AAAI 2022 论文列表

Cross-lingual Machine Reading Comprehension (xMRC) is a challenging task due to the lack of training data in low-resource languages. Recent approaches use training data only in a resource-rich language (such as English) to fine-tune large-scale cross-lingual pre-trained language models, which transfer knowledge from resource-rich languages (source) to low-resource languages (target). Due to the big difference between languages, the model fine-tuned only by the source language may not perform well for target languages. In our study, we make an interesting observation that while the top 1 result predicted by the previous approaches may often fail to hit the ground-truth answer, there are still good chances for the correct answer to be contained in the set of top k predicted results. Intuitively, the previous approaches have empowered the model certain level of capability to roughly distinguish good answers from bad ones. However, without sufficient training data, it is not powerful enough to capture the nuances between the accurate answer and those approximate ones. Based on this observation, we develop a two-stage approach to enhance the model performance. The first stage targets at recall; we design a hard-learning (HL) algorithm to maximize the likelihood that the top k predictions contain the accurate answer. The second stage focuses on precision, where an answer-aware contrastive learning (AA-CL) mechanism is developed to learn the minute difference between the accurate answer and other candidates. Extensive experiments show that our model significantly outperforms strong baselines on two cross-lingual MRC benchmark datasets.

Embedding matrices are key components in neural natural language processing (NLP) models that are responsible to provide numerical representations of input tokens (i.e. words or subwords). In this paper, we analyze the impact and utility of such matrices in the context of neural machine translation (NMT). We show that detracting syntactic and semantic information from word embeddings and running NMT systems with random embeddings is not as damaging as it initially sounds. We also show how incorporating only a limited amount of task-specific knowledge from fully-trained embeddings can boost the performance NMT systems. Our findings demonstrate that in exchange for negligible deterioration in performance, any NMT model can be run with partially random embeddings. Working with such structures means a minimal memory requirement as there is no longer need to store large embedding tables, which is a significant gain in industrial and on-device settings. We evaluated our embeddings in translating English into German and French and achieved a 5.3x compression rate. Despite having a considerably smaller architecture, our models in some cases are even able to outperform state-of-the-art baselines.

Pre-trained Language Models (PLMs) have achieved great success in various Natural Language Processing (NLP) tasks under the pre-training and fine-tuning paradigm. With large quantities of parameters, PLMs are computation-intensive and resource-hungry. Hence, model pruning has been introduced to compress large-scale PLMs. However, most prior approaches only consider task-specific knowledge towards downstream tasks, but ignore the essential task-agnostic knowledge during pruning, which may cause catastrophic forgetting problem and lead to poor generalization ability. To maintain both task-agnostic and task-specific knowledge in our pruned model, we propose ContrAstive Pruning (CAP) under the paradigm of pre-training and fine-tuning. It is designed as a general framework, compatible with both structured and unstructured pruning. Unified in contrastive learn- ing, CAP enables the pruned model to learn from the pre-trained model for task-agnostic knowledge, and fine-tuned model for task-specific knowledge. Besides, to better retain the performance of the pruned model, the snapshots (i.e., the intermediate models at each pruning iteration) also serve as effective supervisions for pruning. Our extensive experiments show that adopting CAP consistently yields significant improvements, especially in extremely high sparsity scenarios. With only 3% model parameters reserved (i.e., 97% sparsity), CAP successfully achieves 99.2% and 96.3% of the original BERT performance in QQP and MNLI tasks. In addition, our probing experiments demonstrate that the model pruned by CAP tends to achieve better generalization ability.

Causality plays a central role in reasoning about observations. In many cases, it might be useful to define the conditions under which a non-deterministic program can be called an actual cause of an effect in a setting where a sequence of programs are executed one after another. There can be two perspectives, one where at least one execution of the program leads to the effect, and another where all executions do so. The former captures a ''weak'' notion of causation and is more general than the latter stronger notion. In this paper, we give a definition of weak potential causes. Our analysis is performed within the situation calculus basic action theories and we consider programs formulated in the logic programming language ConGolog. Within this setting, we show how one can utilize a recently developed abstraction framework to relate causes at various levels of abstraction, which facilitates reasoning about programs as causes.

