Traditional NLP has long held (supervised) syntactic parsing necessary for successful higher-level semantic language understanding (LU). The recent advent of end-to-end neural models, self-supervised via language modeling (LM), and their success on a wide range of LU tasks, however, questions this belief. In this work, we empirically investigate the usefulness of supervised parsing for semantic LU in the context of LM-pretrained transformer networks. Relying on the established fine-tuning paradigm, we first couple a pretrained transformer with a biaffine parsing head, aiming to infuse explicit syntactic knowledge from Universal Dependencies treebanks into the transformer. We then fine-tune the model for LU tasks and measure the effect of the intermediate parsing training (IPT) on downstream LU task performance. Results from both monolingual English and zero-shot language transfer experiments (with intermediate target-language parsing) show that explicit formalized syntax, injected into transformers through IPT, has very limited and inconsistent effect on downstream LU performance. Our results, coupled with our analysis of transformers' representation spaces before and after intermediate parsing, make a significant step towards providing answers to an essential question: how (un)availing is supervised parsing for high-level semantic natural language understanding in the era of large neural models?
In this paper, we study how to leverage pre-trained language models in Text-to-SQL. We argue that previous approaches under utilize the base language models by concatenating all columns together with the NL question and feeding them into the base language model in the encoding stage. We propose a neat approach called Hybrid Ranking Network (HydraNet) which breaks down the problem into column-wise ranking and decoding and finally assembles the column-wise outputs into a SQL query by straightforward rules. In this approach, the encoder is given a NL question and one individual column, which perfectly aligns with the original tasks BERT/RoBERTa is trained on, and hence we avoid any ad-hoc pooling or additional encoding layers which are necessary in prior approaches. Experiments on the WikiSQL dataset show that the proposed approach is very effective, achieving the top place on the leaderboard.
Attention-based sequence-to-sequence (seq2seq) models have achieved promising results in automatic speech recognition (ASR). However, as these models decode in a left-to-right way, they do not have access to context on the right. We leverage both left and right context by applying BERT as an external language model to seq2seq ASR through knowledge distillation. In our proposed method, BERT generates soft labels to guide the training of seq2seq ASR. Furthermore, we leverage context beyond the current utterance as input to BERT. Experimental evaluations show that our method significantly improves the ASR performance from the seq2seq baseline on the Corpus of Spontaneous Japanese (CSJ). Knowledge distillation from BERT outperforms that from a transformer LM that only looks at left context. We also show the effectiveness of leveraging context beyond the current utterance. Our method outperforms other LM application approaches such as n-best rescoring and shallow fusion, while it does not require extra inference cost.
Since the appearance of BERT, recent works including XLNet and RoBERTa utilize sentence embedding models pre-trained by large corpora and a large number of parameters. Because such models have large hardware and a huge amount of data, they take a long time to pre-train. Therefore it is important to attempt to make smaller models that perform comparatively. In this paper, we trained a Korean-specific model KR-BERT, utilizing a smaller vocabulary and dataset. Since Korean is one of the morphologically rich languages with poor resources using non-Latin alphabets, it is also important to capture language-specific linguistic phenomena that the Multilingual BERT model missed. We tested several tokenizers including our BidirectionalWordPiece Tokenizer and adjusted the minimal span of tokens for tokenization ranging from sub-character level to character-level to construct a better vocabulary for our model. With those adjustments, our KR-BERT model performed comparably and even better than other existing pre-trained models using a corpus about 1/10 of the size.
The recent advances in language modeling significantly improved the generative capabilities of deep neural models: in 2019 OpenAI released GPT-2, a pre-trained language model that can autonomously generate coherent, non-trivial and human-like text samples. Since then, ever more powerful text generative models have been developed. Adversaries can exploit these tremendous generative capabilities to enhance social bots that will have the ability to write plausible deepfake messages, hoping to contaminate public debate. To prevent this, it is crucial to develop deepfake social media messages detection systems. However, to the best of our knowledge no one has ever addressed the detection of machine-generated texts on social networks like Twitter or Facebook. With the aim of helping the research in this detection field, we collected the first dataset of \real deepfake tweets, TweepFake. It is real in the sense that each deepfake tweet was actually posted on Twitter. We collected tweets from a total of 23 bots, imitating 17 human accounts. The bots are based on various generation techniques, i.e., Markov Chains, RNN, RNN+Markov, LSTM, GPT-2. We also randomly selected tweets from the humans imitated by the bots to have an overall balanced dataset of 25,572 tweets (half human and half bots generated). The dataset is publicly available on Kaggle. Lastly, we evaluated 13 deepfake text detection methods (based on various state-of-the-art approaches) to both demonstrate the challenges that Tweepfake poses and create a solid baseline of detection techniques. We hope that TweepFake can offer the opportunity to tackle the deepfake detection on social media messages as well.
