Fine-tuning pre-trained large language models (LLMs) on a diverse array of tasks has become a common approach for building models that can solve various natural language processing (NLP) tasks. However, where and to what extent these models retain task-specific knowledge remains largely unexplored. This study investigates the task-specific information encoded in pre-trained LLMs and the effects of instruction tuning on their representations across a diverse set of over 60 NLP tasks. We use a set of matrix analysis tools to examine the differences between the way pre-trained and instruction-tuned LLMs store task-specific information. Our findings reveal that while some tasks are already encoded within the pre-trained LLMs, others greatly benefit from instruction tuning. Additionally, we pinpointed the layers in which the model transitions from high-level general representations to more task-oriented representations. This finding extends our understanding of the governing mechanisms of LLMs and facilitates future research in the fields of parameter-efficient transfer learning and multi-task learning. Our code is available at: https://github.com/zsquaredz/layer_by_layer/

Understanding the inner workings of large language models (LLMs) is crucial for advancing their theoretical foundations and real-world applications. While the attention mechanism and multi-layer perceptrons (MLPs) have been studied independently, their interactions remain largely unexplored. This study investigates how attention heads and next-token neurons interact in LLMs to predict new words. We propose a methodology to identify next-token neurons, find prompts that highly activate them, and determine the upstream attention heads responsible. We then generate and evaluate explanations for the activity of these attention heads in an automated manner. Our findings reveal that some attention heads recognize specific contexts relevant to predicting a token and activate a downstream token-predicting neuron accordingly. This mechanism provides a deeper understanding of how attention heads work with MLP neurons to perform next-token prediction. Our approach offers a foundation for further research into the intricate workings of LLMs and their impact on text generation and understanding.

Knowledge-to-text generators often struggle to faithfully generate descriptions for the input facts: they may produce hallucinations that contradict the input, or describe facts not present in the input. To reduce hallucinations, we propose a decoding-only method, TWEAK (Think While Effectively Articulating Knowledge), which can be integrated with any generator without retraining. TWEAK treats the generated sequences at each decoding step and its future sequences as hypotheses, and ranks each generation candidate based on the extent to which their hypotheses are supported by the input facts using a Hypothesis Verification Model (HVM). We first demonstrate the effectiveness of TWEAK by using a Natural Language Inference (NLI) model as the HVM and report improved faithfulness with a minimal impact on the quality. We then replace the NLI model with a task-specific HVM trained with a first-of-a-kind dataset, FATE (Fact-Aligned Textual Entailment), which pairs input facts with their original and perturbed descriptions. We test TWEAK with two generators, and the best TWEAK variants improve on average for the two models by 2.24/7.17 points in faithfulness (FactKB) in in/out-of-distribution evaluations, respectively, and with only a 0.14/0.32-point decline in quality (BERTScore).

Although large language models (LLMs) exhibit remarkable capacity to leverage in-context demonstrations, it is still unclear to what extent they can learn new facts or concept definitions via prompts. To address this question, we examine the capacity of instruction-tuned LLMs to follow in-context concept annotation guidelines for zero-shot sentence labeling tasks. We design guidelines that present different types of factual and counterfactual concept definitions, which are used as prompts for zero-shot sentence classification tasks. Our results show that although concept definitions consistently help in task performance, only the larger models (with 70B parameters or more) have limited ability to work under counterfactual contexts. Importantly, only proprietary models such as GPT-3.5 can recognize nonsensical guidelines, which we hypothesize is due to more sophisticated alignment methods. Finally, we find that Falcon-180B-chat is outperformed by Llama-2-70B-chat is most cases, which indicates that increasing model scale does not guarantee better adherence to guidelines. Altogether, our simple evaluation method reveals significant gaps in concept understanding between the most capable open-source language models and the leading proprietary APIs.

