ions for Narrative Comprehension Tasks Yi-Chun Chen Justus Robertson Arnav Jhala Department of Computer Science North Carolina State University Raleigh, NC 27695 { ychen74 | jjrobert | ahjhala }@ncsu.edu

Zoom tracking is an essential and critical technology of digital cameras. Zoom tracking is typically achieved by moving the zoom and focus motors in lenses following the “trace curve”. The trace curve shows the in-focus motor positions versus the zoom motor positions for a specific object distance. Considering the limitations of the existing methods in responsiveness and accuracy, we propose a new zoom tracking algorithm based on deep learning, DPZT (Deep Predictive Zoom Tracking). The input of DPZT is a two-dimensional vector, consisting of image distance in two specific focus length. The structure of DPZT is a network with three hidden layers, which consists of 30, 30 and 3 neurons of each layer. Our method optimizes the mean offsets to 0.842 steps. With sufficient amount of field experiments, the newly proposed algorithm has proven its superiority as compared to the existing ones. Particularly, the architecture solves the one-to-many map problem at the root cause. Despite the slight increase in the storage capacity, the unprecedented enhancement of tracking accuracy and responsiveness outweighs the side effect.

Zero-shot transfer learning for multi-domain dialogue state tracking can allow us to handle new domains without incurring the high cost of data acquisition. This paper proposes new zero-short transfer learning technique for dialogue state tracking where the in-domain training data are all synthesized from an abstract dialogue model and the ontology of the domain. We show that data augmentation through synthesized data can improve the accuracy of zero-shot learning for both the TRADE model and the BERT-based SUMBT model on the MultiWOZ 2.1 dataset. We show training with only synthesized in-domain data on the SUMBT model can reach about 2/3 of the accuracy obtained with the full training dataset. We improve the zero-shot learning state of the art on average across domains by 21%.

Word representations are created using analogy context-based statistics and lexical relations on words. Word representations are inputs for the learning models in Natural Language Understanding (NLU) tasks. However, to understand language, knowing only the context is not sufficient. Reading between the lines is a key component of NLU. Embedding deeper word relationships which are not represented in the context enhances the word representation. This paper presents a word embedding which combines an analogy, context-based statistics using Word2Vec, and deeper word relationships using Conceptnet, to create an expanded word representation. In order to fine-tune the word representation, Self-Organizing Map is used to optimize it. The proposed word representation is compared with semantic word representations using Simlex 999. Furthermore, the use of 3D visual representations has shown to be capable of representing the similarity and association between words. The proposed word representation shows a Spearman correlation score of 0.886 and provided the best results when compared to the current state-of-the-art methods, and exceed the human performance of 0.78.

Word embeddings, made widely popular in 2013 with the release of word2vec, have become a mainstay of NLP engineering pipelines. Recently, with the release of BERT, word embeddings have moved from the term-based embedding space to the contextual embedding space—each term is no longer represented by a single low-dimensional vector but instead each term and its context determine the vector weights. BERT’s setup and architecture have been shown to be general enough to be applicable to many natural language tasks. Importantly for Information Retrieval (IR), in contrast to prior deep learning solutions to IR problems which required significant tuning of neural net architectures and training regimes, “vanilla BERT” has been shown to outperform existing retrieval algorithms by a wide margin, including on tasks and corpora that have long resisted retrieval effectiveness gains over traditional IR baselines (such as Robust04). In this paper, we employ the recently proposed axiomatic dataset analysis technique—that is, we create diagnostic datasets that each fulfil a retrieval heuristic (both term matching and semantic-based)—to explore what BERT is able to learn. In contrast to our expectations, we find BERT, when applied to a recently released large-scale web corpus with ad-hoc topics, to not adhere to any of the explored axioms. At the same time, BERT outperforms the traditional query likelihood retrieval model by 40%. This means that the axiomatic approach to IR (and its extension of diagnostic datasets created for retrieval heuristics) may in its current form not be applicable to large-scale corpora. Additional—different—axioms are needed.

