, of which we introduce a graph node compression technique and a graph structure augmentation method to characterize a unified graph representation. Furthermore, we leverage a graph feature network to learn the graph representations of each address and generate the graph embeddings. Finally, we aggregate all graph embeddings of an address into the address-level representation, and engage in a classification model to give the address behavior classification. As a side contribution, we construct and release a large-scale annotated dataset that consists of over 2 million real-world bitcoin addresses and concerns 4 types of address behaviors. Experimental results demonstrate that our proposed framework outperforms state-of-the-art bitcoin address classifiers and existing classification models, where the precision and F1-score are 96% and 95%, respectively. Our implementation and dataset are released, hoping to inspire others.

Sonia K. Murthy, Rosie Zhao, Jennifer Hu et al. · deepmind

Value trade-offs are an integral part of human decision-making and language use, however, current tools for interpreting such dynamic and multi-faceted notions of values in LLMs are limited. In cognitive science, so-called "cognitive models" provide formal accounts of such trade-offs in humans, by modeling the weighting of a speaker's competing utility functions in choosing an action or utterance. Here we use a leading cognitive model of polite speech to systematically evaluate value trade-offs in two encompassing model settings: degrees of reasoning "effort" in frontier black-box models, and RL post-training dynamics of open-source models. Our results highlight patterns of higher informational utility than social utility in reasoning models' default behavior, and demonstrate that these patterns shift in predictable ways when models are prompted to prioritize certain goals over others. Our findings from LLMs' training dynamics suggest large shifts in utility values early on in training with persistent effects of the choice of base model and pretraining data, compared to feedback dataset or alignment method. Our framework offers a flexible tool for probing value trade-offs across diverse model types, providing insights for generating hypotheses about other social behaviors such as sycophancy and for shaping training regimes that better control trade-offs between values during model development.

Ziji Chen, Steven W. D. Chien, Peng Qian et al.

Modern machine learning (ML) has grown into a tightly coupled, full-stack ecosystem that combines hardware, software, network, and applications. Many users rely on cloud providers for elastic, isolated, and cost-efficient resources. Unfortunately, these platforms as a service use virtualization, which means operators have little insight into the users' workloads. This hinders resource optimizations by the operator, which is essential to ensure cost efficiency and minimize execution time. In this paper, we argue that workload knowledge is unnecessary for system-level optimization. We propose Reveal, which takes a hardware-centric approach, relying only on hardware signals - fully accessible by operators. Using low-level signals collected from the system, Reveal detects anomalies through an unsupervised learning pipeline. The pipeline is developed by analyzing over 30 popular ML models on various hardware platforms, ensuring adaptability to emerging workloads and unknown deployment patterns. Using Reveal, we successfully identified both network and system configuration issues, accelerating the DeepSeek model by 5.97%.

Dazhong Rong, Guoyao Yu, Shuheng Shen et al.

To gather a significant quantity of annotated training data for high-performance image classification models, numerous companies opt to enlist third-party providers to label their unlabeled data. This practice is widely regarded as secure, even in cases where some annotated errors occur, as the impact of these minor inaccuracies on the final performance of the models is negligible and existing backdoor attacks require attacker's ability to poison the training images. Nevertheless, in this paper, we propose clean-image backdoor attacks which uncover that backdoors can still be injected via a fraction of incorrect labels without modifying the training images. Specifically, in our attacks, the attacker first seeks a trigger feature to divide the training images into two parts: those with the feature and those without it. Subsequently, the attacker falsifies the labels of the former part to a backdoor class. The backdoor will be finally implanted into the target model after it is trained on the poisoned data. During the inference phase, the attacker can activate the backdoor in two ways: slightly modifying the input image to obtain the trigger feature, or taking an image that naturally has the trigger feature as input. We conduct extensive experiments to demonstrate the effectiveness and practicality of our attacks. According to the experimental results, we conclude that our attacks seriously jeopardize the fairness and robustness of image classification models, and it is necessary to be vigilant about the incorrect labels in outsourced labeling.

Ningyu Xu, Qi Zhang, Menghan Zhang et al.

Probing and enhancing large language models' reasoning capacity remains a crucial open question. Here we re-purpose the reverse dictionary task as a case study to probe LLMs' capacity for conceptual inference. We use in-context learning to guide the models to generate the term for an object concept implied in a linguistic description. Models robustly achieve high accuracy in this task, and their representation space encodes information about object categories and fine-grained features. Further experiments suggest that the conceptual inference ability as probed by the reverse-dictionary task predicts model's general reasoning performance across multiple benchmarks, despite similar syntactic generalization behaviors across models. Explorative analyses suggest that prompting LLMs with description$\Rightarrow$word examples may induce generalization beyond surface-level differences in task construals and facilitate models on broader commonsense reasoning problems.