Scene text spotting aims to detect and recognize the entire word or sentence with multiple characters in natural images. It is still challenging because ambiguity often occurs when the spacing between characters is large or the characters are evenly spread in multiple rows and columns, making many visually plausible groupings of the characters (e.g. "BERLIN" is incorrectly detected as "BERL" and "IN" in Fig. 1(c)). Unlike previous works that merely employed visual features for text detection, this work proposes a novel text spotter, named Ambiguity Eliminating Text Spotter (AE TextSpotter), which learns both visual and linguistic features to significantly reduce ambiguity in text detection. The proposed AE TextSpotter has three important benefits. 1) The linguistic representation is learned together with the visual representation in a framework. To our knowledge, it is the first time to improve text detection by using a language model. 2) A carefully designed language module is utilized to reduce the detection confidence of incorrect text lines, making them easily pruned in the detection stage. 3) Extensive experiments show that AE TextSpotter outperforms other state-of-the-art methods by a large margin. For example, we carefully select a validation set of extremely ambiguous samples from the IC19-ReCTS dataset, where our approach surpasses other methods by more than 4%. The code has been released at https://github.com/whai362/AE_TextSpotter. The image list and evaluation scripts of the validation set have been released at https://github.com/whai362/TDA-ReCTS.
We introduce a deep and light-weight transformer, DeLighT, that delivers similar or better performance than standard transformer-based models with significantly fewer parameters. DeLighT more efficiently allocates parameters both (1) within each Transformer block using the DeLighT transformation, a deep and light-weight transformation, and (2) across blocks using block-wise scaling, which allows for shallower and narrower DeLighT blocks near the input and wider and deeper DeLighT blocks near the output. Overall, DeLighT networks are 2.5 to 4 times deeper than standard transformer models and yet have fewer parameters and operations. Experiments on benchmark machine translation and language modeling tasks show that DeLighT matches or improves the performance of baseline Transformers with 2 to 3 times fewer parameters on average. Our source code is available at: \url{https://github.com/sacmehta/delight}
It has been found that software, like natural language texts, exhibits "naturalness", which can be captured by statistical language models. In recent years, neural language models have been proposed to represent the naturalness of software through deep learning. In this paper, we conduct an experimental evaluation of state-of-the-art neural language models for source code, including RNN-based models and Transformer-XL based models. Through experiments on a large-scale Python code corpus, we find that the Transformer-XL model outperforms RNN-based models (including LSTM and GRU models) in capturing the naturalness of software, with far less computational cost.
Heavily pre-trained transformer models such as BERT have recently shown to be remarkably powerful at language modelling by achieving impressive results on numerous downstream tasks. It has also been shown that they are able to implicitly store factual knowledge in their parameters after pre-training. Understanding what the pre-training procedure of LMs actually learns is a crucial step for using and improving them for Conversational Recommender Systems (CRS). We first study how much off-the-shelf pre-trained BERT "knows" about recommendation items such as books, movies and music. In order to analyze the knowledge stored in BERT's parameters, we use different probes that require different types of knowledge to solve, namely content-based and collaborative-based. Content-based knowledge is knowledge that requires the model to match the titles of items with their content information, such as textual descriptions and genres. In contrast, collaborative-based knowledge requires the model to match items with similar ones, according to community interactions such as ratings. We resort to BERT's Masked Language Modelling head to probe its knowledge about the genre of items, with cloze style prompts. In addition, we employ BERT's Next Sentence Prediction head and representations' similarity to compare relevant and non-relevant search and recommendation query-document inputs to explore whether BERT can, without any fine-tuning, rank relevant items first. Finally, we study how BERT performs in a conversational recommendation downstream task. Overall, our analyses and experiments show that: (i) BERT has knowledge stored in its parameters about the content of books, movies and music; (ii) it has more content-based knowledge than collaborative-based knowledge; and (iii) fails on conversational recommendation when faced with adversarial data.
Existing tools for Question Answering (QA) have challenges that limit their use in practice. They can be complex to set up or integrate with existing infrastructure, do not offer configurable interactive interfaces, and do not cover the full set of subtasks that frequently comprise the QA pipeline (query expansion, retrieval, reading, and explanation/sensemaking). To help address these issues, we introduce NeuralQA - a usable library for QA on large datasets. NeuralQA integrates well with existing infrastructure (e.g., ElasticSearch instances and reader models trained with the HuggingFace Transformers API) and offers helpful defaults for QA subtasks. It introduces and implements contextual query expansion (CQE) using a masked language model (MLM) as well as relevant snippets (RelSnip) - a method for condensing large documents into smaller passages that can be speedily processed by a document reader model. Finally, it offers a flexible user interface to support workflows for research explorations (e.g., visualization of gradient-based explanations to support qualitative inspection of model behaviour) and large scale search deployment. Code and documentation for NeuralQA is available as open source on Github (https://github.com/victordibia/neuralqa}{Github).
