While pre-trained language models (LMs) have brought great improvements in many NLP tasks, there is increasing attention to explore capabilities of LMs and interpret their predictions. However, existing works usually focus only on a certain capability with some downstream tasks. There is a lack of datasets for directly evaluating the masked word prediction performance and the interpretability of pre-trained LMs. To fill in the gap, we propose a novel evaluation benchmark providing with both English and Chinese annotated data. It tests LMs abilities in multiple dimensions, i.e., grammar, semantics, knowledge, reasoning and computation. In addition, it provides carefully annotated token-level rationales that satisfy sufficiency and compactness. It contains perturbed instances for each original instance, so as to use the rationale consistency under perturbations as the metric for faithfulness, a perspective of interpretability. We conduct experiments on several widely-used pre-trained LMs. The results show that they perform very poorly on the dimensions of knowledge and computation. And their plausibility in all dimensions is far from satisfactory, especially when the rationale is short. In addition, the pre-trained LMs we evaluated are not robust on syntax-aware data. We will release this evaluation benchmark at \url{http://xyz}, and hope it can facilitate the research progress of pre-trained LMs.
AI-based protein structure prediction pipelines, such as AlphaFold2, have achieved near-experimental accuracy. These advanced pipelines mainly rely on Multiple Sequence Alignments (MSAs) and templates as inputs to learn the co-evolution information from the homologous sequences. Nonetheless, searching MSAs and templates from protein databases is time-consuming, usually taking dozens of minutes. Consequently, we attempt to explore the limits of fast protein structure prediction by using only primary sequences of proteins. HelixFold-Single is proposed to combine a large-scale protein language model with the superior geometric learning capability of AlphaFold2. Our proposed method, HelixFold-Single, first pre-trains a large-scale protein language model (PLM) with thousands of millions of primary sequences utilizing the self-supervised learning paradigm, which will be used as an alternative to MSAs and templates for learning the co-evolution information. Then, by combining the pre-trained PLM and the essential components of AlphaFold2, we obtain an end-to-end differentiable model to predict the 3D coordinates of atoms from only the primary sequence. HelixFold-Single is validated in datasets CASP14 and CAMEO, achieving competitive accuracy with the MSA-based methods on the targets with large homologous families. Furthermore, HelixFold-Single consumes much less time than the mainstream pipelines for protein structure prediction, demonstrating its potential in tasks requiring many predictions. The code of HelixFold-Single is available at https://github.com/PaddlePaddle/PaddleHelix/tree/dev/apps/protein_folding/helixfold-single, and we also provide stable web services on https://paddlehelix.baidu.com/app/drug/protein-single/forecast.
Generative open-domain dialogue systems can benefit from external knowledge, but the lack of external knowledge resources and the difficulty in finding relevant knowledge limit the development of this technology. To this end, we propose a knowledge-driven dialogue task using dynamic service information. Specifically, we use a large number of service APIs that can provide high coverage and spatiotemporal sensitivity as external knowledge sources. The dialogue system generates queries to request external services along with user information, get the relevant knowledge, and generate responses based on this knowledge. To implement this method, we collect and release the first open domain Chinese service knowledge dialogue dataset DuSinc. At the same time, we construct a baseline model PLATO-SINC, which realizes the automatic utilization of service information for dialogue. Both automatic evaluation and human evaluation show that our proposed new method can significantly improve the effect of open-domain conversation, and the session-level overall score in human evaluation is improved by 59.29% compared with the dialogue pre-training model PLATO-2. The dataset and benchmark model will be open sourced.
We introduce Bi-SimCut: a simple but effective training strategy to boost neural machine translation (NMT) performance. It consists of two procedures: bidirectional pretraining and unidirectional finetuning. Both procedures utilize SimCut, a simple regularization method that forces the consistency between the output distributions of the original and the cutoff sentence pairs. Without leveraging extra dataset via back-translation or integrating large-scale pretrained model, Bi-SimCut achieves strong translation performance across five translation benchmarks (data sizes range from 160K to 20.2M): BLEU scores of 31.16 for en -> de and 38.37 for de -> en on the IWSLT14 dataset, 30.78 for en -> de and 35.15 for de -> en on the WMT14 dataset, and 27.17 for zh -> en on the WMT17 dataset. SimCut is not a new method, but a version of Cutoff (Shen et al., 2020) simplified and adapted for NMT, and it could be considered as a perturbation-based method. Given the universality and simplicity of SimCut and Bi-SimCut, we believe they can serve as strong baselines for future NMT research.
Many recent works indicate that the deep neural networks tend to take dataset biases as shortcuts to make decision, rather than understand the tasks, which results in failures on the real-world applications. In this work, we focus on the spurious correlation between feature and label, which derive from the biased data distribution in the training data, and analyze it concretely. In particular, we define the word highly co-occurring with a specific label as biased word, and the example containing biased word as biased example. Our analysis reveals that the biased examples with spurious correlations are easier for models to learn, and when predicting, the biased words make significantly higher contributions to models' predictions than other words, and the models tend to assign the labels over-relying on the spurious correlation between words and labels. To mitigate the model's over-reliance on the shortcut, we propose a training strategy Less-Learn-Shortcut (LLS): we quantify the biased degree of the biased examples, and down-weight them with the biased degree. Experimental results on QM and NLI tasks show that the models improve the performances both on in-domain and adversarial data (1.57% on DuQM and 2.12% on HANS) with our LLS.