Low-frequency word prediction remains a challenge in modern neural machine translation (NMT) systems. Recent adaptive training methods promote the output of infrequent words by emphasizing their weights in the overall training objectives. Despite the improved recall of low-frequency words, their prediction precision is unexpectedly hindered by the adaptive objectives. Inspired by the observation that low-frequency words form a more compact embedding space, we tackle this challenge from a representation learning perspective. Specifically, we propose a frequency-aware token-level contrastive learning method, in which the hidden state of each decoding step is pushed away from the counterparts of other target words, in a soft contrastive way based on the corresponding word frequencies. We conduct experiments on widely used NIST Chinese-English and WMT14 English-German translation tasks. Empirical results show that our proposed methods can not only significantly improve the translation quality but also enhance lexical diversity and optimize word representation space. Further investigation reveals that, comparing with related adaptive training strategies, the superiority of our method on low-frequency word prediction lies in the robustness of token-level recall across different frequencies without sacrificing precision.

Various deepfake detectors have been proposed, but challenges still exist to detect images of unknown categories or GAN models outside of the training settings. Such issues arise from the overfitting issue, which we discover from our own analysis and the previous studies to originate from the frequency-level artifacts in generated images. We find that ignoring the frequency-level artifacts can improve the detector's generalization across various GAN models, but it can reduce the model's performance for the trained GAN models. Thus, we design a framework to generalize the deepfake detector for both the known and unseen GAN models. Our framework generates the frequency-level perturbation maps to make the generated images indistinguishable from the real images. By updating the deepfake detector along with the training of the perturbation generator, our model is trained to detect the frequency-level artifacts at the initial iterations and consider the image-level irregularities at the last iterations. For experiments, we design new test scenarios varying from the training settings in GAN models, color manipulations, and object categories. Numerous experiments validate the state-of-the-art performance of our deepfake detector.

This work introduces Fractional Adaptive Linear Units (FALUs), a flexible generalization of adaptive activation functions. Leveraging principles from fractional calculus, FALUs define a diverse family of activation functions (AFs) that encompass many traditional and state-of-the-art activation functions. This family includes the Sigmoid, Gaussian, ReLU, GELU, and Swish functions, as well as a large variety of smooth interpolations between these functions. Our technique requires only a small number of additional trainable parameters, and needs no further specialized optimization or initialization procedures. For this reason, FALUs present a seamless and rich automated solution to the problem of activation function optimization. Through experiments on a variety of conventional tasks and network architectures, we demonstrate the effectiveness of FALUs when compared to traditional and state-of-the-art AFs. To facilitate practical use of this work, we plan to make our code publicly available

Optimization of real-world black-box functions defined over purely categorical variables is an active area of research. In particular, optimization and design of biological sequences with specific functional or structural properties have a profound impact in medicine, materials science, and biotechnology. Standalone search algorithms, such as simulated annealing (SA) and Monte Carlo tree search (MCTS), are typically used for such optimization problems. In order to improve the performance and sample efficiency of such algorithms, we propose to use existing methods in conjunction with a surrogate model for the black-box evaluations over purely categorical variables. To this end, we present two different representations, a group-theoretic Fourier expansion and an abridged one-hot encoded Boolean Fourier expansion. To learn such representations, we consider two different settings to update our surrogate model. First, we utilize an adversarial online regression setting where Fourier characters of each representation are considered as experts and their respective coefficients are updated via an exponential weight update rule each time the black box is evaluated. Second, we consider a Bayesian setting where queries are selected via Thompson sampling and the posterior is updated via a sparse Bayesian regression model (over our proposed representation) with a regularized horseshoe prior. Numerical experiments over synthetic benchmarks as well as real-world RNA sequence optimization and design problems demonstrate the representational power of the proposed methods, which achieve competitive or superior performance compared to state-of-the-art counterparts, while improving the computation cost and/or sample efficiency, substantially.

In recent years, pretrained language models have revolutionized the NLP world, while achieving state of the art performance in various downstream tasks. However, in many cases, these models do not perform well when labeled data is scarce and the model is expected to perform in the zero or few shot setting. Recently, several works have shown that continual pretraining or performing a second phase of pretraining (inter-training) which is better aligned with the downstream task, can lead to improved results, especially in the scarce data setting. Here, we propose to leverage sentiment-carrying discourse markers to generate large-scale weakly-labeled data, which in turn can be used to adapt language models for sentiment analysis. Extensive experimental results show the value of our approach on various benchmark datasets, including the finance domain. Code, models and data are available at https://github.com/ibm/tslm-discourse-markers.

We introduce the formula-synthesis problem for Propositional Dynamic Logic with Shuffle (PDL || ). This problem, which generalises the model-checking problem againsts PDL || is the following: given a finite transition system and a regular term-grammar that generates (possibly infinitely many) PDL || formulas, find a formula generated by the grammar that is true in the structure (or return that there is none). We prove that the problem is undecidable in general, but add certain restrictions on the input structure or on the input grammar to yield decidability. In particular, we prove that (1) if the grammar only generates formulas in PDL (without shuffle), then the problem is EXPTIME-complete, and a further restriction to linear grammars is PSPACE-complete, and a further restriction to non-recursive grammars is NP-complete, and (2) if one restricts the input structure to have only simple paths then the problem is in 2-EXPTIME. This work is motivated by and opens up connections to other forms of synthesis from hierarchical descriptions, including HTN problems in Planning and Attack-tree Synthesis problems in Security.

