Cognate alignment models purport to enable decipherment, but their speed and need for clean data can make them unsuitable for realistic decipherment problems. We seek to draw attention to these shortcomings in the hopes that future work may avoid them, and we outline two techniques which begin to overcome the described problems.
Despite remarkable strides made in the development of entity linking systems in recent years, a comprehensive comparative analysis of these systems using a unified framework is notably absent. This paper addresses this oversight by introducing a new black-box benchmark and conducting a comprehensive evaluation of all state-of-the-art entity linking methods. We use an ablation study to investigate the impact of candidate sets on the performance of entity linking. Our findings uncover exactly how much such entity linking systems depend on candidate sets, and how much this limits the general applicability of each system. We present an alternative approach to candidate sets, demonstrating that leveraging the entire in-domain candidate set can serve as a viable substitute for certain models. We show the trade-off between less restrictive candidate sets, increased inference time and memory footprint for some models.
Multilingual neural translation models exploit cross-lingual transfer to perform zero-shot translation between unseen language pairs. Past efforts to improve cross-lingual transfer have focused on aligning contextual sentence-level representations. This paper introduces three novel contributions to allow exploiting nearest neighbours at the token level during training, including: (i) an efficient, gradient-friendly way to share representations between neighboring tokens; (ii) an attentional semantic layer which extracts latent features from shared embeddings; and (iii) an agreement loss to harmonize predictions across different sentence representations. Experiments on two multilingual datasets demonstrate consistent gains in zero shot translation over strong baselines.
The decoder in simultaneous neural machine translation receives limited information from the source while having to balance the opposing requirements of latency versus translation quality. In this paper, we use an auxiliary target-side language model to augment the training of the decoder model. Under this notion of target adaptive training, generating rare or difficult tokens is rewarded which improves the translation quality while reducing latency. The predictions made by a language model in the decoder are combined with the traditional cross entropy loss which frees up the focus on the source side context. Our experimental results over multiple language pairs show that compared to previous state of the art methods in simultaneous translation, we can use an augmented target side context to improve BLEU scores significantly. We show improvements over the state of the art in the low latency range with lower average lagging values (faster output).
A numeration system encodes abstract numeric quantities as concrete strings of written characters. The numeration systems used by modern scripts tend to be precise and unambiguous, but this was not so for the ancient and partially-deciphered proto-Elamite (PE) script, where written numerals can have up to four distinct readings depending on the system that is used to read them. We consider the task of disambiguating between these readings in order to determine the values of the numeric quantities recorded in this corpus. We algorithmically extract a list of possible readings for each PE numeral notation, and contribute two disambiguation techniques based on structural properties of the original documents and classifiers learned with the bootstrapping algorithm. We also contribute a test set for evaluating disambiguation techniques, as well as a novel approach to cautious rule selection for bootstrapped classifiers. Our analysis confirms existing intuitions about this script and reveals previously-unknown correlations between tablet content and numeral magnitude. This work is crucial to understanding and deciphering PE, as the corpus is heavily accounting-focused and contains many more numeric tokens than tokens of text.
A crucial step in deciphering a text is to identify what set of characters were used to write it. This requires grouping character tokens according to visual and contextual features, which can be challenging for human analysts when the number of tokens or underlying types is large. Prior work has shown that this process can be automated by clustering dense representations of character images, in a task which we call “script clustering”. In this work, we present novel architectures which exploit varying degrees of contextual and visual information to learn representations for use in script clustering. We evaluate on a range of modern and ancient scripts, and find that our models produce representations which are more effective for script recovery than the current state-of-the-art, despite using just ~2% as many parameters. Our analysis fruitfully applies these models to assess hypotheses about the character inventory of the partially-deciphered proto-Elamite script.
