Large pre-trained sentence encoders like BERT start a new chapter in natural language processing. A common practice to apply pre-trained BERT to sequence classification tasks (e.g., classification of sentences or sentence pairs) is by feeding the embedding of [CLS] token (in the last layer) to a task-specific classification layer, and then fine tune the model parameters of BERT and classifier jointly. In this paper, we conduct systematic analysis over several sequence classification datasets to examine the embedding values of [CLS] token before the fine tuning phase, and present the biased embedding distribution issue---i.e., embedding values of [CLS] concentrate on a few dimensions and are non-zero centered. Such biased embedding brings challenge to the optimization process during fine-tuning as gradients of [CLS] embedding may explode and result in degraded model performance. We further propose several simple yet effective normalization methods to modify the [CLS] embedding during the fine-tuning. Compared with the previous practice, neural classification model with the normalized embedding shows improvements on several text classification tasks, demonstrates the effectiveness of our method.
Rational humans can generate sentences that cover a certain set of concepts while describing natural and common scenes. For example, given {apple(noun), tree(noun), pick(verb)}, humans can easily come up with scenes like "a boy is picking an apple from a tree" via their generative commonsense reasoning ability. However, we find this capacity has not been well learned by machines. Most prior works in machine commonsense focus on discriminative reasoning tasks with a multi-choice question answering setting. Herein, we present CommonGen: a challenging dataset for testing generative commonsense reasoning with a constrained text generation task. We collect 37k concept-sets as inputs and 90k human-written sentences as associated outputs. Additionally, we also provide high-quality rationales behind the reasoning process for the development and test sets from the human annotators. We demonstrate the difficulty of the task by examining a wide range of sequence generation methods with both automatic metrics and human evaluation. The state-of-the-art pre-trained generation model, UniLM, is still far from human performance in this task. Our data and code is publicly available at http://inklab.usc.edu/CommonGen/ .
The impressive performance of neural networks on natural language processing tasks attributes to their ability to model complicated word and phrase interactions. Existing flat, word level explanations of predictions hardly unveil how neural networks handle compositional semantics to reach predictions. To tackle the challenge, we study hierarchical explanation of neural network predictions. We identify non-additivity and independent importance attributions within hierarchies as two desirable properties for highlighting word and phrase interactions. We show prior efforts on hierarchical explanations, e.g. contextual decomposition, however, do not satisfy the desired properties mathematically. In this paper, we propose a formal way to quantify the importance of each word or phrase for hierarchical explanations. Following the formulation, we propose Sampling and Contextual Decomposition (SCD) algorithm and Sampling and Occlusion (SOC) algorithm. Human and metrics evaluation on both LSTM models and BERT Transformer models on multiple datasets show that our algorithms outperform prior hierarchical explanation algorithms. Our algorithms apply to hierarchical visualization of compositional semantics, extraction of classification rules and improving human trust of models.
Deep neural networks usually require massive labeled data, which restricts their applications in scenarios where data annotation is expensive. Natural language (NL) explanations have been demonstrated very useful additional supervision, which can provide sufficient domain knowledge for generating more labeled data over new instances, while the annotation time only doubles. However, directly applying them for augmenting model learning encounters two challenges: (1) NL explanations are unstructured and inherently compositional. (2) NL explanations often have large numbers of linguistic variants, resulting in low recall and limited generalization ability. In this paper, we propose a novel Neural EXecution Tree (NEXT) framework to augment training data for text classification using NL explanations. After transforming NL explanations into executable logical forms by semantic parsing, NEXT generalizes different types of actions specified by the logical forms for labeling data instances, which substantially increases the coverage of each NL explanation. Experiments on two NLP tasks (relation extraction and sentiment analysis) demonstrate its superiority over baseline methods. Its extension to multi-hop question answering achieves performance gain with light annotation effort.
