In recent geospatial research, the importance of modeling large-scale human mobility data and predicting trajectories is rising, in parallel with progress in text generation using large-scale corpora in natural language processing. Whereas there are already plenty of feasible approaches applicable to geospatial sequence modeling itself, there seems to be room to improve with regard to evaluation, specifically about measuring the similarity between generated and reference trajectories. In this work, we propose a novel similarity measure, GEO-BLEU, which can be especially useful in the context of geospatial sequence modeling and generation. As the name suggests, this work is based on BLEU, one of the most popular measures used in machine translation research, while introducing spatial proximity to the idea of n-gram. We compare this measure with an established baseline, dynamic time warping, applying it to actual generated geospatial sequences. Using crowdsourced annotated data on the similarity between geospatial sequences collected from over 12,000 cases, we quantitatively and qualitatively show the proposed method's superiority.
Training neural machine translation (NMT) models in federated learning (FL) settings could be inefficient both computationally and communication-wise, due to the large size of translation engines as well as the multiple rounds of updates required to train clients and a central server. In this paper, we explore how to efficiently build NMT models in an FL setup by proposing a novel solution. In order to reduce the communication overhead, out of all neural layers we only exchange what we term "Controller" layers. Controllers are a small number of additional neural components connected to our pre-trained architectures. These new components are placed in between original layers. They act as liaisons to communicate with the central server and learn minimal information that is sufficient enough to update clients. We evaluated the performance of our models on five datasets from different domains to translate from German into English. We noted that the models equipped with Controllers preform on par with those trained in a central and non-FL setting. In addition, we observed a substantial reduction in the communication traffic of the FL pipeline, which is a direct consequence of using Controllers. Based on our experiments, Controller-based models are ~6 times less expensive than their other peers. This reduction is significantly important when we consider the number of parameters in large models and it becomes even more critical when such parameters need to be exchanged for multiple rounds in FL settings.
This study aims to develop a semi-automatically labelled prosody database for Hindi, for enhancing the intonation component in ASR and TTS systems, which is also helpful for building Speech to Speech Machine Translation systems. Although no single standard for prosody labelling exists in Hindi, researchers in the past have employed perceptual and statistical methods in literature to draw inferences about the behaviour of prosody patterns in Hindi. Based on such existing research and largely agreed upon theories of intonation in Hindi, this study attempts to first develop a manually annotated prosodic corpus of Hindi speech data, which is then used for training prediction models for generating automatic prosodic labels. A total of 5,000 sentences (23,500 words) for declarative and interrogative types have been labelled. The accuracy of the trained models for pitch accent, intermediate phrase boundaries and accentual phrase boundaries is 73.40%, 93.20%, and 43% respectively.
Continuously-growing data volumes lead to larger generic models. Specific use-cases are usually left out, since generic models tend to perform poorly in domain-specific cases. Our work addresses this gap with a method for selecting in-domain data from generic-domain (parallel text) corpora, for the task of machine translation. The proposed method ranks sentences in parallel general-domain data according to their cosine similarity with a monolingual domain-specific data set. We then select the top K sentences with the highest similarity score to train a new machine translation system tuned to the specific in-domain data. Our experimental results show that models trained on this in-domain data outperform models trained on generic or a mixture of generic and domain data. That is, our method selects high-quality domain-specific training instances at low computational cost and data size.
Recent approaches to Sign Language Production (SLP) have adopted spoken language Neural Machine Translation (NMT) architectures, applied without sign-specific modifications. In addition, these works represent sign language as a sequence of skeleton pose vectors, projected to an abstract representation with no inherent skeletal structure. In this paper, we represent sign language sequences as a skeletal graph structure, with joints as nodes and both spatial and temporal connections as edges. To operate on this graphical structure, we propose Skeletal Graph Self-Attention (SGSA), a novel graphical attention layer that embeds a skeleton inductive bias into the SLP model. Retaining the skeletal feature representation throughout, we directly apply a spatio-temporal adjacency matrix into the self-attention formulation. This provides structure and context to each skeletal joint that is not possible when using a non-graphical abstract representation, enabling fluid and expressive sign language production. We evaluate our Skeletal Graph Self-Attention architecture on the challenging RWTH-PHOENIX-Weather-2014T(PHOENIX14T) dataset, achieving state-of-the-art back translation performance with an 8% and 7% improvement over competing methods for the dev and test sets.