We present a simple and concise semantics for temporal planning. Our semantics are developed and formalised in the logic of the interactive theorem prover Isabelle/HOL. We derive from those semantics a validation algorithm for temporal planning and show, using a formal proof in Isabelle/HOL, that this validation algorithm implements our semantics. We experimentally evaluate our verified validation algorithm and show that it is practical.

Financial assets exhibit dependence structures, i.e., movements of their prices or returns show various correlations. Knowledge of assets’ price dependencies can help investors to create a diversified portfolio, aiming to reduce portfolio risk due to the high volatility of the financial market. Since asset dependency changes with time in complex patterns, asset dependency forecast is an essential problem in finance. In this paper, we organize pairwise assets dependencies in an Asset Dependency Matrix (ADM) and formulate the problem of assets dependencies forecast to predict the future ADM given a sequence of past ADMs. We propose a novel idea viewing a sequence of ADMs as a sequence of images to capture the spatial and temporal dependencies among the assets. Inspired by video prediction tasks, we develop a novel Asset Dependency Neural Network (ADNN) to tackle the ADM prediction problem. Experiments show that our proposed framework consistently outperforms baselines on both future ADM prediction and portfolio risk reduction tasks.

Unconstrained lip-to-speech aims to generate corresponding speeches based on silent facial videos with no restriction to head pose or vocabulary. It is desirable to generate intelligible and natural speech with a fast speed in unconstrained settings. Currently, to handle the more complicated scenarios, most existing methods adopt the autoregressive architecture, which is optimized with the MSE loss. Although these methods have achieved promising performance, they are prone to bring issues including high inference latency and mel-spectrogram over-smoothness. To tackle these problems, we propose a novel flow-based non-autoregressive lip-to-speech model (GlowLTS) to break autoregressive constraints and achieve faster inference. Concretely, we adopt a flow-based decoder which is optimized by maximizing the likelihood of the training data and is capable of more natural and fast speech generation. Moreover, we devise a condition module to improve the intelligibility of generated speech. We demonstrate the superiority of our proposed method through objective and subjective evaluation on Lip2Wav-Chemistry-Lectures and Lip2Wav-Chess-Analysis datasets. Our demo video can be found at https://glowlts.github.io/.

Recent research on model interpretability in natural language processing extensively uses feature scoring methods for identifying which parts of the input are the most important for a model to make a prediction (i.e. explanation or rationale). However, previous research has shown that there is no clear best scoring method across various text classification tasks while practitioners typically have to make several other ad-hoc choices regarding the length and the type of the rationale (e.g. short or long, contiguous or not). Inspired by this, we propose a simple yet effective and flexible method that allows selecting optimally for each data instance: (1) a feature scoring method; (2) the length; and (3) the type of the rationale. Our method is inspired by input erasure approaches to interpretability which assume that the most faithful rationale for a prediction should be the one with the highest difference between the model's output distribution using the full text and the text after removing the rationale as input respectively. Evaluation on four standard text classification datasets shows that our proposed method provides more faithful, comprehensive and highly sufficient explanations compared to using a fixed feature scoring method, rationale length and type. More importantly, we demonstrate that a practitioner is not required to make any ad-hoc choices in order to extract faithful rationales using our approach.

Multi-Agent Path Finding (MAPF) is the problem of finding collision-free paths for multiple agents that minimize the sum of path costs. EECBS is a leading two-level algorithm that solves MAPF bounded-suboptimally, that is, within some factor w of the minimum sum of path costs C*. It uses focal search to find bounded-suboptimal paths on the low level and Explicit Estimation Search (EES) to resolve collisions on the high level. EES keeps track of a lower bound LB on C* to find paths whose sum of path costs is at most w LB in order to solve MAPF bounded-suboptimally. However, the costs of many paths are often much smaller than w times their minimum path costs, meaning that the sum of path costs is much smaller than w C*. In this paper, we therefore propose Flexible EECBS (FEECBS), which uses a flex(ible) distribution of the path costs (that relaxes the requirement to find bounded-suboptimal paths on the low level) in order to reduce the number of collisions that need to be resolved on the high level while still guaranteeing to solve MAPF bounded suboptimally. We address the drawbacks of flex distribution via techniques such as restrictions on the flex distribution, restarts of the high-level search with EECBS, and low-level focal-A* search. Our empirical evaluation shows that FEECBS substantially improves the efficiency of EECBS on MAPF instances with large maps and large numbers of agents.