Solving substitution ciphers involves mapping sequences of cipher symbols to fluent text in a target language. This has conventionally been formulated as a search problem, to find the decipherment key using a character-level language model to constrain the search space. This work instead frames decipherment as a sequence prediction task, using a Transformer-based causal language model to learn recurrences between characters in a ciphertext. We introduce a novel technique for transcribing arbitrary substitution ciphers into a common recurrence encoding. By leveraging this technique, we (i) create a large synthetic dataset of homophonic ciphers using random keys, and (ii) train a decipherment model that predicts the plaintext sequence given a recurrence-encoded ciphertext. Our method achieves strong results on synthetic 1:1 and homophonic ciphers, and cracks several real historic homophonic ciphers. Our analysis shows that the model learns recurrence relations between cipher symbols and recovers decipherment keys in its self-attention.
Entity linking is a prominent thread of research focused on structured data creation by linking spans of text to an ontology or knowledge source. We revisit the use of structured prediction for entity linking which classifies each individual input token as an entity, and aggregates the token predictions. Our system, called SpEL (Structured prediction for Entity Linking) is a state-of-the-art entity linking system that uses some new ideas to apply structured prediction to the task of entity linking including: two refined fine-tuning steps; a context sensitive prediction aggregation strategy; reduction of the size of the model’s output vocabulary, and; we address a common problem in entity-linking systems where there is a training vs. inference tokenization mismatch. Our experiments show that we can outperform the state-of-the-art on the commonly used AIDA benchmark dataset for entity linking to Wikipedia. Our method is also very compute efficient in terms of number of parameters and speed of inference.
This work describes the first thorough analysis of “header” signs in proto-Elamite, an undeciphered script from 3100-2900 BCE. Headers are a category of signs which have been provisionally identified through painstaking manual analysis of this script by domain experts. We use unsupervised neural and statistical sequence modeling techniques to provide new and independent evidence for the existence of headers, without supervision from domain experts. Having affirmed the existence of headers as a legitimate structural feature, we next arrive at a richer understanding of their possible meaning and purpose by (i) examining which features predict their presence; (ii) identifying correlations between these features and other document properties; and (iii) examining cases where these features predict the presence of a header in texts where domain experts do not expect one (or vice versa). We provide more concrete processes for labeling headers in this corpus and a clearer justification for existing intuitions about document structure in proto-Elamite.
We propose a novel technique that combines alternative subword tokenizations of a single source-target language pair that allows us to leverage multilingual neural translation training methods. These alternate segmentations function like related languages in multilingual translation. Overall this improves translation accuracy for low-resource languages and produces translations that are lexically diverse and morphologically rich. We also introduce a cross-teaching technique which yields further improvements in translation accuracy and cross-lingual transfer between high- and low-resource language pairs. Compared to other strong multilingual baselines, our approach yields average gains of +1.7 BLEU across the four low-resource datasets from the multilingual TED-talks dataset. Our technique does not require additional training data and is a drop-in improvement for any existing neural translation system.
We propose a novel data-augmentation technique for neural machine translation based on ROT-k ciphertexts. ROT-k is a simple letter substitution cipher that replaces a letter in the plaintext with the kth letter after it in the alphabet. We first generate multiple ROT-k ciphertexts using different values of k for the plaintext which is the source side of the parallel data. We then leverage this enciphered training data along with the original parallel data via multi-source training to improve neural machine translation. Our method, CipherDAug, uses a co-regularization-inspired training procedure, requires no external data sources other than the original training data, and uses a standard Transformer to outperform strong data augmentation techniques on several datasets by a significant margin. This technique combines easily with existing approaches to data augmentation, and yields particularly strong results in low-resource settings.
In simultaneous machine translation, finding an agent with the optimal action sequence of reads and writes that maintain a high level of translation quality while minimizing the average lag in producing target tokens remains an extremely challenging problem. We propose a novel supervised learning approach for training an agent that can detect the minimum number of reads required for generating each target token by comparing simultaneous translations against full-sentence translations during training to generate oracle action sequences. These oracle sequences can then be used to train a supervised model for action generation at inference time. Our approach provides an alternative to current heuristic methods in simultaneous translation by introducing a new training objective, which is easier to train than previous attempts at training the agent using reinforcement learning techniques for this task. Our experimental results show that our novel training method for action generation produces much higher quality translations while minimizing the average lag in simultaneous translation.