Natural language processing researchers have identified limitations of evaluation methodology for generation tasks, with new questions raised about the validity of automatic metrics and of crowdworker judgments. Meanwhile, efforts to improve generation models tend to depend on simple n-gram overlap metrics (e.g., BLEU, ROUGE). We argue that new advances on models and metrics should each more directly benefit and inform the other. We therefore propose a generalization of leaderboards, bidimensional leaderboards (Billboards), that simultaneously tracks progress in language generation models and metrics for their evaluation. Unlike conventional unidimensional leaderboards that sort submitted systems by predetermined metrics, a Billboard accepts both generators and evaluation metrics as competing entries. A Billboard automatically creates an ensemble metric that selects and linearly combines a few metrics based on a global analysis across generators. Further, metrics are ranked based on their correlation with human judgments. We release four Billboards for machine translation, summarization, and image captioning. We demonstrate that a linear ensemble of a few diverse metrics sometimes substantially outperforms existing metrics in isolation. Our mixed-effects model analysis shows that most automatic metrics, especially the reference-based ones, overrate machine over human generation, demonstrating the importance of updating metrics as generation models become stronger (and perhaps more similar to humans) in the future. Our project website is available at https://nlp.cs.washington.edu/billboard/.
We show that deep learning models, and especially architectures like the Transformer, originally intended for natural language, can be trained on randomly generated datasets to predict to very high accuracy both the qualitative and quantitative features of metabolic networks. Using standard mathematical techniques, we create large sets (40 million elements) of random networks that can be used to train our models. These trained models can predict network equilibrium on random graphs in more than 99% of cases. They can also generalize to graphs with different structure than those encountered at training. Finally, they can predict almost perfectly the equilibria of a small set of known biological networks. Our approach is both very economical in experimental data and uses only small and shallow deep-learning model, far from the large architectures commonly used in machine translation. Such results pave the way for larger use of deep learning models for problems related to biological networks in key areas such as quantitative systems pharmacology, systems biology, and synthetic biology.
Hashtag segmentation, also known as hashtag decomposition, is a common step in preprocessing pipelines for social media datasets. It usually precedes tasks such as sentiment analysis and hate speech detection. For sentiment analysis in medium to low-resourced languages, previous research has demonstrated that a multilingual approach that resorts to machine translation can be competitive or superior to previous approaches to the task. We develop a zero-shot hashtag segmentation framework and demonstrate how it can be used to improve the accuracy of multilingual sentiment analysis pipelines. Our zero-shot framework establishes a new state-of-the-art for hashtag segmentation datasets, surpassing even previous approaches that relied on feature engineering and language models trained on in-domain data.
Nowadays, due to the breakthrough in natural language generation (NLG), including machine translation, document summarization, image captioning, etc NLG models have been encapsulated in cloud APIs to serve over half a billion people worldwide and process over one hundred billion word generations per day. Thus, NLG APIs have already become essential profitable services in many commercial companies. Due to the substantial financial and intellectual investments, service providers adopt a pay-as-you-use policy to promote sustainable market growth. However, recent works have shown that cloud platforms suffer from financial losses imposed by model extraction attacks, which aim to imitate the functionality and utility of the victim services, thus violating the intellectual property (IP) of cloud APIs. This work targets at protecting IP of NLG APIs by identifying the attackers who have utilized watermarked responses from the victim NLG APIs. However, most existing watermarking techniques are not directly amenable for IP protection of NLG APIs. To bridge this gap, we first present a novel watermarking method for text generation APIs by conducting lexical modification to the original outputs. Compared with the competitive baselines, our watermark approach achieves better identifiable performance in terms of p-value, with fewer semantic losses. In addition, our watermarks are more understandable and intuitive to humans than the baselines. Finally, the empirical studies show our approach is also applicable to queries from different domains, and is effective on the attacker trained on a mixture of the corpus which includes less than 10\% watermarked samples.
In this paper, we present a corpus based study of politeness across two languages-English and Hindi. It studies the politeness in a translated parallel corpus of Hindi and English and sees how politeness in a Hindi text is translated into English. We provide a detailed theoretical background in which the comparison is carried out, followed by a brief description of the translated data within this theoretical model. Since politeness may become one of the major reasons of conflict and misunderstanding, it is a very important phenomenon to be studied and understood cross-culturally, particularly for such purposes as machine translation.