Word embeddings have been shown to be useful across state-of-the-art systems in many natural language processing tasks, ranging from question answering systems to dependency parsing. (Herbelot and Vecchi, 2015) explored word embeddings and their utility for modeling language semantics. In particular, they presented an approach to automatically map a standard distributional semantic space onto a set-theoretic model using partial least squares regression. We show in this paper that a simple baseline achieves a +51% relative improvement compared to their model on one of the two datasets they used, and yields competitive results on the second dataset.

Word embeddings are traditionally trained on a large corpus in an unsupervised setting, with no specific design for incorporating domain knowledge. This can lead to unsatisfactory performances when training data originate from heterogeneous domains. In this paper, we propose two novel mechanisms for domain-aware word embedding training, namely domain indicator and domain attention, which integrate domain-specific knowledge into the widely used SG and CBOW models, respectively. The two methods are based on a joint learning paradigm and ensure that words in a target domain are intensively focused when trained on a source domain corpus. Qualitative and quantitative evaluation confirm the validity and effectiveness of our models. Compared to baseline methods, our method is particularly effective in near-cold-start scenarios.

With the tremendous advances made by Convolutional Neural Networks (ConvNets) on object recognition, we can now easily obtain adequately reliable machine-labeled annotations easily from predictions by off-the-shelf ConvNets. In this work, we present an abstraction memory based framework for few-shot learning, building upon machine-labeled image annotations. Our method takes large-scale machine-annotated dataset (e.g., OpenImages) as an external memory bank. In the external memory bank, the information is stored in the memory slots in the form of key-value, in which image feature is regarded as the key and the label embedding serves as the value. When queried by the few-shot examples, our model selects visually similar data from the external memory bank and writes the useful information obtained from related external data into another memory bank, i.e., abstraction memory. Long Short-Term Memory (LSTM) controllers and attention mechanisms are utilized to guarantee the data written to the abstraction memory correlates with the query example. The abstraction memory concentrates information from the external memory bank to make the few-shot recognition effective. In the experiments, we first confirm that our model can learn to conduct few-shot object recognition on clean human-labeled data from the ImageNet dataset. Then, we demonstrate that with our model, machine-labeled image annotations are very effective and abundant resources for performing object recognition on novel categories. Experimental results show that our proposed model with machine-labeled annotations achieves great results, with only a 1% difference in accuracy between the machine-labeled annotations and the human-labeled annotations.

With the successful development of artificial intelligence using deep learning, there has been growing interest in its deployment. The mobile environment is the closest hardware platform to real life, and it has become an important platform for the success or failure of artificial intelligence. Memory-augmented neural networks (MANNs) are neural networks proposed to efficiently handle question-and-answer (Q&A) tasks, well-suited for mobile devices. As a MANN requires various types of operations and recurrent data paths, it is difficult to accelerate the inference in the structure designed for other conventional neural network models, which is one of the biggest obstacles to deploying MANNs in mobile environments. To address the aforementioned issues, we propose Streaming MANN. This is the first attempt to implement and demonstrate the architecture for energy-efficient inference of MANNs with the concept of streaming processing. To achieve the full potential of the streaming process, we propose a novel approach, called inference thresholding, using Bayesian approach considering the characteristics of natural language processing (NLP) tasks. To evaluate our proposed approaches, we implemented the architecture and method in a field-programmable gate array (FPGA) which is suitable for streaming processing. We measured the execution time and power consumption of the inference for the bAbI dataset. The experimental results showed that the performance efficiency per energy (FLOPS/kJ) of the Streaming MANN increased by a factor of up to about 126 compared to the results of NVIDIA TITAN V, and up to 140 if inference thresholding is applied.

With the rise in mobile and voice search, answer passage retrieval acts as a critical component of an effective information retrieval system for open domain question answering. Currently, there are no comparable collections that address non-factoid question answering within larger documents while simultaneously providing enough examples sufficient to train a deep neural network. In this paper, we introduce a new Wikipedia based collection specific for non-factoid answer passage retrieval containing thousands of questions with annotated answers and show benchmark results on a variety of state of the art neural architectures and retrieval models. The experimental results demonstrate the unique challenges presented by answer passage retrieval within topically relevant documents for future research.

