Pengxiang Cheng

Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro et al.

Fine-tuning pre-trained language models has become the prevalent paradigm for building downstream NLP models. Oftentimes fine-tuned models are readily available but their training data is not, due to data privacy or intellectual property concerns. This creates a barrier to fusing knowledge across individual models to yield a better single model. In this paper, we study the problem of merging individual models built on different training data sets to obtain a single model that performs well both across all data set domains and can generalize on out-of-domain data. We propose a dataless knowledge fusion method that merges models in their parameter space, guided by weights that minimize prediction differences between the merged model and the individual models. Over a battery of evaluation settings, we show that the proposed method significantly outperforms baselines such as Fisher-weighted averaging or model ensembling. Further, we find that our method is a promising alternative to multi-task learning that can preserve or sometimes improve over the individual models without access to the training data. Finally, model merging is more efficient than training a multi-task model, thus making it applicable to a wider set of scenarios.

Niklas Stoehr, Pengxiang Cheng, Jing Wang et al. · eth-zurich

Language models contain ranking-based knowledge and are powerful solvers of in-context ranking tasks. For instance, they may have parametric knowledge about the ordering of countries by size or may be able to rank product reviews by sentiment. We compare pairwise, pointwise and listwise prompting techniques to elicit a language model's ranking knowledge. However, we find that even with careful calibration and constrained decoding, prompting-based techniques may not always be self-consistent in the rankings they produce. This motivates us to explore an alternative approach that is inspired by an unsupervised probing method called Contrast-Consistent Search (CCS). The idea is to train a probe guided by a logical constraint: a language model's representation of a statement and its negation must be mapped to contrastive true-false poles consistently across multiple statements. We hypothesize that similar constraints apply to ranking tasks where all items are related via consistent, pairwise or listwise comparisons. To this end, we extend the binary CCS method to Contrast-Consistent Ranking (CCR) by adapting existing ranking methods such as the Max-Margin Loss, Triplet Loss and an Ordinal Regression objective. Across different models and datasets, our results confirm that CCR probing performs better or, at least, on a par with prompting.

Dugang Liu, Pengxiang Cheng, Hong Zhu et al.

Tabular data is one of the most common data storage formats behind many real-world web applications such as retail, banking, and e-commerce. The success of these web applications largely depends on the ability of the employed machine learning model to accurately distinguish influential features from all the predetermined features in tabular data. Intuitively, in practical business scenarios, different instances should correspond to different sets of influential features, and the set of influential features of the same instance may vary in different scenarios. However, most existing methods focus on global feature selection assuming that all instances have the same set of influential features, and few methods considering instance-wise feature selection ignore the variability of influential features in different scenarios. In this paper, we first introduce a new perspective based on the influence function for instance-wise feature selection, and give some corresponding theoretical insights, the core of which is to use the influence function as an indicator to measure the importance of an instance-wise feature. We then propose a new solution for discovering instance-wise influential features in tabular data (DIWIFT), where a self-attention network is used as a feature selection model and the value of the corresponding influence function is used as an optimization objective to guide the model. Benefiting from the advantage of the influence function, i.e., its computation does not depend on a specific architecture and can also take into account the data distribution in different scenarios, our DIWIFT has better flexibility and robustness. Finally, we conduct extensive experiments on both synthetic and real-world datasets to validate the effectiveness of our DIWIFT.

Heyuan Huang, Xingyu Lou, Chaochao Chen et al.

Cross-Domain Recommendation (CDR) have received widespread attention due to their ability to utilize rich information across domains. However, most existing CDR methods assume an ideal static condition that is not practical in industrial recommendation systems (RS). Therefore, simply applying existing CDR methods in the industrial RS environment may lead to low effectiveness and efficiency. To fill this gap, we propose DIIT, an end-to-end Domain-Invariant Information Transfer method for industrial cross-domain recommendation. Specifically, We first simulate the industrial RS environment that maintains respective models in multiple domains, each of them is trained in the incremental mode. Then, for improving the effectiveness, we design two extractors to fully extract domain-invariant information from the latest source domain models at the domain level and the representation level respectively. Finally, for improving the efficiency, we design a migrator to transfer the extracted information to the latest target domain model, which only need the target domain model for inference. Experiments conducted on one production dataset and two public datasets verify the effectiveness and efficiency of DIIT.

Shuyang Cao, Karthik Radhakrishnan, David Rosenberg et al.