Deep neural networks usually require massive labeled data, which restricts their applications in scenarios where data annotation is expensive. Natural language (NL) explanations have been demonstrated very useful additional supervision, which can provide sufficient domain knowledge for generating more labeled data over new instances, while the annotation time only doubles. However, directly applying them for augmenting model learning encounters two challenges: (1) NL explanations are unstructured and inherently compositional. (2) NL explanations often have large numbers of linguistic variants, resulting in low recall and limited generalization ability. In this paper, we propose a novel Neural EXecution Tree (NEXT) framework to augment training data for text classification using NL explanations. After transforming NL explanations into executable logical forms by semantic parsing, NEXT generalizes different types of actions specified by the logical forms for labeling data instances, which substantially increases the coverage of each NL explanation. Experiments on two NLP tasks (relation extraction and sentiment analysis) demonstrate its superiority over baseline methods. Its extension to multi-hop question answering achieves performance gain with light annotation effort.
Taxonomies are of great value to many knowledge-rich applications. As the manual taxonomy curation costs enormous human effects, automatic taxonomy construction is in great demand. However, most existing automatic taxonomy construction methods can only build hypernymy taxonomies wherein each edge is limited to expressing the "is-a" relation. Such a restriction limits their applicability to more diverse real-world tasks where the parent-child may carry different relations. In this paper, we aim to construct a task-guided taxonomy from a domain-specific corpus and allow users to input a "seed" taxonomy, serving as the task guidance. We propose an expansion-based taxonomy construction framework, namely HiExpan, which automatically generates key term list from the corpus and iteratively grows the seed taxonomy. Specifically, HiExpan views all children under each taxonomy node forming a coherent set and builds the taxonomy by recursively expanding all these sets. Furthermore, HiExpan incorporates a weakly-supervised relation extraction module to extract the initial children of a newly-expanded node and adjusts the taxonomy tree by optimizing its global structure. Our experiments on three real datasets from different domains demonstrate the effectiveness of HiExpan for building task-guided taxonomies.
Corpus-based set expansion (i.e., finding the "complete" set of entities belonging to the same semantic class, based on a given corpus and a tiny set of seeds) is a critical task in knowledge discovery. It may facilitate numerous downstream applications, such as information extraction, taxonomy induction, question answering, and web search. To discover new entities in an expanded set, previous approaches either make one-time entity ranking based on distributional similarity, or resort to iterative pattern-based bootstrapping. The core challenge for these methods is how to deal with noisy context features derived from free-text corpora, which may lead to entity intrusion and semantic drifting. In this study, we propose a novel framework, SetExpan, which tackles this problem, with two techniques: (1) a context feature selection method that selects clean context features for calculating entity-entity distributional similarity, and (2) a ranking-based unsupervised ensemble method for expanding entity set based on denoised context features. Experiments on three datasets show that SetExpan is robust and outperforms previous state-of-the-art methods in terms of mean average precision.
Sequence labeling is a fundamental framework for various natural language processing problems. Its performance is largely influenced by the annotation quality and quantity in supervised learning scenarios. In many cases, ground truth labels are costly and time-consuming to collect or even non-existent, while imperfect ones could be easily accessed or transferred from different domains. In this paper, we propose a novel framework named consensus Network (ConNet) to conduct training with imperfect annotations from multiple sources. It learns the representation for every weak supervision source and dynamically aggregates them by a context-aware attention mechanism. Finally, it leads to a model reflecting the consensus among multiple sources. We evaluate the proposed framework in two practical settings of multisource learning: learning with crowd annotations and unsupervised cross-domain model adaptation. Extensive experimental results show that our model achieves significant improvements over existing methods in both settings.
While deep neural models have gained successes on information extraction tasks, they become less reliable when the amount of labeled data is limited. In this paper, we study relation extraction (RE) under low-resource setting, where only some (hand-built) labeling rules are provided for learning a neural model over a large, unlabeled corpus. To overcome the low-coverage issue of current bootstrapping methods (i.e., hard grounding of rules), we propose a Neural Rule Grounding (REGD) framework for jointly learning a relation extraction module (with flexible neural architecture) and a sentence-rule soft matching module. The soft matching module extends the coverage of rules on semantically similar instances and augments the learning on unlabeled corpus. Experiments on two public datasets demonstrate the effectiveness of REGD when compared with both rule-based and semi-supervised baselines. Additionally, the learned soft matching module is able to predict on new relations with unseen rules, and can provide interpretation on matching results.