The Moran process is a classic stochastic process that models invasion dynamics on graphs. A single mutant (e.g., a new opinion, strain, social trait etc.) invades a population of residents spread over the nodes of a graph. The mutant fitness advantage δ>=0 determines how aggressively mutants propagate to their neighbors. The quantity of interest is the fixation probability, i.e., the probability that the initial mutant eventually takes over the whole population. However, in realistic settings, the invading mutant has an advantage only in certain locations. E.g., the ability to metabolize a certain sugar is an advantageous trait to bacteria only when the sugar is actually present in their surroundings. In this paper we introduce the positional Moran process, a natural generalization in which the mutant fitness advantage is only realized on specific nodes called active nodes, and study the problem of fixation maximization: given a budget k, choose a set of k active nodes that maximize the fixation probability of the invading mutant. We show that the problem is NP-hard, while the optimization function is not submodular, thus indicating strong computational hardness. We focus on two natural limits. In the limit of δ to infinity (strong selection), although the problem remains NP-hard, the optimization function becomes submodular and thus admits a constant-factor approximation using a simple greedy algorithm. In the limit of δ to 0 (weak selection), we show that we can obtain a tight approximation in O(n^{2×ω}) time, where ω is the matrix-multiplication exponent. An experimental evaluation of the new algorithms along with some proposed heuristics corroborates our results.

Conventional convolution neural networks (CNNs) trained on narrow Field-of-View (FoV) images are the state-of-the art approaches for object recognition tasks. Some methods proposed the adaptation of CNNs to ultra-wide FoV images by learning deformable kernels. However, they are limited by the Euclidean geometry and their accuracy degrades under strong distortions caused by fisheye projections. In this work, we demonstrate that learning the shape of convolution kernels in non-Euclidean spaces is better than existing deformable kernel methods. In particular, we propose a new approach that learns deformable kernel parameters (positions) in hyperbolic space. FisheyeHDK is a hybrid CNN architecture combining hyperbolic and Euclidean convolution layers for positions and features learning. First, we provide intuition of hyperbolic space for wide FoV images. Using synthetic distortion profiles, we demonstrate the effectiveness of our approach. We select two datasets - Cityscapes and BDD100K 2020 - of perspective images which we transform to fisheye equivalents at different scaling factors (analogue to focal lengths). Finally, we provide an experiment on data collected by a real fisheye camera. Validations and experiments show that our approach improves existing deformable kernel methods for CNN adaptation on fisheye images.

This paper studies a function fitting problem which we coin first-order convex fitting (FCF): given any two vector sequences x1, ..., xT and p1, ..., pT, when is it possible to efficiently construct a convex function f(x) that ``fits'' the two sequences in the first-order sense, i.e, the (sub)gradient of f(xi) equals precisely pi, for all i = 1, ..., T? Despite a basic question of convex analysis, FCF has surprisingly been overlooked in the past literature. With an efficient constructive proof, we provide a clean answer to this question: FCF is possible if and only if the two sequences are permutation stable: p1 * x1 + ... + pT * xT is greater than or equal to p1 * x’1 + ... + pT * x’T where x’1, ..., x’T is any permutation of x1, ..., xT. We demonstrate the usefulness of FCF in two applications. First, we study how it can be used as an empirical risk minimization procedure to learn the original convex function. We provide efficient PAC-learnability bounds for special classes of convex functions learned via FCF, and demonstrate its application to multiple economic problems where only function gradients (as opposed to function values) can be observed. Second, we empirically show how it can be used as a surrogate to significantly accelerate the minimization of the original convex function.

We consider first-order (FO) rewritability for query answering in ontology mediated querying (OMQ) in which ontologies are formulated in Horn fragments of description logics (DLs). In general, OMQ approaches for such logics rely on non-FO rewriting of the query and/or on non-FO completion of the data, called a ABox. Specifically, we consider the problem of FO rewritability in terms of Beth definability, and show how Craig interpolation can then be used to effectively construct the rewritings, when they exist, from the Clark’s completion of Datalog-like programs encoding a given DL TBox and optionally a query. We show how this approach to FO rewritability can also be used to (a) capture integrity constraints commonly available in backend relational data sources, (b) capture constraints inherent in mapping such sources to an ABox , and (c) can be used an alternative to deriving so-called perfect rewritings of queries in the case of DL-Lite ontologies.