Transformer based language models have led to impressive results across all domains in Natural Language Processing. Pretraining these models on language modeling tasks and finetuning them on downstream tasks such as Text Classification, Question Answering and Neural Machine Translation has consistently shown exemplary results. In this work, we propose a Multitask Finetuning methodology which combines the Bilingual Machine Translation task with an auxiliary Causal Language Modeling task to improve performance on the former task on Indian Languages. We conduct an empirical study on three language pairs, Marathi-Hindi, Marathi-English and Hindi-English, where we compare the multitask finetuning approach to the standard finetuning approach, for which we use the mBART50 model. Our study indicates that the multitask finetuning method could be a better technique than standard finetuning, and could improve Bilingual Machine Translation across language pairs.
The inception of modeling contextual information using models such as BERT, ELMo, and Flair has significantly improved representation learning for words. It has also given SOTA results in almost every NLP task - Machine Translation, Text Summarization and Named Entity Recognition, to name a few. In this work, in addition to using these dominant context-aware representations, we propose a Knowledge Aware Representation Learning (KARL) Network for Named Entity Recognition (NER). We discuss the challenges of using existing methods in incorporating world knowledge for NER and show how our proposed methods could be leveraged to overcome those challenges. KARL is based on a Transformer Encoder that utilizes large knowledge bases represented as fact triplets, converts them to a graph context, and extracts essential entity information residing inside to generate contextualized triplet representation for feature augmentation. Experimental results show that the augmentation done using KARL can considerably boost the performance of our NER system and achieve significantly better results than existing approaches in the literature on three publicly available NER datasets, namely CoNLL 2003, CoNLL++, and OntoNotes v5. We also observe better generalization and application to a real-world setting from KARL on unseen entities.
In this paper, we explore machine translation improvement via Generative Adversarial Network (GAN) architecture. We take inspiration from RelGAN, a model for text generation, and NMT-GAN, an adversarial machine translation model, to implement a model that learns to transform awkward, non-fluent English sentences to fluent ones, while only being trained on monolingual corpora. We utilize a parameter $\lambda$ to control the amount of deviation from the input sentence, i.e. a trade-off between keeping the original tokens and modifying it to be more fluent. Our results improved upon phrase-based machine translation in some cases. Especially, GAN with a transformer generator shows some promising results. We suggests some directions for future works to build upon this proof-of-concept.
Spoken language understanding (SLU) tasks are usually solved by first transcribing an utterance with automatic speech recognition (ASR) and then feeding the output to a text-based model. Recent advances in self-supervised representation learning for speech data have focused on improving the ASR component. We investigate whether representation learning for speech has matured enough to replace ASR in SLU. We compare learned speech features from wav2vec 2.0, state-of-the-art ASR transcripts, and the ground truth text as input for a novel speech-based named entity recognition task, a cardiac arrest detection task on real-world emergency calls and two existing SLU benchmarks. We show that learned speech features are superior to ASR transcripts on three classification tasks. For machine translation, ASR transcripts are still the better choice. We highlight the intrinsic robustness of wav2vec 2.0 representations to out-of-vocabulary words as key to better performance.
Inducing latent tree structures from sequential data is an emerging trend in the NLP research landscape today, largely popularized by recent methods such as Gumbel LSTM and Ordered Neurons (ON-LSTM). This paper proposes FASTTREES, a new general purpose neural module for fast sequence encoding. Unlike most previous works that consider recurrence to be necessary for tree induction, our work explores the notion of parallel tree induction, i.e., imbuing our model with hierarchical inductive biases in a parallelizable, non-autoregressive fashion. To this end, our proposed FASTTREES achieves competitive or superior performance to ON-LSTM on four well-established sequence modeling tasks, i.e., language modeling, logical inference, sentiment analysis and natural language inference. Moreover, we show that the FASTTREES module can be applied to enhance Transformer models, achieving performance gains on three sequence transduction tasks (machine translation, subject-verb agreement and mathematical language understanding), paving the way for modular tree induction modules. Overall, we outperform existing state-of-the-art models on logical inference tasks by +4% and mathematical language understanding by +8%.
A tournament graph is a complete directed graph, which can be used to model a round-robin tournament between $n$ players. In this paper, we address the problem of finding a champion of the tournament, also known as Copeland winner, which is a player that wins the highest number of matches. In detail, we aim to investigate algorithms that find the champion by playing a low number of matches. Solving this problem allows us to speed up several Information Retrieval and Recommender System applications, including question answering, conversational search, etc. Indeed, these applications often search for the champion inducing a round-robin tournament among the players by employing a machine learning model to estimate who wins each pairwise comparison. Our contribution, thus, allows finding the champion by performing a low number of model inferences. We prove that any deterministic or randomized algorithm finding a champion with constant success probability requires $\Omega(\ell n)$ comparisons, where $\ell$ is the number of matches lost by the champion. We then present an asymptotically-optimal deterministic algorithm matching this lower bound without knowing $\ell$, and we extend our analysis to three variants of the problem. Lastly, we conduct a comprehensive experimental assessment of the proposed algorithms on a question answering task on public data. Results show that our proposed algorithms speed up the retrieval of the champion up to $13\times$ with respect to the state-of-the-art algorithm that perform the full tournament.