With the increasing computational demands of the neural networks, many hardware accelerators for the neural networks have been proposed. Such existing neural network accelerators often focus on popular neural network types such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs); however, not much attention has been paid to attention mechanisms, an emerging neural network primitive that enables neural networks to retrieve most relevant information from a knowledge-base, external memory, or past states. The attention mechanism is widely adopted by many state-of-the-art neural networks for computer vision, natural language processing, and machine translation, and accounts for a large portion of total execution time. We observe today's practice of implementing this mechanism using matrix-vector multiplication is suboptimal as the attention mechanism is semantically a content-based search where a large portion of computations ends up not being used. Based on this observation, we design and architect A3, which accelerates attention mechanisms in neural networks with algorithmic approximation and hardware specialization. Our proposed accelerator achieves multiple orders of magnitude improvement in energy efficiency (performance/watt) as well as substantial speedup over the state-of-the-art conventional hardware.

With the growing amount of textual information in recent years, it has become quite challenging to keep up with content produced by humans. Many models have been proposed that can perform reading comprehension on a variety of texts; however, past models either excel at information retrieval on complex texts or inference on simple texts. In this paper, we propose a model called ALICE that can perform information retrieval as well as inference tasks on any text. It is scalable to any document size and can be used to aid professionals in quickly finding answers to their problems using natural language queries. We will explore how ALICE achieves this and test it on some common datasets.

With the ever-growing mass of published text, natural language understanding stands as one of the most sought-after goal of artificial intelligence. In natural language, not every fact expressed in the text is necessarily explicit: human readers naturally infer what is missing through various intuitive linguistic skills, common sense or domain-specific knowledge, and life experiences. Natural Language Processing (NLP) systems do not have these initial capabilities. Unable to draw inferences to fill the gaps in the text, they cannot truly understand it. This dissertation focuses on this problem and presents our work on the automatic resolution of textual inferences in the context of machine reading. A textual inference is simply defined as a relation between two fragments of text: a human reading the first can reasonably infer that the second is true. A lot of different NLP tasks more or less directly evaluate systems on their ability to recognize textual inference. Among this multiplicity of evaluation frameworks, inferences themselves are not one and the same and also present a wide variety of different types. We reflect on inferences for NLP from a theoretical standpoint and present two contributions addressing these levels of diversity: an abstract contextualized inference task encompassing most NLP inference-related tasks, and a novel hierchical taxonomy of textual inferences based on their difficulty.Automatically recognizing textual inference currently almost always involves a machine learning model, trained to use various linguistic features on a labeled dataset of samples of textual inference. However, specific data on complex inference phenomena is not currently abundant enough that systems can directly learn world knowledge and commonsense reasoning. Instead, systems focus on learning how to use the syntactic structure of sentences to align the words of two semantically related sentences. To extend what systems know of the world, they include external background knowledge, often improving their results. But this addition is often made on top of other features, and rarely well integrated to sentence structure. The main contributions of our thesis address the previous concern, with the aim of solving complex natural language understanding tasks. With the hypothesis that a simpler lexicon should make easier to compare the sense of two sentences, we present a passage retrieval method using structured lexical expansion backed up by a simplifying dictionary. This simplification hypothesis is tested again in a contribution on textual entailment: syntactical paraphrases are extracted from the same dictionary and repeatedly applied on the first sentence to turn it into the second. We then present a machine learning kernel-based method recognizing sentence rewritings, with a notion of types able to encode lexical-semantic knowledge. This approach is effective on three tasks: paraphrase identification, textual entailment and question answering. We address its lack of scalability while keeping most of its strengths in our last contribution. Reading comprehension tests are used for evaluation: these multiple-choice questions on short text constitute the most practical way to assess textual inference within a complete context. Our system is founded on a efficient tree edit algorithm, and the features extracted from edit sequences are used to build two classifiers for the validation and invalidation of answer candidates. This approach reaches second place at the "Entrance Exams" CLEF 2015 challenge.

With machine learning successfully applied to new daunting problems almost every day, general AI starts looking like an attainable goal. However, most current research focuses instead on important but narrow applications, such as image classification or machine translation. We believe this to be largely due to the lack of objective ways to measure progress towards broad machine intelligence. In order to fill this gap, we propose here a set of concrete desiderata for general AI, together with a platform to test machines on how well they satisfy such desiderata, while keeping all further complexities to a minimum.