While there is increasing concern about the interpretability of neural models, the evaluation of interpretability remains an open problem, due to the lack of proper evaluation datasets and metrics. In this paper, we present a novel benchmark to evaluate the interpretability of both neural models and saliency methods. This benchmark covers three representative NLP tasks: sentiment analysis, textual similarity and reading comprehension, each provided with both English and Chinese annotated data. In order to precisely evaluate the interpretability, we provide token-level rationales that are carefully annotated to be sufficient, compact and comprehensive. We also design a new metric, i.e., the consistency between the rationales before and after perturbations, to uniformly evaluate the interpretability of models and saliency methods on different tasks. Based on this benchmark, we conduct experiments on three typical models with three saliency methods, and unveil their strengths and weakness in terms of interpretability. We will release this benchmark at \url{https://xyz} and hope it can facilitate the research in building trustworthy systems.
Neural retrievers based on pre-trained language models (PLMs), such as dual-encoders, have achieved promising performance on the task of open-domain question answering (QA). Their effectiveness can further reach new state-of-the-arts by incorporating cross-architecture knowledge distillation. However, most of the existing studies just directly apply conventional distillation methods. They fail to consider the particular situation where the teacher and student have different structures. In this paper, we propose a novel distillation method that significantly advances cross-architecture distillation for dual-encoders. Our method 1) introduces a self on-the-fly distillation method that can effectively distill late interaction (i.e., ColBERT) to vanilla dual-encoder, and 2) incorporates a cascade distillation process to further improve the performance with a cross-encoder teacher. Extensive experiments are conducted to validate that our proposed solution outperforms strong baselines and establish a new state-of-the-art on open-domain QA benchmarks.
Accurate ADMET (an abbreviation for "absorption, distribution, metabolism, excretion, and toxicity") predictions can efficiently screen out undesirable drug candidates in the early stage of drug discovery. In recent years, multiple comprehensive ADMET systems that adopt advanced machine learning models have been developed, providing services to estimate multiple endpoints. However, those ADMET systems usually suffer from weak extrapolation ability. First, due to the lack of labelled data for each endpoint, typical machine learning models perform frail for the molecules with unobserved scaffolds. Second, most systems only provide fixed built-in endpoints and cannot be customised to satisfy various research requirements. To this end, we develop a robust and endpoint extensible ADMET system, HelixADMET (H-ADMET). H-ADMET incorporates the concept of self-supervised learning to produce a robust pre-trained model. The model is then fine-tuned with a multi-task and multi-stage framework to transfer knowledge between ADMET endpoints, auxiliary tasks, and self-supervised tasks. Our results demonstrate that H-ADMET achieves an overall improvement of 4%, compared with existing ADMET systems on comparable endpoints. Additionally, the pre-trained model provided by H-ADMET can be fine-tuned to generate new and customised ADMET endpoints, meeting various demands of drug research and development requirements.
Recent years have witnessed the significant advance in dense retrieval (DR) based on powerful pre-trained language models (PLM). DR models have achieved excellent performance in several benchmark datasets, while they are shown to be not as competitive as traditional sparse retrieval models (e.g., BM25) in a zero-shot retrieval setting. However, in the related literature, there still lacks a detailed and comprehensive study on zero-shot retrieval. In this paper, we present the first thorough examination of the zero-shot capability of DR models. We aim to identify the key factors and analyze how they affect zero-shot retrieval performance. In particular, we discuss the effect of several key factors related to source training set, analyze the potential bias from the target dataset, and review and compare existing zero-shot DR models. Our findings provide important evidence to better understand and develop zero-shot DR models.
Emotional support is a crucial skill for many real-world scenarios, including caring for the elderly, mental health support, and customer service chats. This paper presents a novel task of empathetic dialog generation with positive emotion elicitation to promote users' positive emotions, similar to that of emotional support between humans. In this task, the agent conducts empathetic responses along with the target of eliciting the user's positive emotions in the multi-turn dialog. To facilitate the study of this task, we collect a large-scale emotional dialog dataset with positive emotion elicitation, called PosEmoDial (about 820k dialogs, 3M utterances). In these dialogs, the agent tries to guide the user from any possible initial emotional state, e.g., sadness, to a positive emotional state. Then we present a positive-emotion-guided dialog generation model with a novel loss function design. This loss function encourages the dialog model to not only elicit positive emotions from users but also ensure smooth emotional transitions along with the whole dialog. Finally, we establish benchmark results on PosEmoDial, and we will release this dataset and related source code to facilitate future studies.