Retrieval-augmented generation (RAG) generally enhances large language models' (LLMs) ability to solve knowledge-intensive tasks. But RAG may also lead to performance degradation due to imperfect retrieval and the model's limited ability to leverage retrieved content. In this work, we evaluate the robustness of LLMs in practical RAG setups (henceforth retrieval robustness). We focus on three research questions: (1) whether RAG is always better than non-RAG; (2) whether more retrieved documents always lead to better performance; (3) and whether document orders impact results. To facilitate this study, we establish a benchmark of 1500 open-domain questions, each with retrieved documents from Wikipedia. We introduce three robustness metrics, each corresponds to one research question. Our comprehensive experiments, involving 11 LLMs and 3 prompting strategies, reveal that all of these LLMs exhibit surprisingly high retrieval robustness; nonetheless, different degrees of imperfect robustness hinders them from fully utilizing the benefits of RAG.

Aldo Porco, Dhruv Mehra, Igor Malioutov et al.

Learned Sparse Retrieval (LSR) models encode text as weighted term vectors, which need to be sparse to leverage inverted index structures during retrieval. SPLADE, the most popular LSR model, uses FLOPS regularization to encourage vector sparsity during training. However, FLOPS regularization does not ensure sparsity among terms - only within a given query or document. Terms with very high Document Frequencies (DFs) substantially increase latency in production retrieval engines, such as Apache Solr, due to their lengthy posting lists. To address the issue of high DFs, we present a new variant of FLOPS regularization: DF-FLOPS. This new regularization technique penalizes the usage of high-DF terms, thereby shortening posting lists and reducing retrieval latency. Unlike other inference-time sparsification methods, such as stopword removal, DF-FLOPS regularization allows for the selective inclusion of high-frequency terms in cases where the terms are truly salient. We find that DF-FLOPS successfully reduces the prevalence of high-DF terms and lowers retrieval latency (around 10x faster) in a production-grade engine while maintaining effectiveness both in-domain (only a 2.2-point drop in MRR@10) and cross-domain (improved performance in 12 out of 13 tasks on which we tested). With retrieval latencies on par with BM25, this work provides an important step towards making LSR practical for deployment in production-grade search engines.

Ozan İrsoy, Pengxiang Cheng, Jennifer L. Chen et al.

Instruct models, obtained from various instruction tuning or post-training steps, are commonly deemed superior and more usable than their base counterpart. While the model gains instruction following ability, instruction tuning may lead to forgetting the knowledge from pre-training or it may encourage the model to become overly conversational or verbose. This, in turn, can lead to degradation of in-context few-shot learning performance. In this work, we study the performance trajectory between base and instruct models by scaling down the strength of instruction-tuning via the partial adaption method. We show that, across several model families and model sizes, reducing the strength of instruction-tuning results in material improvement on a few-shot in-context learning benchmark covering a variety of classic natural language tasks. This comes at the cost of losing some degree of instruction following ability as measured by AlpacaEval. Our study shines light on the potential trade-off between in-context learning and instruction following abilities that is worth considering in practice.

Genta Indra Winata, Lingjue Xie, Karthik Radhakrishnan et al.

Real-life multilingual systems should be able to efficiently incorporate new languages as data distributions fed to the system evolve and shift over time. To do this, systems need to handle the issue of catastrophic forgetting, where the model performance drops for languages or tasks seen further in its past. In this paper, we study catastrophic forgetting, as well as methods to minimize this, in a massively multilingual continual learning framework involving up to 51 languages and covering both classification and sequence labeling tasks. We present LR ADJUST, a learning rate scheduling method that is simple, yet effective in preserving new information without strongly overwriting past knowledge. Furthermore, we show that this method is effective across multiple continual learning approaches. Finally, we provide further insights into the dynamics of catastrophic forgetting in this massively multilingual setup.

Mi Luo, Fei Chen, Pengxiang Cheng et al.

Recommender systems often face heterogeneous datasets containing highly personalized historical data of users, where no single model could give the best recommendation for every user. We observe this ubiquitous phenomenon on both public and private datasets and address the model selection problem in pursuit of optimizing the quality of recommendation for each user. We propose a meta-learning framework to facilitate user-level adaptive model selection in recommender systems. In this framework, a collection of recommenders is trained with data from all users, on top of which a model selector is trained via meta-learning to select the best single model for each user with the user-specific historical data. We conduct extensive experiments on two public datasets and a real-world production dataset, demonstrating that our proposed framework achieves improvements over single model baselines and sample-level model selector in terms of AUC and LogLoss. In particular, the improvements may lead to huge profit gain when deployed in online recommender systems.

