We introduce a new large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure. We show that training models on this new dataset leads to state-of-the-art performance on a variety of popular NLI benchmarks, while posing a more difficult challenge with its new test set. Our analysis sheds light on the shortcomings of current state-of-the-art models, and shows that non-expert annotators are successful at finding their weaknesses. The data collection method can be applied in a never-ending learning scenario, becoming a moving target for NLU, rather than a static benchmark that will quickly saturate.
These days, vast amounts of knowledge are available online, most of it in written form. Search engines help us access this knowledge, but aggregating, relating and reasoning with it is still a predominantly human effort. One of the key challenges for automated reasoning based on natural-language texts is the need to extract meaning (semantics) from texts. Natural language understanding (NLU) systems describe the conversion from a set of natural language utterances to terms in a particular logic. Tools for the co-development of grammar and target logic are currently largely missing. We will describe the Grammatical Logical Framework (GLF), a combination of two existing frameworks, in which large parts of a symbolic, rule-based NLU system can be developed and implemented: the Grammatical Framework (GF) and MMT. GF is a tool for syntactic analysis, generation, and translation with complex natural language grammars and MMT can be used to specify logical systems and to represent knowledge in them. Combining these tools is possible, because they are based on compatible logical frameworks: Martin-L\"of type theory and LF. The flexibility of logical frameworks is needed, as NLU research has not settled on a particular target logic for meaning representation. Instead, new logics are developed all the time to handle various language phenomena. GLF allows users to develop the logic and the language parsing components in parallel, and to connect them for experimentation with the entire pipeline.
Spoken Language Understanding (SLU) typically comprises of an automatic speech recognition (ASR) followed by a natural language understanding (NLU) module. The two modules process signals in a blocking sequential fashion, i.e., the NLU often has to wait for the ASR to finish processing on an utterance basis, potentially leading to high latencies that render the spoken interaction less natural. In this paper, we propose recurrent neural network (RNN) based incremental processing towards the SLU task of intent detection. The proposed methodology offers lower latencies than a typical SLU system, without any significant reduction in system accuracy. We introduce and analyze different recurrent neural network architectures for incremental and online processing of the ASR transcripts and compare it to the existing offline systems. A lexical End-of-Sentence (EOS) detector is proposed for segmenting the stream of transcript into sentences for intent classification. Intent detection experiments are conducted on benchmark ATIS, Snips and Facebook's multilingual task oriented dialog datasets modified to emulate a continuous incremental stream of words with no utterance demarcation. We also analyze the prospects of early intent detection, before EOS, with our proposed system.
Entity Coreference Resolution is the task of resolving all mentions in a document that refer to the same real world entity and is considered as one of the most difficult tasks in natural language understanding. It is of great importance for downstream natural language processing tasks such as entity linking, machine translation, summarization, chatbots, etc. This work aims to give a detailed review of current progress on solving Coreference Resolution using neural-based approaches. It also provides a detailed appraisal of the datasets and evaluation metrics in the field, as well as the subtask of Pronoun Resolution that has seen various improvements in the recent years. We highlight the advantages and disadvantages of the approaches, the challenges of the task, the lack of agreed-upon standards in the task and propose a way to further expand the boundaries of the field.
Expanding new functionalities efficiently is an ongoing challenge for single-turn task-oriented dialogue systems. In this work, we explore functionality-specific semi-supervised learning via self-training. We consider methods that augment training data automatically from unlabeled data sets in a functionality-targeted manner. In addition, we examine multiple techniques for efficient selection of augmented utterances to reduce training time and increase diversity. First, we consider paraphrase detection methods that attempt to find utterance variants of labeled training data with good coverage. Second, we explore sub-modular optimization based on n-grams features for utterance selection. Experiments show that functionality-specific self-training is very effective for improving system performance. In addition, methods optimizing diversity can reduce training data in many cases to 50% with little impact on performance.