A key distinguishing feature of conversational recommender systems over traditional recommender systems is their ability to elicit user preferences using natural language. Currently, the predominant approach to preference elicitation is to ask questions directly about items or item attributes. Users searching for recommendations may not have deep knowledge of the available options in a given domain. As such, they might not be aware of key attributes or desirable values for them. However, in many settings, talking about the planned use of items does not present any difficulties, even for those that are new to a domain. In this paper, we propose a novel approach to preference elicitation by asking implicit questions based on item usage. As one of the main contributions of this work, we develop a multi-stage data annotation protocol using crowdsourcing, to create a high-quality labeled training dataset. Another main contribution is the development of four models for the question generation task: two template-based baseline models and two neural text-to-text models. The template-based models use heuristically extracted common patterns found in the training data, while the neural models use the training data to learn to generate questions automatically. Using common metrics from machine translation for automatic evaluation, we show that our approaches are effective in generating elicitation questions, even with limited training data. We further employ human evaluation for comparing the generated questions using both pointwise and pairwise evaluation designs. We find that the human evaluation results are consistent with the automatic ones, allowing us to draw conclusions about the quality of the generated questions with certainty. Finally, we provide a detailed analysis of cases where the models show their limitations.
Syntax knowledge contributes its powerful strength in Neural machine translation (NMT) tasks. Early NMT works supposed that syntax details can be automatically learned from numerous texts via attention networks. However, succeeding researches pointed out that limited by the uncontrolled nature of attention computation, the NMT model requires an external syntax to capture the deep syntactic awareness. Although existing syntax-aware NMT methods have born great fruits in combining syntax, the additional workloads they introduced render the model heavy and slow. Particularly, these efforts scarcely involve the Transformer-based NMT and modify its core self-attention network (SAN). To this end, we propose a parameter-free, Dependency-scaled Self-Attention Network (Deps-SAN) for syntax-aware Transformer-based NMT. A quantified matrix of dependency closeness between tokens is constructed to impose explicit syntactic constraints into the SAN for learning syntactic details and dispelling the dispersion of attention distributions. Two knowledge sparsing techniques are further integrated to avoid the model overfitting the dependency noises introduced by the external parser. Experiments and analyses on IWSLT14 German-to-English and WMT16 German-to-English benchmark NMT tasks verify the effectiveness of our approach.
In this paper, we describe the submission of the joint Samsung Research Philippines-Konvergen AI team for the WMT'21 Large Scale Multilingual Translation Task - Small Track 2. We submit a standard Seq2Seq Transformer model to the shared task without any training or architecture tricks, relying mainly on the strength of our data preprocessing techniques to boost performance. Our final submission model scored 22.92 average BLEU on the FLORES-101 devtest set, and scored 22.97 average BLEU on the contest's hidden test set, ranking us sixth overall. Despite using only a standard Transformer, our model ranked first in Indonesian to Javanese, showing that data preprocessing matters equally, if not more, than cutting edge model architectures and training techniques.
There has been a rapid advance of custom hardware (HW) for accelerating the inference speed of deep neural networks (DNNs). Previously, the softmax layer was not a main concern of DNN accelerating HW, because its portion is relatively small in multi-layer perceptron or convolutional neural networks. However, as the attention mechanisms are widely used in various modern DNNs, a cost-efficient implementation of softmax layer is becoming very important. In this paper, we propose two methods to approximate softmax computation, which are based on the usage of LookUp Tables (LUTs). The required size of LUT is quite small (about 700 Bytes) because ranges of numerators and denominators of softmax are stable if normalization is applied to the input. We have validated the proposed technique over different AI tasks (object detection, machine translation, sentiment analysis, and semantic equivalence) and DNN models (DETR, Transformer, BERT) by a variety of benchmarks (COCO17, WMT14, WMT17, GLUE). We showed that 8-bit approximation allows to obtain acceptable accuracy loss below $1.0\%$.