This paper studies composer style classification of piano sheet music images. Previous approaches to the composer classification task have been limited by a scarcity of data. We address this issue in two ways: (1) we recast the problem to be based on raw sheet music images rather than a symbolic music format, and (2) we propose an approach that can be trained on unlabeled data. Our approach first converts the sheet music image into a sequence of musical "words" based on the bootleg feature representation, and then feeds the sequence into a text classifier. We show that it is possible to significantly improve classifier performance by first training a language model on a set of unlabeled data, initializing the classifier with the pretrained language model weights, and then finetuning the classifier on a small amount of labeled data. We train AWD-LSTM, GPT-2, and RoBERTa language models on all piano sheet music images in IMSLP. We find that transformer-based architectures outperform CNN and LSTM models, and pretraining boosts classification accuracy for the GPT-2 model from 46\% to 70\% on a 9-way classification task. The trained model can also be used as a feature extractor that projects piano sheet music into a feature space that characterizes compositional style.
Neural text decoding is important for generating high-quality texts using language models. To generate high-quality text, popular decoding algorithms like top-k, top-p (nucleus), and temperature-based sampling truncate or distort the unreliable low probability tail of the language model. Though these methods generate high-quality text after parameter tuning, they are ad hoc. Not much is known about the control they provide over the statistics of the output, which is important since recent reports show text quality is highest for a specific range of likelihoods. Here, first we provide a theoretical analysis of perplexity in top-k, top-p, and temperature sampling, finding that cross-entropy behaves approximately linearly as a function of p in top-p sampling whereas it is a nonlinear function of k in top-k sampling, under Zipfian statistics. We use this analysis to design a feedback-based adaptive top-k text decoding algorithm called mirostat that generates text (of any length) with a predetermined value of perplexity, and thereby high-quality text without any tuning. Experiments show that for low values of k and p in top-k and top-p sampling, perplexity drops significantly with generated text length, which is also correlated with excessive repetitions in the text (the boredom trap). On the other hand, for large values of k and p, we find that perplexity increases with generated text length, which is correlated with incoherence in the text (confusion trap). Mirostat avoids both traps: experiments show that cross-entropy has a near-linear relation with repetition in generated text. This relation is almost independent of the sampling method but slightly dependent on the model used. Hence, for a given language model, control over perplexity also gives control over repetitions. Experiments with human raters for fluency, coherence, and quality further verify our findings.
In natural language processing (NLP), enormous pre-trained models like BERT have become the standard starting point for training on a range of downstream tasks, and similar trends are emerging in other areas of deep learning. In parallel, work on the lottery ticket hypothesis has shown that models for NLP and computer vision contain smaller matching subnetworks capable of training in isolation to full accuracy and transferring to other tasks. In this work, we combine these observations to assess whether such trainable, transferrable subnetworks exist in pre-trained BERT models. For a range of downstream tasks, we indeed find matching subnetworks at 40% to 90% sparsity. We find these subnetworks at (pre-trained) initialization, a deviation from prior NLP research where they emerge only after some amount of training. Subnetworks found on the masked language modeling task (the same task used to pre-train the model) transfer universally; those found on other tasks transfer in a limited fashion if at all. As large-scale pre-training becomes an increasingly central paradigm in deep learning, our results demonstrate that the main lottery ticket observations remain relevant in this context. Codes available at https://github.com/VITA-Group/BERT-Tickets.
This work describes our two approaches for the background linking task of TREC 2020 News Track. The main objective of this task is to recommend a list of relevant articles that the reader should refer to in order to understand the context and gain background information of the query article. Our first approach focuses on building an effective search query by combining weighted keywords extracted from the query document and uses BM25 for retrieval. The second approach leverages the capability of SBERT (Nils Reimers et al.) to learn contextual representations of the query in order to perform semantic search over the corpus. We empirically show that employing a language model benefits our approach in understanding the context as well as the background of the query article. The proposed approaches are evaluated on the TREC 2018 Washington Post dataset and our best model outperforms the TREC median as well as the highest scoring model of 2018 in terms of the nDCG@5 metric. We further propose a diversity measure to evaluate the effectiveness of the various approaches in retrieving a diverse set of documents. This would potentially motivate researchers to work on introducing diversity in their recommended list. We have open sourced our implementation on Github and plan to submit our runs for the background linking task in TREC 2020.