In Natural Language Generation (NLG), End-to-End (E2E) systems trained through deep learning have recently gained a strong interest. Such deep models need a large amount of carefully annotated data to reach satisfactory performance. However, acquiring such datasets for every new NLG application is a tedious and time-consuming task. In this paper, we propose a semi-supervised deep learning scheme that can learn from non-annotated data and annotated data when available. It uses an NLG and a Natural Language Understanding (NLU) sequence-to-sequence models which are learned jointly to compensate for the lack of annotation. Experiments on two benchmark datasets show that, with limited amount of annotated data, the method can achieve very competitive results while not using any pre-processing or re-scoring tricks. These findings open the way to the exploitation of non-annotated datasets which is the current bottleneck for the E2E NLG system development to new applications.
One of the goals of natural language understanding is to develop models that map sentences into meaning representations. However, training such models requires expensive annotation of complex structures, which hinders their adoption. Learning to actively-learn (LTAL) is a recent paradigm for reducing the amount of labeled data by learning a policy that selects which samples should be labeled. In this work, we examine LTAL for learning semantic representations, such as QA-SRL. We show that even an oracle policy that is allowed to pick examples that maximize performance on the test set (and constitutes an upper bound on the potential of LTAL), does not substantially improve performance compared to a random policy. We investigate factors that could explain this finding and show that a distinguishing characteristic of successful applications of LTAL is the interaction between optimization and the oracle policy selection process. In successful applications of LTAL, the examples selected by the oracle policy do not substantially depend on the optimization procedure, while in our setup the stochastic nature of optimization strongly affects the examples selected by the oracle. We conclude that the current applicability of LTAL for improving data efficiency in learning semantic meaning representations is limited.
What information from an act of sentence understanding is robustly represented in the human brain? We investigate this question by comparing sentence encoding models on a brain decoding task, where the sentence that an experimental participant has seen must be predicted from the fMRI signal evoked by the sentence. We take a pre-trained BERT architecture as a baseline sentence encoding model and fine-tune it on a variety of natural language understanding (NLU) tasks, asking which lead to improvements in brain-decoding performance. We find that none of the sentence encoding tasks tested yield significant increases in brain decoding performance. Through further task ablations and representational analyses, we find that tasks which produce syntax-light representations yield significant improvements in brain decoding performance. Our results constrain the space of NLU models that could best account for human neural representations of language, but also suggest limits on the possibility of decoding fine-grained syntactic information from fMRI human neuroimaging.
We present a new neural architecture for wide-coverage Natural Language Understanding in Spoken Dialogue Systems. We develop a hierarchical multi-task architecture, which delivers a multi-layer representation of sentence meaning (i.e., Dialogue Acts and Frame-like structures). The architecture is a hierarchy of self-attention mechanisms and BiLSTM encoders followed by CRF tagging layers. We describe a variety of experiments, showing that our approach obtains promising results on a dataset annotated with Dialogue Acts and Frame Semantics. Moreover, we demonstrate its applicability to a different, publicly available NLU dataset annotated with domain-specific intents and corresponding semantic roles, providing overall performance higher than state-of-the-art tools such as RASA, Dialogflow, LUIS, and Watson. For example, we show an average 4.45% improvement in entity tagging F-score over Rasa, Dialogflow and LUIS.
The identification of syllables within phonetic sequences is known as syllabification. This task is thought to play an important role in natural language understanding, speech production, and the development of speech recognition systems. The concept of the syllable is cross-linguistic, though formal definitions are rarely agreed upon, even within a language. In response, data-driven syllabification methods have been developed to learn from syllabified examples. These methods often employ classical machine learning sequence labeling models. In recent years, recurrence-based neural networks have been shown to perform increasingly well for sequence labeling tasks such as named entity recognition (NER), part of speech (POS) tagging, and chunking. We present a novel approach to the syllabification problem which leverages modern neural network techniques. Our network is constructed with long short-term memory (LSTM) cells, a convolutional component, and a conditional random field (CRF) output layer. Existing syllabification approaches are rarely evaluated across multiple language families. To demonstrate cross-linguistic generalizability, we show that the network is competitive with state of the art systems in syllabifying English, Dutch, Italian, French, Manipuri, and Basque datasets.