Adding linguistic information (syntax or semantics) to neural machine translation (NMT) have mostly focused on using point estimates from pre-trained models. Directly using the capacity of massive pre-trained contextual word embedding models such as BERT(Devlin et al., 2019) has been marginally useful in NMT because effective fine-tuning is difficult to obtain for NMT without making training brittle and unreliable. We augment NMT by extracting dense fine-tuned vector-based linguistic information from BERT instead of using point estimates. Experimental results show that our method of incorporating linguistic information helps NMT to generalize better in a variety of training contexts and is no more difficult to train than conventional Transformer-based NMT.
While the attention heatmaps produced by neural machine translation (NMT) models seem insightful, there is little evidence that they reflect a model’s true internal reasoning. We provide a measure of faithfulness for NMT based on a variety of stress tests where attention weights which are crucial for prediction are perturbed and the model should alter its predictions if the learned weights are a faithful explanation of the predictions. We show that our proposed faithfulness measure for NMT models can be improved using a novel differentiable objective that rewards faithful behaviour by the model through probability divergence. Our experimental results on multiple language pairs show that our objective function is effective in increasing faithfulness and can lead to a useful analysis of NMT model behaviour and more trustworthy attention heatmaps. Our proposed objective improves faithfulness without reducing the translation quality and has a useful regularization effect on the NMT model and can even improve translation quality in some cases.
Directly translating from speech to text using an end-to-end approach is still challenging for many language pairs due to insufficient data. Although pretraining the encoder parameters using the Automatic Speech Recognition (ASR) task improves the results in low resource settings, attempting to use pretrained parameters from the Neural Machine Translation (NMT) task has been largely unsuccessful in previous works. In this paper, we will show that by using an adversarial regularizer, we can bring the encoder representations of the ASR and NMT tasks closer even though they are in different modalities, and how this helps us effectively use a pretrained NMT decoder for speech translation.
Can we trust that the attention heatmaps produced by a neural machine translation (NMT) model reflect its true internal reasoning? We isolate and examine in detail the notion of faithfulness in NMT models. We provide a measure of faithfulness for NMT based on a variety of stress tests where model parameters are perturbed and measuring faithfulness based on how often the model output changes. We show that our proposed faithfulness measure for NMT models can be improved using a novel differentiable objective that rewards faithful behaviour by the model through probability divergence. Our experimental results on multiple language pairs show that our objective function is effective in increasing faithfulness and can lead to a useful analysis of NMT model behaviour and more trustworthy attention heatmaps. Our proposed objective improves faithfulness without reducing the translation quality and it also seems to have a useful regularization effect on the NMT model and can even improve translation quality in some cases.
We describe a first attempt at using techniques from computational linguistics to analyze the undeciphered proto-Elamite script. Using hierarchical clustering, n-gram frequencies, and LDA topic models, we both replicate results obtained by manual decipherment and reveal previously-unobserved relationships between signs. This demonstrates the utility of these techniques as an aid to manual decipherment.
Supertagging is a sequence prediction task where each word is assigned a piece of complex syntactic structure called a supertag. We provide a novel approach to multi-task learning for Tree Adjoining Grammar (TAG) supertagging by deconstructing these complex supertags in order to define a set of related but auxiliary sequence prediction tasks. Our multi-task prediction framework is trained over the exactly same training data used to train the original supertagger where each auxiliary task provides an alternative view on the original prediction task. Our experimental results show that our multi-task approach significantly improves TAG supertagging with a new state-of-the-art accuracy score of 91.39% on the Penn Treebank supertagging dataset.
