Model robustness to bias is often determined by the generalization on carefully designed out-of-distribution datasets. Recent debiasing methods in natural language understanding (NLU) improve performance on such datasets by pressuring models into making unbiased predictions. An underlying assumption behind such methods is that this also leads to the discovery of more robust features in the model's inner representations. We propose a general probing-based framework that allows for post-hoc interpretation of biases in language models, and use an information-theoretic approach to measure the extractability of certain biases from the model's representations. We experiment with several NLU datasets and known biases, and show that, counter-intuitively, the more a language model is pushed towards a debiased regime, the more bias is actually encoded in its inner representations.
Dialogue-based relation extraction (RE) aims to extract relation(s) between two arguments that appear in a dialogue. Because dialogues have the characteristics of high personal pronoun occurrences and low information density, and since most relational facts in dialogues are not supported by any single sentence, dialogue-based relation extraction requires a comprehensive understanding of dialogue. In this paper, we propose the TUrn COntext awaRE Graph Convolutional Network (TUCORE-GCN) modeled by paying attention to the way people understand dialogues. In addition, we propose a novel approach which treats the task of emotion recognition in conversations (ERC) as a dialogue-based RE. Experiments on a dialogue-based RE dataset and three ERC datasets demonstrate that our model is very effective in various dialogue-based natural language understanding tasks. In these experiments, TUCORE-GCN outperforms the state-of-the-art models on most of the benchmark datasets. Our code is available at https://github.com/BlackNoodle/TUCORE-GCN.
Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively. In this work, leveraging the best of both worlds, we propose an acquisition function that opts for selecting \textit{contrastive examples}, i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods. We compare our approach, CAL (Contrastive Active Learning), with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. Our experiments show that CAL performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. We also conduct an extensive ablation study of our method and we further analyze all actively acquired datasets showing that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies.
Combining a pretrained language model (PLM) with textual patterns has been shown to help in both zero- and few-shot settings. For zero-shot performance, it makes sense to design patterns that closely resemble the text seen during self-supervised pretraining because the model has never seen anything else. Supervised training allows for more flexibility. If we allow for tokens outside the PLM's vocabulary, patterns can be adapted more flexibly to a PLM's idiosyncrasies. Contrasting patterns where a "token" can be any continuous vector vs. those where a discrete choice between vocabulary elements has to be made, we call our method CONtinuous pAtterNs (CONAN). We evaluate CONAN on two established benchmarks for lexical inference in context (LIiC) a.k.a. predicate entailment, a challenging natural language understanding task with relatively small training sets. In a direct comparison with discrete patterns, CONAN consistently leads to improved performance, setting a new state of the art. Our experiments give valuable insights into the kind of pattern that enhances a PLM's performance on LIiC and raise important questions regarding our understanding of PLMs using text patterns.
Existing Natural Language Understanding (NLU) models have been shown to incorporate dataset biases leading to strong performance on in-distribution (ID) test sets but poor performance on out-of-distribution (OOD) ones. We introduce a simple yet effective debiasing framework whereby the shallow representations of the main model are used to derive a bias model and both models are trained simultaneously. We demonstrate on three well studied NLU tasks that despite its simplicity, our method leads to competitive OOD results. It significantly outperforms other debiasing approaches on two tasks, while still delivering high in-distribution performance.
In this paper, we present our Alexa Prize Grand Challenge 4 socialbot: Proto. Leveraging diverse sources of world knowledge, and powered by a suite of neural and rule-based natural language understanding modules, state-of-the-art neural generators, novel state-based deterministic generators, an ensemble of neural re-rankers, a robust post-processing algorithm, and an efficient overall conversation strategy, Proto strives to be able to converse coherently about a diverse range of topics of interest to humans, and provide a memorable experience to the user. In this paper we dissect and analyze the different components and conversation strategies implemented by our socialbot, which enables us to generate colloquial, empathetic, engaging, self-rectifying, factually correct, and on-topic response, which has helped us achieve consistent scores throughout the competition.
Large-scale conversational assistants like Alexa, Siri, Cortana and Google Assistant process every utterance using multiple models for domain, intent and named entity recognition. Given the decoupled nature of model development and large traffic volumes, it is extremely difficult to identify utterances processed erroneously by such systems. We address this challenge to detect domain classification errors using offline Transformer models. We combine utterance encodings from a RoBERTa model with the Nbest hypothesis produced by the production system. We then fine-tune end-to-end in a multitask setting using a small dataset of humanannotated utterances with domain classification errors. We tested our approach for detecting misclassifications from one domain that accounts for <0.5% of the traffic in a large-scale conversational AI system. Our approach achieves an F1 score of 30% outperforming a bi- LSTM baseline by 16.9% and a standalone RoBERTa model by 4.8%. We improve this further by 2.2% to 32.2% by ensembling multiple models.