With advances in deep learning, neural network variants are becoming the dominant architecture for many NLP tasks. In this project, we apply several deep learning approaches to question answering, with a focus on the bAbI dataset.

While question answering (QA) with neural network, i.e. neural QA, has achieved promising results in recent years, lacking of large scale real-word QA dataset is still a challenge for developing and evaluating neural QA system. To alleviate this problem, we propose a large scale human annotated real-world QA dataset WebQA with more than 42k questions and 556k evidences. As existing neural QA methods resolve QA either as sequence generation or classification/ranking problem, they face challenges of expensive softmax computation, unseen answers handling or separate candidate answer generation component. In this work, we cast neural QA as a sequence labeling problem and propose an end-to-end sequence labeling model, which overcomes all the above challenges. Experimental results on WebQA show that our model outperforms the baselines significantly with an F1 score of 74.69% with word-based input, and the performance drops only 3.72 F1 points with more challenging character-based input.

While pre-trained language models (PLMs) are the go-to solution to tackle many natural language processing problems, they are still very limited in their ability to capture and to use common-sense knowledge. In fact, even if information is available in the form of approximate (soft) logical rules, it is not clear how to transfer it to a PLM in order to improve its performance for deductive reasoning tasks. Here, we aim to bridge this gap by teaching PLMs how to reason with soft Horn rules. We introduce a classification task where, given facts and soft rules, the PLM should return a prediction with a probability for a given hypothesis. We release the first dataset for this task, and we propose a revised loss function that enables the PLM to learn how to predict precise probabilities for the task. Our evaluation results show that the resulting fine-tuned models achieve very high performance, even on logical rules that were unseen at training. Moreover, we demonstrate that logical notions expressed by the rules are transferred to the finetuned model, yielding state-of-the-art results on external datasets.

While neural language models often perform surprisingly well on natural language understanding (NLU) tasks, their strengths and limitations remain poorly understood. Controlled synthetic tasks are thus an increasingly important resource for diagnosing model behavior. In this work we focus on story understanding, a core competency for NLU systems. However, the main synthetic resource for story understanding, the bAbI benchmark, lacks such a systematic mechanism for controllable task generation. We develop Dyna-bAbI, a dynamic framework providing fine-grained control over task generation in bAbI. We demonstrate our ideas by constructing three new tasks requiring compositional generalization, an important evaluation setting absent from the original benchmark. We tested both special-purpose models developed for bAbI as well as stateof-the-art pre-trained methods, and found that while both approaches solve the original tasks (>99% accuracy), neither approach succeeded in the compositional generalization setting, indicating the limitations of the original training data. We explored ways to augment the original data, and found that though diversifying training data was far more useful than simply increasing dataset size, it was still insufficient for driving robust compositional generalization (with <70% accuracy for complex compositions). Our results underscore the importance of highly controllable task generators for creating robust NLU systems through a virtuous cycle of model and data development.1

While large-scale knowledge graphs provide vast amounts of structured facts about entities, a short textual description can often be useful to succinctly characterize an entity and its type. Unfortunately, many knowledge graph entities lack such textual descriptions. In this paper, we introduce a dynamic memory-based network that generates a short open vocabulary description of an entity by jointly leveraging induced fact embeddings as well as the dynamic context of the generated sequence of words. We demonstrate the ability of our architecture to discern relevant information for more accurate generation of type description by pitting the system against several strong baselines.

While evaluating an answer choice for Reading Comprehension task, other answer choices available for the question and the answers of related questions about the same paragraph often provide valuable information. In this paper, we propose a method to leverage the natural language relations between the answer choices, such as entailment and contradiction, to improve the performance of machine comprehension. We use a stand-alone question answering (QA) system to perform QA task and a Natural Language Inference (NLI) system to identify the relations between the choice pairs. Then we perform inference using an Integer Linear Programming (ILP)-based relational framework to re-evaluate the decisions made by the standalone QA system in light of the relations identified by the NLI system. We also propose a multitask learning model that learns both the tasks jointly.