Yifan Jia, Yanbin Wang, Jianguo Sun et al.

Current Ethereum fraud detection methods rely on context-independent, numerical transaction sequences, failing to capture semantic of account transactions. Furthermore, the pervasive homogeneity in Ethereum transaction records renders it challenging to learn discriminative account embeddings. Moreover, current self-supervised graph learning methods primarily learn node representations through graph reconstruction, resulting in suboptimal performance for node-level tasks like fraud account detection, while these methods also encounter scalability challenges. To tackle these challenges, we propose LMAE4Eth, a multi-view learning framework that fuses transaction semantics, masked graph embedding, and expert knowledge. We first propose a transaction-token contrastive language model (TxCLM) that transforms context-independent numerical transaction records into logically cohesive linguistic representations. To clearly characterize the semantic differences between accounts, we also use a token-aware contrastive learning pre-training objective together with the masked transaction model pre-training objective, learns high-expressive account representations. We then propose a masked account graph autoencoder (MAGAE) using generative self-supervised learning, which achieves superior node-level account detection by focusing on reconstructing account node features. To enable MAGAE to scale for large-scale training, we propose to integrate layer-neighbor sampling into the graph, which reduces the number of sampled vertices by several times without compromising training quality. Finally, using a cross-attention fusion network, we unify the embeddings of TxCLM and MAGAE to leverage the benefits of both. We evaluate our method against 21 baseline approaches on three datasets. Experimental results show that our method outperforms the best baseline by over 10% in F1-score on two of the datasets.

Yiwen Wang, Jennifer Hu, Roger Levy et al.

Prior work has shown that structural supervision helps English language models learn generalizations about syntactic phenomena such as subject-verb agreement. However, it remains unclear if such an inductive bias would also improve language models' ability to learn grammatical dependencies in typologically different languages. Here we investigate this question in Mandarin Chinese, which has a logographic, largely syllable-based writing system; different word order; and sparser morphology than English. We train LSTMs, Recurrent Neural Network Grammars, Transformer language models, and Transformer-parameterized generative parsing models on two Mandarin Chinese datasets of different sizes. We evaluate the models' ability to learn different aspects of Mandarin grammar that assess syntactic and semantic relationships. We find suggestive evidence that structural supervision helps with representing syntactic state across intervening content and improves performance in low-data settings, suggesting that the benefits of hierarchical inductive biases in acquiring dependency relationships may extend beyond English.

Peng Qian, Tahira Naseem, Roger Levy et al.

Transformer-based language models pre-trained on large amounts of text data have proven remarkably successful in learning generic transferable linguistic representations. Here we study whether structural guidance leads to more human-like systematic linguistic generalization in Transformer language models without resorting to pre-training on very large amounts of data. We explore two general ideas. The "Generative Parsing" idea jointly models the incremental parse and word sequence as part of the same sequence modeling task. The "Structural Scaffold" idea guides the language model's representation via additional structure loss that separately predicts the incremental constituency parse. We train the proposed models along with a vanilla Transformer language model baseline on a 14 million-token and a 46 million-token subset of the BLLIP dataset, and evaluate models' syntactic generalization performances on SG Test Suites and sized BLiMP. Experiment results across two benchmarks suggest converging evidence that generative structural supervisions can induce more robust and humanlike linguistic generalization in Transformer language models without the need for data intensive pre-training.

Zhenguang Liu, Peng Qian, Xiaoyang Wang et al.

Smart contract vulnerability detection draws extensive attention in recent years due to the substantial losses caused by hacker attacks. Existing efforts for contract security analysis heavily rely on rigid rules defined by experts, which are labor-intensive and non-scalable. More importantly, expert-defined rules tend to be error-prone and suffer the inherent risk of being cheated by crafty attackers. Recent researches focus on the symbolic execution and formal analysis of smart contracts for vulnerability detection, yet to achieve a precise and scalable solution. Although several methods have been proposed to detect vulnerabilities in smart contracts, there is still a lack of effort that considers combining expert-defined security patterns with deep neural networks. In this paper, we explore using graph neural networks and expert knowledge for smart contract vulnerability detection. Specifically, we cast the rich control- and data- flow semantics of the source code into a contract graph. To highlight the critical nodes in the graph, we further design a node elimination phase to normalize the graph. Then, we propose a novel temporal message propagation network to extract the graph feature from the normalized graph, and combine the graph feature with designed expert patterns to yield a final detection system. Extensive experiments are conducted on all the smart contracts that have source code in Ethereum and VNT Chain platforms. Empirical results show significant accuracy improvements over the state-of-the-art methods on three types of vulnerabilities, where the detection accuracy of our method reaches 89.15%, 89.02%, and 83.21% for reentrancy, timestamp dependence, and infinite loop vulnerabilities, respectively.