Attention models have become a crucial component in neural machine translation (NMT). They are often implicitly or explicitly used to justify the model’s decision in generating a specific token but it has not yet been rigorously established to what extent attention is a reliable source of information in NMT. To evaluate the explanatory power of attention for NMT, we examine the possibility of yielding the same prediction but with counterfactual attention models that modify crucial aspects of the trained attention model. Using these counterfactual attention mechanisms we assess the extent to which they still preserve the generation of function and content words in the translation process. Compared to a state of the art attention model, our counterfactual attention models produce 68% of function words and 21% of content words in our German-English dataset. Our experiments demonstrate that attention models by themselves cannot reliably explain the decisions made by a NMT model.
Automated filters are commonly used by online services to stop users from sending age-inappropriate, bullying messages, or asking others to expose personal information. Previous work has focused on rules or classifiers to detect and filter offensive messages, but these are vulnerable to cleverly disguised plaintext and unseen expressions especially in an adversarial setting where the users can repeatedly try to bypass the filter. In this paper, we model the disguised messages as if they are produced by encrypting the original message using an invented cipher. We apply automatic decipherment techniques to decode the disguised malicious text, which can be then filtered using rules or classifiers. We provide experimental results on three different datasets and show that decipherment is an effective tool for this task.
Rapidly expanding volume of publications in the biomedical domain makes it increasingly difficult for a timely evaluation of the latest literature. That, along with a push for automated evaluation of clinical reports, present opportunities for effective natural language processing methods. In this study we target the problem of named entity recognition, where texts are processed to annotate terms that are relevant for biomedical studies. Terms of interest in the domain include gene and protein names, and cell lines and types. Here we report on a pipeline built on Embeddings from Language Models (ELMo) and a deep learning package for natural language processing (AllenNLP). We trained context-aware token embeddings on a dataset of biomedical papers using ELMo, and incorporated these embeddings in the LSTM-CRF model used by AllenNLP for named entity recognition. We show these representations improve named entity recognition for different types of biomedical named entities. We also achieve a new state of the art in gene mention detection on the BioCreative II gene mention shared task.
The addition of syntax-aware decoding in Neural Machine Translation (NMT) systems requires an effective tree-structured neural network, a syntax-aware attention model and a language generation model that is sensitive to sentence structure. Recent approaches resort to sequential decoding by adding additional neural network units to capture bottom-up structural information, or serialising structured data into sequence. We exploit a top-down tree-structured model called DRNN (Doubly-Recurrent Neural Networks) first proposed by Alvarez-Melis and Jaakola (2017) to create an NMT model called Seq2DRNN that combines a sequential encoder with tree-structured decoding augmented with a syntax-aware attention model. Unlike previous approaches to syntax-based NMT which use dependency parsing models our method uses constituency parsing which we argue provides useful information for translation. In addition, we use the syntactic structure of the sentence to add new connections to the tree-structured decoder neural network (Seq2DRNN+SynC). We compare our NMT model with sequential and state of the art syntax-based NMT models and show that our model produces more fluent translations with better reordering. Since our model is capable of doing translation and constituency parsing at the same time we also compare our parsing accuracy against other neural parsing models.
Decipherment of homophonic substitution ciphers using language models is a well-studied task in NLP. Previous work in this topic scores short local spans of possible plaintext decipherments using n-gram language models. The most widely used technique is the use of beam search with n-gram language models proposed by Nuhn et al.(2013). We propose a beam search algorithm that scores the entire candidate plaintext at each step of the decipherment using a neural language model. We augment beam search with a novel rest cost estimation that exploits the prediction power of a neural language model. We compare against the state of the art n-gram based methods on many different decipherment tasks. On challenging ciphers such as the Beale cipher we provide significantly better error rates with much smaller beam sizes.
Simultaneous speech translation aims to maintain translation quality while minimizing the delay between reading input and incrementally producing the output. We propose a new general-purpose prediction action which predicts future words in the input to improve quality and minimize delay in simultaneous translation. We train this agent using reinforcement learning with a novel reward function. Our agent with prediction has better translation quality and less delay compared to an agent-based simultaneous translation system without prediction.