Pengxiang Cheng, Katrin Erk

Recent progress in NLP witnessed the development of large-scale pre-trained language models (GPT, BERT, XLNet, etc.) based on Transformer (Vaswani et al. 2017), and in a range of end tasks, such models have achieved state-of-the-art results, approaching human performance. This demonstrates the power of the stacked self-attention architecture when paired with a sufficient number of layers and a large amount of pre-training data. However, on tasks that require complex and long-distance reasoning where surface-level cues are not enough, there is still a large gap between the pre-trained models and human performance. Strubell et al. (2018) recently showed that it is possible to inject knowledge of syntactic structure into a model through supervised self-attention. We conjecture that a similar injection of semantic knowledge, in particular, coreference information, into an existing model would improve performance on such complex problems. On the LAMBADA (Paperno et al. 2016) task, we show that a model trained from scratch with coreference as auxiliary supervision for self-attention outperforms the largest GPT-2 model, setting the new state-of-the-art, while only containing a tiny fraction of parameters compared to GPT-2. We also conduct a thorough analysis of different variants of model architectures and supervision configurations, suggesting future directions on applying similar techniques to other problems.

Pengxiang Cheng, Katrin Erk

Implicit arguments, which cannot be detected solely through syntactic cues, make it harder to extract predicate-argument tuples. We present a new model for implicit argument prediction that draws on reading comprehension, casting the predicate-argument tuple with the missing argument as a query. We also draw on pointer networks and multi-hop computation. Our model shows good performance on an argument cloze task as well as on a nominal implicit argument prediction task.

Pengxiang Cheng, Katrin Erk

Implicit arguments are not syntactically connected to their predicates, and are therefore hard to extract. Previous work has used models with large numbers of features, evaluated on very small datasets. We propose to train models for implicit argument prediction on a simple cloze task, for which data can be generated automatically at scale. This allows us to use a neural model, which draws on narrative coherence and entity salience for predictions. We show that our model has superior performance on both synthetic and natural data.

Yalin Sun, Pengxiang Cheng, Shengwei Wang et al.

Information extraction is a critical step in the practice of conducting biomedical systematic literature reviews. Extracted structured data can be aggregated via methods such as statistical meta-analysis. Typically highly trained domain experts extract data for systematic reviews. The high expense of conducting biomedical systematic reviews has motivated researchers to explore lower cost methods that achieve similar rigor without compromising quality. Crowdsourcing represents one such promising approach. In this work-in-progress study, we designed a crowdsourcing task for biomedical information extraction. We briefly report the iterative design process and the results of two pilot testings. We found that giving more concrete examples in the task instruction can help workers better understand the task, especially for concepts that are abstract and confusing. We found a few workers completed most of the work, and our payment level appeared more attractive to workers from low-income countries. In the future, we will further evaluate our results with reference to gold standard extractions, thus assessing the feasibility of tasking crowd workers with extracting biomedical intervention information for systematic reviews.

I. Beltagy, Stephen Roller, Pengxiang Cheng et al.

NLP tasks differ in the semantic information they require, and at this time no single se- mantic representation fulfills all requirements. Logic-based representations characterize sentence structure, but do not capture the graded aspect of meaning. Distributional models give graded similarity ratings for words and phrases, but do not capture sentence structure in the same detail as logic-based approaches. So it has been argued that the two are complementary. We adopt a hybrid approach that combines logic-based and distributional semantics through probabilistic logic inference in Markov Logic Networks (MLNs). In this paper, we focus on the three components of a practical system integrating logical and distributional models: 1) Parsing and task representation is the logic-based part where input problems are represented in probabilistic logic. This is quite different from representing them in standard first-order logic. 2) For knowledge base construction we form weighted inference rules. We integrate and compare distributional information with other sources, notably WordNet and an existing paraphrase collection. In particular, we use our system to evaluate distributional lexical entailment approaches. We use a variant of Robinson resolution to determine the necessary inference rules. More sources can easily be added by mapping them to logical rules; our system learns a resource-specific weight that corrects for scaling differences between resources. 3) In discussing probabilistic inference, we show how to solve the inference problems efficiently. To evaluate our approach, we use the task of textual entailment (RTE), which can utilize the strengths of both logic-based and distributional representations. In particular we focus on the SICK dataset, where we achieve state-of-the-art results.