Zhenguang Liu, Peng Qian, Xiang Wang et al.

Smart contracts hold digital coins worth billions of dollars, their security issues have drawn extensive attention in the past years. Towards smart contract vulnerability detection, conventional methods heavily rely on fixed expert rules, leading to low accuracy and poor scalability. Recent deep learning approaches alleviate this issue but fail to encode useful expert knowledge. In this paper, we explore combining deep learning with expert patterns in an explainable fashion. Specifically, we develop automatic tools to extract expert patterns from the source code. We then cast the code into a semantic graph to extract deep graph features. Thereafter, the global graph feature and local expert patterns are fused to cooperate and approach the final prediction, while yielding their interpretable weights. Experiments are conducted on all available smart contracts with source code in two platforms, Ethereum and VNT Chain. Empirically, our system significantly outperforms state-of-the-art methods. Our code is released.

Mycal Tucker, Peng Qian, Roger Levy

Neural language models exhibit impressive performance on a variety of tasks, but their internal reasoning may be difficult to understand. Prior art aims to uncover meaningful properties within model representations via probes, but it is unclear how faithfully such probes portray information that the models actually use. To overcome such limitations, we propose a technique, inspired by causal analysis, for generating counterfactual embeddings within models. In experiments testing our technique, we produce evidence that suggests some BERT-based models use a tree-distance-like representation of syntax in downstream prediction tasks.

Ethan Wilcox, Peng Qian, Richard Futrell et al.

Humans can learn structural properties about a word from minimal experience, and deploy their learned syntactic representations uniformly in different grammatical contexts. We assess the ability of modern neural language models to reproduce this behavior in English and evaluate the effect of structural supervision on learning outcomes. First, we assess few-shot learning capabilities by developing controlled experiments that probe models' syntactic nominal number and verbal argument structure generalizations for tokens seen as few as two times during training. Second, we assess invariance properties of learned representation: the ability of a model to transfer syntactic generalizations from a base context (e.g., a simple declarative active-voice sentence) to a transformed context (e.g., an interrogative sentence). We test four models trained on the same dataset: an n-gram baseline, an LSTM, and two LSTM-variants trained with explicit structural supervision (Dyer et al.,2016; Charniak et al., 2016). We find that in most cases, the neural models are able to induce the proper syntactic generalizations after minimal exposure, often from just two examples during training, and that the two structurally supervised models generalize more accurately than the LSTM model. All neural models are able to leverage information learned in base contexts to drive expectations in transformed contexts, indicating that they have learned some invariance properties of syntax.

Ethan Gotlieb Wilcox, Jon Gauthier, Jennifer Hu et al.

Human reading behavior is tuned to the statistics of natural language: the time it takes human subjects to read a word can be predicted from estimates of the word's probability in context. However, it remains an open question what computational architecture best characterizes the expectations deployed in real time by humans that determine the behavioral signatures of reading. Here we test over two dozen models, independently manipulating computational architecture and training dataset size, on how well their next-word expectations predict human reading time behavior on naturalistic text corpora. We find that across model architectures and training dataset sizes the relationship between word log-probability and reading time is (near-)linear. We next evaluate how features of these models determine their psychometric predictive power, or ability to predict human reading behavior. In general, the better a model's next-word expectations, the better its psychometric predictive power. However, we find nontrivial differences across model architectures. For any given perplexity, deep Transformer models and n-gram models generally show superior psychometric predictive power over LSTM or structurally supervised neural models, especially for eye movement data. Finally, we compare models' psychometric predictive power to the depth of their syntactic knowledge, as measured by a battery of syntactic generalization tests developed using methods from controlled psycholinguistic experiments. Once perplexity is controlled for, we find no significant relationship between syntactic knowledge and predictive power. These results suggest that different approaches may be required to best model human real-time language comprehension behavior in naturalistic reading versus behavior for controlled linguistic materials designed for targeted probing of syntactic knowledge.

Jennifer Hu, Jon Gauthier, Peng Qian et al.