We show that an epsilon-free, chain-free synchronous context-free grammar (SCFG) can be converted into a weakly equivalent synchronous tree-adjoining grammar (STAG) which is prefix lexicalized. This transformation at most doubles the grammar’s rank and cubes its size, but we show that in practice the size increase is only quadratic. Our results extend Greibach normal form from CFGs to SCFGs and prove new formal properties about SCFG, a formalism with many applications in natural language processing.
Current word alignment models do not distinguish between different types of alignment links. In this paper, we provide a new probabilistic model for word alignment where word alignments are associated with linguistically motivated alignment types. We propose a novel task of joint prediction of word alignment and alignment types and propose novel semi-supervised learning algorithms for this task. We also solve a sub-task of predicting the alignment type given an aligned word pair. In our experimental results, the generative models we introduce to model alignment types significantly outperform the models without alignment types.
Phrase-based and hierarchical phrase-based (Hiero) translation models differ radically in the way reordering is modeled. Lexicalized reordering models play an important role in phrase-based MT and such models have been added to CKY-based decoders for Hiero. Watanabe et al. (2006) proposed a promising decoding algorithm for Hiero (LR-Hiero) that visits input spans in arbitrary order and produces the translation in left to right (LR) order which leads to far fewer language model calls and leads to a considerable speedup in decoding. We introduce a novel shift-reduce algorithm to LR-Hiero to decode with our lexicalized reordering model (LRM) and show that it improves translation quality for Czech-English, Chinese-English and German-English.
Left-to-right (LR) decoding Watanabe et al. (2006) is a promising decoding algorithm for hierarchical phrase-based translation (Hiero) that visits input spans in arbitrary order producing the output translation in left to right order. This leads to far fewer language model calls. But the constrained SCFG grammar used in LR-Hiero (GNF) with at most two non-terminals is unable to account for some complex phrasal reordering. Allowing more non-terminals in the rules results in a more expressive grammar. LR-decoding can be used to decode with SCFGs with more than two non-terminals, but the CKY decoders used for Hiero systems cannot deal with such expressive grammars due to a blowup in computational complexity. In this paper we present a dynamic programming algorithm for GNF rule extraction which efficiently extracts sentence level SCFG rule sets with an arbitrary number of non-terminals. We analyze the performance of the obtained grammar for statistical machine translation on three language pairs.
The typical training of a hierarchical phrase-based machine translation involves a pipeline of multiple steps where mistakes in early steps of the pipeline are propagated without any scope for rectifying them. Additionally the alignments are trained independent of and without being informed of the end goal and hence are not optimized for translation. We introduce a novel Bayesian iterative-cascade framework for training Hiero-style model that learns the alignments together with the synchronous translation grammar in an iterative setting. Our framework addresses the above mentioned issues and provides an elegant and principled alternative to the existing training pipeline. Based on the validation experiments involving two language pairs, our proposed iterative-cascade framework shows consistent gains over the traditional training pipeline for hierarchical translation.
This paper conducts a comprehensive study on the use of triangulation for four very low-resource languages: Mawukakan and Maninkakan, Haitian Kreyol and Malagasy. To the best of our knowledge, ours is the first effective translation system for the first two of these languages. We improve translation quality by adding data using pivot languages and exper- imentally compare previously proposed triangulation design options. Furthermore, since the low-resource language pair and pivot language pair data typically come from very different domains, we use insights from domain adaptation to tune the weighted mixture of direct and pivot based phrase pairs to improve translation quality.
This paper introduces two novel approaches for extracting compact grammars for hierarchical phrase-based translation. The first is a combinatorial optimization approach and the second is a Bayesian model over Hiero grammars using Variational Bayes for inference. In contrast to the conventional Hiero (Chiang, 2007) rule extraction algorithm , our methods extract compact models reducing model size by 17.8% to 57.6% without impacting translation quality across several language pairs. The Bayesian model is particularly effective for resource-poor languages with evidence from Korean-English translation. To our knowledge, this is the first alternative to Hiero-style rule extraction that finds a more compact synchronous grammar without hurting translation performance.