The state-of-the-art neural network architectures make it possible to create spoken language understanding systems with high quality and fast processing time. One major challenge for real-world applications is the high latency of these systems caused by triggered actions with high executions times. If an action can be separated into subactions, the reaction time of the systems can be improved through incremental processing of the user utterance and starting subactions while the utterance is still being uttered. In this work, we present a model-agnostic method to achieve high quality in processing incrementally produced partial utterances. Based on clean and noisy versions of the ATIS dataset, we show how to create datasets with our method to create low-latency natural language understanding components. We get improvements of up to 47.91 absolute percentage points in the metric F1-score.
Recently, pre-trained language representation flourishes as the mainstay of the natural language understanding community, e.g., BERT. These pre-trained language representations can create state-of-the-art results on a wide range of downstream tasks. Along with continuous significant performance improvement, the size and complexity of these pre-trained neural models continue to increase rapidly. Is it possible to compress these large-scale language representation models? How will the pruned language representation affect the downstream multi-task transfer learning objectives? In this paper, we propose Reweighted Proximal Pruning (RPP), a new pruning method specifically designed for a large-scale language representation model. Through experiments on SQuAD and the GLUE benchmark suite, we show that proximal pruned BERT keeps high accuracy for both the pre-training task and the downstream multiple fine-tuning tasks at high prune ratio. RPP provides a new perspective to help us analyze what large-scale language representation might learn. Additionally, RPP makes it possible to deploy a large state-of-the-art language representation model such as BERT on a series of distinct devices (e.g., online servers, mobile phones, and edge devices).
Adversarial training, which minimizes the maximal risk for label-preserving input perturbations, has proved to be effective for improving the generalization of language models. In this work, we propose a novel adversarial training algorithm, FreeLB, that promotes higher invariance in the embedding space, by adding adversarial perturbations to word embeddings and minimizing the resultant adversarial risk inside different regions around input samples. To validate the effectiveness of the proposed approach, we apply it to Transformer-based models for natural language understanding and commonsense reasoning tasks. Experiments on the GLUE benchmark show that when applied only to the finetuning stage, it is able to improve the overall test scores of BERT-base model from 78.3 to 79.4, and RoBERTa-large model from 88.5 to 88.8. In addition, the proposed approach achieves state-of-the-art single-model test accuracies of 85.44\% and 67.75\% on ARC-Easy and ARC-Challenge. Experiments on CommonsenseQA benchmark further demonstrate that FreeLB can be generalized and boost the performance of RoBERTa-large model on other tasks as well. Code is available at \url{https://github.com/zhuchen03/FreeLB .
Large neural language models trained on massive amounts of text have emerged as a formidable strategy for Natural Language Understanding tasks. However, the strength of these models as Natural Language Generators is less clear. Though anecdotal evidence suggests that these models generate better quality text, there has been no detailed study characterizing their generation abilities. In this work, we compare the performance of an extensively pretrained model, OpenAI GPT2-117 (Radford et al., 2019), to a state-of-the-art neural story generation model (Fan et al., 2018). By evaluating the generated text across a wide variety of automatic metrics, we characterize the ways in which pretrained models do, and do not, make better storytellers. We find that although GPT2-117 conditions more strongly on context, is more sensitive to ordering of events, and uses more unusual words, it is just as likely to produce repetitive and under-diverse text when using likelihood-maximizing decoding algorithms.
Language model pre-training, such as BERT, has significantly improved the performances of many natural language processing tasks. However, pre-trained language models are usually computationally expensive, so it is difficult to efficiently execute them on resource-restricted devices. To accelerate inference and reduce model size while maintaining accuracy, we first propose a novel Transformer distillation method that is specially designed for knowledge distillation (KD) of the Transformer-based models. By leveraging this new KD method, the plenty of knowledge encoded in a large teacher BERT can be effectively transferred to a small student Tiny-BERT. Then, we introduce a new two-stage learning framework for TinyBERT, which performs Transformer distillation at both the pretraining and task-specific learning stages. This framework ensures that TinyBERT can capture he general-domain as well as the task-specific knowledge in BERT. TinyBERT with 4 layers is empirically effective and achieves more than 96.8% the performance of its teacher BERTBASE on GLUE benchmark, while being 7.5x smaller and 9.4x faster on inference. TinyBERT with 4 layers is also significantly better than 4-layer state-of-the-art baselines on BERT distillation, with only about 28% parameters and about 31% inference time of them. Moreover, TinyBERT with 6 layers performs on-par with its teacher BERTBASE.