While state-of-the-art neural network models continue to achieve lower perplexity scores on language modeling benchmarks, it remains unknown whether optimizing for broad-coverage predictive performance leads to human-like syntactic knowledge. Furthermore, existing work has not provided a clear picture about the model properties required to produce proper syntactic generalizations. We present a systematic evaluation of the syntactic knowledge of neural language models, testing 20 combinations of model types and data sizes on a set of 34 English-language syntactic test suites. We find substantial differences in syntactic generalization performance by model architecture, with sequential models underperforming other architectures. Factorially manipulating model architecture and training dataset size (1M--40M words), we find that variability in syntactic generalization performance is substantially greater by architecture than by dataset size for the corpora tested in our experiments. Our results also reveal a dissociation between perplexity and syntactic generalization performance.

Aixiu An, Peng Qian, Ethan Wilcox et al.

Neural language models have achieved state-of-the-art performances on many NLP tasks, and recently have been shown to learn a number of hierarchically-sensitive syntactic dependencies between individual words. However, equally important for language processing is the ability to combine words into phrasal constituents, and use constituent-level features to drive downstream expectations. Here we investigate neural models' ability to represent constituent-level features, using coordinated noun phrases as a case study. We assess whether different neural language models trained on English and French represent phrase-level number and gender features, and use those features to drive downstream expectations. Our results suggest that models use a linear combination of NP constituent number to drive CoordNP/verb number agreement. This behavior is highly regular and even sensitive to local syntactic context, however it differs crucially from observed human behavior. Models have less success with gender agreement. Models trained on large corpora perform best, and there is no obvious advantage for models trained using explicit syntactic supervision.

Richard Futrell, Ethan Wilcox, Takashi Morita et al.

We deploy the methods of controlled psycholinguistic experimentation to shed light on the extent to which the behavior of neural network language models reflects incremental representations of syntactic state. To do so, we examine model behavior on artificial sentences containing a variety of syntactically complex structures. We test four models: two publicly available LSTM sequence models of English (Jozefowicz et al., 2016; Gulordava et al., 2018) trained on large datasets; an RNNG (Dyer et al., 2016) trained on a small, parsed dataset; and an LSTM trained on the same small corpus as the RNNG. We find evidence that the LSTMs trained on large datasets represent syntactic state over large spans of text in a way that is comparable to the RNNG, while the LSTM trained on the small dataset does not or does so only weakly.

Ethan Wilcox, Peng Qian, Richard Futrell et al.

State-of-the-art LSTM language models trained on large corpora learn sequential contingencies in impressive detail and have been shown to acquire a number of non-local grammatical dependencies with some success. Here we investigate whether supervision with hierarchical structure enhances learning of a range of grammatical dependencies, a question that has previously been addressed only for subject-verb agreement. Using controlled experimental methods from psycholinguistics, we compare the performance of word-based LSTM models versus two models that represent hierarchical structure and deploy it in left-to-right processing: Recurrent Neural Network Grammars (RNNGs) (Dyer et al., 2016) and a incrementalized version of the Parsing-as-Language-Modeling configuration from Chariak et al., (2016). Models are tested on a diverse range of configurations for two classes of non-local grammatical dependencies in English---Negative Polarity licensing and Filler--Gap Dependencies. Using the same training data across models, we find that structurally-supervised models outperform the LSTM, with the RNNG demonstrating best results on both types of grammatical dependencies and even learning many of the Island Constraints on the filler--gap dependency. Structural supervision thus provides data efficiency advantages over purely string-based training of neural language models in acquiring human-like generalizations about non-local grammatical dependencies.

Peng Qian, Xipeng Qiu, Xuanjing Huang

Recently, the long short-term memory neural network (LSTM) has attracted wide interest due to its success in many tasks. LSTM architecture consists of a memory cell and three gates, which looks similar to the neuronal networks in the brain. However, there still lacks the evidence of the cognitive plausibility of LSTM architecture as well as its working mechanism. In this paper, we study the cognitive plausibility of LSTM by aligning its internal architecture with the brain activity observed via fMRI when the subjects read a story. Experiment results show that the artificial memory vector in LSTM can accurately predict the observed sequential brain activities, indicating the correlation between LSTM architecture and the cognitive process of story reading.

Xipeng Qiu, Peng Qian, Liusong Yin et al.

In this paper, we give an overview for the shared task at the 4th CCF Conference on Natural Language Processing \& Chinese Computing (NLPCC 2015): Chinese word segmentation and part-of-speech (POS) tagging for micro-blog texts. Different with the popular used newswire datasets, the dataset of this shared task consists of the relatively informal micro-texts. The shared task has two sub-tasks: (1) individual Chinese word segmentation and (2) joint Chinese word segmentation and POS Tagging. Each subtask has three tracks to distinguish the systems with different resources. We first introduce the dataset and task, then we characterize the different approaches of the participating systems, report the test results, and provide a overview analysis of these results. An online system is available for open registration and evaluation at http://nlp.fudan.edu.cn/nlpcc2015.