Most benchmark datasets targeting commonsense reasoning focus on everyday scenarios: physical knowledge like knowing that you could fill a cup under a waterfall [Talmor et al., 2019], social knowledge like bumping into someone is awkward [Sap et al., 2019], and other generic situations. However, there is a rich space of commonsense inferences anchored to knowledge about specific entities: for example, deciding the truthfulness of a claim "Harry Potter can teach classes on how to fly on a broomstick." Can models learn to combine entity knowledge with commonsense reasoning in this fashion? We introduce CREAK, a testbed for commonsense reasoning about entity knowledge, bridging fact-checking about entities (Harry Potter is a wizard and is skilled at riding a broomstick) with commonsense inferences (if you're good at a skill you can teach others how to do it). Our dataset consists of 13k human-authored English claims about entities that are either true or false, in addition to a small contrast set. Crowdworkers can easily come up with these statements and human performance on the dataset is high (high 90s); we argue that models should be able to blend entity knowledge and commonsense reasoning to do well here. In our experiments, we focus on the closed-book setting and observe that a baseline model finetuned on existing fact verification benchmark struggles on CREAK. Training a model on CREAK improves accuracy by a substantial margin, but still falls short of human performance. Our benchmark provides a unique probe into natural language understanding models, testing both its ability to retrieve facts (e.g., who teaches at the University of Chicago?) and unstated commonsense knowledge (e.g., butlers do not yell at guests).
It is often observed in knowledge-centric tasks (e.g., common sense question and answering, relation classification) that the integration of external knowledge such as entity representation into language models can help provide useful information to boost the performance. However, it is still unclear whether this benefit can extend to general natural language understanding (NLU) tasks. In this work, we empirically investigated the contribution of external knowledge by measuring the end-to-end performance of language models with various knowledge integration methods. We find that the introduction of knowledge can significantly improve the results on certain tasks while having no adverse effects on other tasks. We then employ mutual information to reflect the difference brought by knowledge and a neural interpretation model to reveal how a language model utilizes external knowledge. Our study provides valuable insights and guidance for practitioners to equip NLP models with knowledge.
Language understanding in speech-based systems have attracted much attention in recent years with the growing demand for voice interface applications. However, the robustness of natural language understanding (NLU) systems to errors introduced by automatic speech recognition (ASR) is under-examined. %To facilitate the research on ASR-robust general language understanding, In this paper, we propose ASR-GLUE benchmark, a new collection of 6 different NLU tasks for evaluating the performance of models under ASR error across 3 different levels of background noise and 6 speakers with various voice characteristics. Based on the proposed benchmark, we systematically investigate the effect of ASR error on NLU tasks in terms of noise intensity, error type and speaker variants. We further purpose two ways, correction-based method and data augmentation-based method to improve robustness of the NLU systems. Extensive experimental results and analysises show that the proposed methods are effective to some extent, but still far from human performance, demonstrating that NLU under ASR error is still very challenging and requires further research.
Building machine learning models for natural language understanding (NLU) tasks relies heavily on labeled data. Weak supervision has been proven valuable when large amount of labeled data is unavailable or expensive to obtain. Existing works studying weak supervision for NLU either mostly focus on a specific task or simulate weak supervision signals from ground-truth labels. It is thus hard to compare different approaches and evaluate the benefit of weak supervision without access to a unified and systematic benchmark with diverse tasks and real-world weak labeling rules. In this paper, we propose such a benchmark, named WALNUT (semi-WeAkly supervised Learning for Natural language Understanding Testbed), to advocate and facilitate research on weak supervision for NLU. WALNUT consists of NLU tasks with different types, including document-level and token-level prediction tasks. WALNUT is the first semi-weakly supervised learning benchmark for NLU, where each task contains weak labels generated by multiple real-world weak sources, together with a small set of clean labels. We conduct baseline evaluations on WALNUT to systematically evaluate the effectiveness of various weak supervision methods and model architectures. Our results demonstrate the benefit of weak supervision for low-resource NLU tasks and highlight interesting patterns across tasks. We expect WALNUT to stimulate further research on methodologies to leverage weak supervision more effectively. The benchmark and code for baselines are available at \url{aka.ms/walnut_benchmark}.