Recently, pretrained language models (e.g., BERT) have achieved great success on many downstream natural language understanding tasks and exhibit a certain level of commonsense reasoning ability. However, their performance on commonsense tasks is still far from that of humans. As a preliminary attempt, we propose a simple yet effective method to teach pretrained models with commonsense reasoning by leveraging the structured knowledge in ConceptNet, the largest commonsense knowledge base (KB). Specifically, the structured knowledge in KB allows us to construct various logical forms, and then generate multiple-choice questions requiring commonsense logical reasoning. Experimental results demonstrate that, when refined on these training examples, the pretrained models consistently improve their performance on tasks that require commonsense reasoning, especially in the few-shot learning setting. Besides, we also perform analysis to understand which logical relations are more relevant to commonsense reasoning.
As an attempt to combine extractive and abstractive summarization, Sentence Rewriting models adopt the strategy of extracting salient sentences from a document first and then paraphrasing the selected ones to generate a summary. However, the existing models in this framework mostly rely on sentence-level rewards or suboptimal labels, causing a mismatch between a training objective and evaluation metric. In this paper, we present a novel training signal that directly maximizes summary-level ROUGE scores through reinforcement learning. In addition, we incorporate BERT into our model, making good use of its ability on natural language understanding. In extensive experiments, we show that a combination of our proposed model and training procedure obtains new state-of-the-art performance on both CNN/Daily Mail and New York Times datasets. We also demonstrate that it generalizes better on DUC-2002 test set.
Natural Language Understanding (NLU) is a core component of dialog systems. It typically involves two tasks - intent classification (IC) and slot labeling (SL), which are then followed by a dialogue management (DM) component. Such NLU systems cater to utterances in isolation, thus pushing the problem of context management to DM. However, contextual information is critical to the correct prediction of intents and slots in a conversation. Prior work on contextual NLU has been limited in terms of the types of contextual signals used and the understanding of their impact on the model. In this work, we propose a context-aware self-attentive NLU (CASA-NLU) model that uses multiple signals, such as previous intents, slots, dialog acts and utterances over a variable context window, in addition to the current user utterance. CASA-NLU outperforms a recurrent contextual NLU baseline on two conversational datasets, yielding a gain of up to 7% on the IC task for one of the datasets. Moreover, a non-contextual variant of CASA-NLU achieves state-of-the-art performance for IC task on standard public datasets - Snips and ATIS.
Grounding is crucial for natural language understanding. An important subtask is to understand modified color expressions, such as 'dirty blue'. We present a model of color modifiers that, compared with previous additive models in RGB space, learns more complex transformations. In addition, we present a model that operates in the HSV color space. We show that certain adjectives are better modeled in that space. To account for all modifiers, we train a hard ensemble model that selects a color space depending on the modifier color pair. Experimental results show significant and consistent improvements compared to the state-of-the-art baseline model.
In this work we present Ludwig, a flexible, extensible and easy to use toolbox which allows users to train deep learning models and use them for obtaining predictions without writing code. Ludwig implements a novel approach to deep learning model building based on two main abstractions: data types and declarative configuration files. The data type abstraction allows for easier code and sub-model reuse, and the standardized interfaces imposed by this abstraction allow for encapsulation and make the code easy to extend. Declarative model definition configuration files enable inexperienced users to obtain effective models and increase the productivity of expert users. Alongside these two innovations, Ludwig introduces a general modularized deep learning architecture called Encoder-Combiner-Decoder that can be instantiated to perform a vast amount of machine learning tasks. These innovations make it possible for engineers, scientists from other fields and, in general, a much broader audience to adopt deep learning models for their tasks, concretely helping in its democratization.