In this paper, we present our approach for sentiment classification on Spanish-English code-mixed social media data in the SemEval-2020 Task 9. We investigate performance of various pre-trained Transformer models by using different fine-tuning strategies. We explore both monolingual and multilingual models with the standard fine-tuning method. Additionally, we propose a custom model that we fine-tune in two steps: once with a language modeling objective, and once with a task-specific objective. Although two-step fine-tuning improves sentiment classification performance over the base model, the large multilingual XLM-RoBERTa model achieves best weighted F1-score with 0.537 on development data and 0.739 on test data. With this score, our team jupitter placed tenth overall in the competition.
This paper describes our submissions to SemEval 2020 Task 11: Detection of Propaganda Techniques in News Articles for each of the two subtasks of Span Identification and Technique Classification. We make use of pre-trained BERT language model enhanced with tagging techniques developed for the task of Named Entity Recognition (NER), to develop a system for identifying propaganda spans in the text. For the second subtask, we incorporate contextual features in a pre-trained RoBERTa model for the classification of propaganda techniques. We were ranked 5th in the propaganda technique classification subtask.
BERT (Bidirectional Encoder Representations from Transformers) and ALBERT (A Lite BERT) are methods for pre-training language models which can later be fine-tuned for a variety of Natural Language Understanding tasks. These methods have been applied to a number of such tasks (mostly in English), achieving results that outperform the state-of-the-art. In this paper, our contribution is twofold. First, we make available our trained BERT and Albert model for Portuguese. Second, we compare our monolingual and the standard multilingual models using experiments in semantic textual similarity, recognizing textual entailment, textual category classification, sentiment analysis, offensive comment detection, and fake news detection, to assess the effectiveness of the generated language representations. The results suggest that both monolingual and multilingual models are able to achieve state-of-the-art and the advantage of training a single language model, if any, is small.
Humans can infer a great deal about the meaning of a word, using the syntax and semantics of surrounding words even if it is their first time reading or hearing it. We can also generalise the learned concept of the word to new tasks. Despite great progress in achieving human-level performance in certain tasks (Silver et al., 2016), learning from one or few examples remains a key challenge in machine learning, and has not thoroughly been explored in Natural Language Processing (NLP). In this work we tackle the problem of oneshot learning for an NLP task by employing ideas from recent developments in machine learning: embeddings, attention mechanisms (softmax) and similarity measures (cosine, Euclidean, Poincare, and Minkowski). We adapt the framework suggested in matching networks (Vinyals et al., 2016), and explore the effectiveness of the aforementioned methods in one, two and three-shot learning problems on the task of predicting missing word explored in (Vinyals et al., 2016) by using the WikiText-2 dataset. Our work contributes in two ways: Our first contribution is that we explore the effectiveness of different distance metrics on k-shot learning, and show that there is no single best distance metric for k-shot learning, which challenges common belief. We found that the performance of a distance metric depends on the number of shots used during training. The second contribution of our work is that we establish a benchmark for one, two, and three-shot learning on a language task with a publicly available dataset that can be used to benchmark against in future research.
While mainstream machine learning methods are known to have limited ability to compositionally generalize, new architectures and techniques continue to be proposed to address this limitation. We investigate state-of-the-art techniques and architectures in order to assess their effectiveness in improving compositional generalization in semantic parsing tasks based on the SCAN and CFQ datasets. We show that masked language model (MLM) pre-training rivals SCAN-inspired architectures on primitive holdout splits. On a more complex compositional task, we show that pre-training leads to significant improvements in performance vs. comparable non-pre-trained models, whereas architectures proposed to encourage compositional generalization on SCAN or in the area of algorithm learning fail to lead to significant improvements. We establish a new state of the art on the CFQ compositional generalization benchmark using MLM pre-training together with an intermediate representation.
Arabic dialect identification is a complex problem for a number of inherent properties of the language itself. In this paper, we present the experiments conducted, and the models developed by our competing team, Mawdoo3 AI, along the way to achieving our winning solution to subtask 1 of the Nuanced Arabic Dialect Identification (NADI) shared task. The dialect identification subtask provides 21,000 country-level labeled tweets covering all 21 Arab countries. An unlabeled corpus of 10M tweets from the same domain is also presented by the competition organizers for optional use. Our winning solution itself came in the form of an ensemble of different training iterations of our pre-trained BERT model, which achieved a micro-averaged F1-score of 26.78% on the subtask at hand. We publicly release the pre-trained language model component of our winning solution under the name of Multi-dialect-Arabic-BERT model, for any interested researcher out there.