The successes of Artificial Intelligence in recent years in areas such as image analysis, natural language understanding and strategy games have sparked interest from the world of finance. Specifically, there are high expectations, and ongoing engineering projects, regarding the creation of artificial agents, known as robotic traders, capable of juggling the financial markets with the skill of experienced human traders. Obvious economic implications aside, this is certainly an area of great scientific interest, due to the challenges that such a real context poses to the use of AI techniques. Precisely for this reason, we must be aware that artificial agents capable of operating at such levels are not just round the corner, and that there will be no simple answers, but rather a concurrence of various technologies and methods to the success of the effort. In the course of this article, we review the issues inherent in the design of effective robotic traders as well as the consequently applicable solutions, having in view the general objective of bringing the current state of the art of robo-trading up to the next level of intelligence, which we refer to as Cognitive Trading. Key to our approach is the joining of two methodological and technological directions which, although both deeply rooted in the disciplinary field of artificial intelligence, have so far gone their separate ways: heuristics and learning.
Building systems with capability of natural language understanding (NLU) has been one of the oldest areas of AI. An essential component of NLU is to detect logical succession of events contained in a text. The task of sentence ordering is proposed to learn succession of events with applications in AI tasks. The performance of previous works employing statistical methods is poor, while the neural networks-based approaches are in serious need of large corpora for model learning. In this paper, we propose a method for sentence ordering which does not need a training phase and consequently a large corpus for learning. To this end, we generate sentence embedding using BERT pre-trained model and measure sentence similarity using cosine similarity score. We suggest this score as an indicator of sequential events' level of coherence. We finally sort the sentences through brute-force search to maximize overall similarities of the sequenced sentences. Our proposed method outperformed other baselines on ROCStories, a corpus of 5-sentence human-made stories. The method is specifically more efficient than neural network-based methods when no huge corpus is available. Among other advantages of this method are its interpretability and needlessness to linguistic knowledge.
Recent advances in text representation have shown that training on large amounts of text is crucial for natural language understanding. However, models trained without predefined notions of topical interest typically require careful fine-tuning when transferred to specialized domains. When a sufficient amount of within-domain text may not be available, expanding a seed corpus of relevant documents from large-scale web data poses several challenges. First, corpus expansion requires scoring and ranking each document in the collection, an operation that can quickly become computationally expensive as the web corpora size grows. Relying on dense vector spaces and pairwise similarity adds to the computational expense. Secondly, as the domain concept becomes more nuanced, capturing the long tail of domain-specific rare terms becomes non-trivial, especially under limited seed corpora scenarios. In this paper, we consider the problem of fast approximate corpus expansion given a small seed corpus with a few relevant documents as a query, with the goal of capturing the long tail of a domain-specific set of concept terms. To efficiently collect large-scale domain-specific corpora with limited relevance feedback, we propose a novel truncated sparse document bit-vector representation, termed Signature Assisted Unsupervised Corpus Expansion (SAUCE). Experimental results show that SAUCE can reduce the computational burden while ensuring high within-domain lexical coverage.
Utterance-level intent detection and token-level slot filling are two key tasks for natural language understanding (NLU) in task-oriented systems. Most existing approaches assume that only a single intent exists in an utterance. However, there are often multiple intents within an utterance in real-life scenarios. In this paper, we propose a multi-intent NLU framework, called SLIM, to jointly learn multi-intent detection and slot filling based on BERT. To fully exploit the existing annotation data and capture the interactions between slots and intents, SLIM introduces an explicit slot-intent classifier to learn the many-to-one mapping between slots and intents. Empirical results on three public multi-intent datasets demonstrate (1) the superior performance of SLIM compared to the current state-of-the-art for NLU with multiple intents and (2) the benefits obtained from the slot-intent classifier.
Discovering the logical sequence of events is one of the cornerstones in Natural Language Understanding. One approach to learn the sequence of events is to study the order of sentences in a coherent text. Sentence ordering can be applied in various tasks such as retrieval-based Question Answering, document summarization, storytelling, text generation, and dialogue systems. Furthermore, we can learn to model text coherence by learning how to order a set of shuffled sentences. Previous research has relied on RNN, LSTM, and BiLSTM architecture for learning text language models. However, these networks have performed poorly due to the lack of attention mechanisms. We propose an algorithm for sentence ordering in a corpus of short stories. Our proposed method uses a language model based on Universal Transformers (UT) that captures sentences' dependencies by employing an attention mechanism. Our method improves the previous state-of-the-art in terms of Perfect Match Ratio (PMR) score in the ROCStories dataset, a corpus of nearly 100K short human-made stories. The proposed model includes three components: Sentence Encoder, Language Model, and Sentence Arrangement with Brute Force Search. The first component generates sentence embeddings using SBERT-WK pre-trained model fine-tuned on the ROCStories data. Then a Universal Transformer network generates a sentence-level language model. For decoding, the network generates a candidate sentence as the following sentence of the current sentence. We use cosine similarity as a scoring function to assign scores to the candidate embedding and the embeddings of other sentences in the shuffled set. Then a Brute Force Search is employed to maximize the sum of similarities between pairs of consecutive sentences.
This thesis investigates the controllability of deep learning-based, end-to-end, generative dialogue systems in both task-oriented and chit-chat scenarios. In particular, we study the different aspects of controlling generative dialogue systems, including controlling styles and topics and continuously adding and combining dialogue skills. In the three decades since the first dialogue system was commercialized, the basic architecture of such systems has remained substantially unchanged, consisting of four pipelined basic components, namely, natural language understanding (NLU), dialogue state tracking (DST), a dialogue manager (DM) and natural language generation (NLG). The dialogue manager, which is the critical component of the modularized system, controls the response content and style. This module is usually programmed by rules and is designed to be highly controllable and easily extendable. With the emergence of powerful "deep learning" architectures, end-to-end generative dialogue systems have been proposed to optimize overall system performance and simplify training. However, these systems cannot be easily controlled and extended as the modularized dialogue manager can. This is because a single neural system is used, which is usually a large pre-trained language model (e.g., GPT-2), and thus it is hard to surgically change desirable attributes (e.g., style, topics, etc.). More importantly, uncontrollable dialogue systems can generate offensive and even toxic responses. Therefore, in this thesis, we study controllable methods for end-to-end generative dialogue systems in task-oriented and chit-chat scenarios. Throughout the chapters, we describe 1) how to control the style and topics of chit-chat models, 2) how to continuously control and extend task-oriented dialogue systems, and 3) how to compose and control multi-skill dialogue models.
Twitch chats pose a unique problem in natural language understanding due to a large presence of neologisms, specifically emotes. There are a total of 8.06 million emotes, over 400k of which were used in the week studied. There is virtually no information on the meaning or sentiment of emotes, and with a constant influx of new emotes and drift in their frequencies, it becomes impossible to maintain an updated manually-labeled dataset. Our paper makes a two fold contribution. First we establish a new baseline for sentiment analysis on Twitch data, outperforming the previous supervised benchmark by 7.9% points. Secondly, we introduce a simple but powerful unsupervised framework based on word embeddings and k-NN to enrich existing models with out-of-vocabulary knowledge. This framework allows us to auto-generate a pseudo-dictionary of emotes and we show that we can nearly match the supervised benchmark above even when injecting such emote knowledge into sentiment classifiers trained on extraneous datasets such as movie reviews or Twitter.
Intent detection and slot filling are two main tasks in natural language understanding (NLU) for identifying users' needs from their utterances. These two tasks are highly related and often trained jointly. However, most previous works assume that each utterance only corresponds to one intent, ignoring the fact that a user utterance in many cases could include multiple intents. In this paper, we propose a novel Self-Distillation Joint NLU model (SDJN) for multi-intent NLU. First, we formulate multiple intent detection as a weakly supervised problem and approach with multiple instance learning (MIL). Then, we design an auxiliary loop via self-distillation with three orderly arranged decoders: Initial Slot Decoder, MIL Intent Decoder, and Final Slot Decoder. The output of each decoder will serve as auxiliary information for the next decoder. With the auxiliary knowledge provided by the MIL Intent Decoder, we set Final Slot Decoder as the teacher model that imparts knowledge back to Initial Slot Decoder to complete the loop. The auxiliary loop enables intents and slots to guide mutually in-depth and further boost the overall NLU performance. Experimental results on two public multi-intent datasets indicate that our model achieves strong performance compared to others.
Most previous methods for text data augmentation are limited to simple tasks and weak baselines. We explore data augmentation on hard tasks (i.e., few-shot natural language understanding) and strong baselines (i.e., pretrained models with over one billion parameters). Under this setting, we reproduced a large number of previous augmentation methods and found that these methods bring marginal gains at best and sometimes degrade the performance much. To address this challenge, we propose a novel data augmentation method FlipDA that jointly uses a generative model and a classifier to generate label-flipped data. Central to the idea of FlipDA is the discovery that generating label-flipped data is more crucial to the performance than generating label-preserved data. Experiments show that FlipDA achieves a good tradeoff between effectiveness and robustness -- it substantially improves many tasks while not negatively affecting the others.