Advances in model editing through neuron pruning hold promise for removing undesirable concepts from large language models. However, it remains unclear whether models have the capacity to reacquire pruned concepts after editing. To investigate this, we evaluate concept relearning in models by tracking concept saliency and similarity in pruned neurons during retraining for named entity recognition tasks. Our findings reveal that models can quickly regain performance post-pruning by relocating advanced concepts to earlier layers and reallocating pruned concepts to primed neurons with similar semantics. This suggests that models exhibit polysemantic capacities and can blend old and new concepts in individual neurons. While neuron pruning provides interpretability into model concepts, our results highlight the challenges of permanent concept removal for improved model *safety*. Monitoring concept reemergence and developing techniques to mitigate relearning of unsafe concepts will be important directions for more robust model editing. Overall, our work strongly demonstrates the resilience and fluidity of concept representations in LLMs post concept removal.

In this work, we investigate the controllability of large language models (LLMs) on scientific summarization tasks. We identify key stylistic and content coverage factors that characterize different types of summaries such as paper reviews, abstracts, and lay summaries. By controlling stylistic features, we find that non-fine-tuned LLMs outperform humans in the MuP review generation task, both in terms of similarity to reference summaries and human preferences. Also, we show that we can improve the controllability of LLMs with keyword-based classifier-free guidance (CFG) while achieving lexical overlap comparable to strong fine-tuned baselines on arXiv and PubMed. However, our results also indicate that LLMs cannot consistently generate long summaries with more than 8 sentences. Furthermore, these models exhibit limited capacity to produce highly abstractive lay summaries. Although LLMs demonstrate strong generic summarization competency, sophisticated content control without costly fine-tuning remains an open problem for domain-specific applications.

We introduce a dataset comprising commercial machine translations, gathered weekly over six years across 12 translation directions. Since human A/B testing is commonly used, we assume commercial systems improve over time, which enables us to evaluate machine translation (MT) metrics based on their preference for more recent translations. Our study not only confirms several prior findings, such as the advantage of neural metrics over non-neural ones, but also explores the debated issue of how MT quality affects metric reliability—an investigation that smaller datasets in previous research could not sufficiently explore. Overall, our research demonstrates the dataset’s value as a testbed for metric evaluation. We release our code.

The diffusion of financial news into market prices is a complex process, making it challenging to evaluate the connections between news events and market movements. This paper introduces FININ (Financial Interconnected News Influence Network), a novel market prediction model that captures not only the links between news and prices but also the interactions among news items themselves. FININ effectively integrates multi-modal information from both market data and news articles. We conduct extensive experiments on two datasets, encompassing the S&P 500 and NASDAQ 100 indices over a 15-year period and over 2.7 million news articles. The results demonstrate FININ’s effectiveness, outperforming advanced market prediction models with an improvement of 0.429 and 0.341 in the daily Sharpe ratio for the two markets respectively. Moreover, our results reveal insights into the financial news, including the delayed market pricing of news, the long memory effect of news, and the limitations of financial sentiment analysis in fully extracting predictive power from news data.

Are Large Language Models (LLMs) temporally grounded? Since LLMs cannot perceive and interact with the environment, it is impossible to answer this question directly. Instead, we provide LLMs with textual narratives and probe them with respect to their common-sense knowledge of the structure and duration of events, their ability to order events along a timeline, and self-consistency within their temporal model (e.g., temporal relations such as after and before are mutually exclusive for any pair of events). We evaluate state-of-the-art LLMs (such as LLaMA 2 and GPT-4) on three tasks reflecting these abilities. Generally, we find that LLMs lag significantly behind both human performance as well as small-scale, specialised LMs. In-context learning, instruction tuning, and chain-of-thought prompting reduce this gap only to a limited degree. Crucially, LLMs struggle the most with self-consistency, displaying incoherent behaviour in at least 27.23% of their predictions. Contrary to expectations, we also find that scaling the model size does not guarantee positive gains in performance. To explain these results, we study the sources from which LLMs may gather temporal information: we find that sentence ordering in unlabelled texts, available during pre-training, is only weakly correlated with event ordering. Moreover, public instruction tuning mixtures contain few temporal tasks. Hence, we conclude that current LLMs lack a consistent temporal model of textual narratives.

Large language models (LLMs) often struggle with complex logical reasoning due to logical inconsistencies and the inherent difficulty ofsuch reasoning. We use Lean, a theorem proving framework, to address these challenges. By formalizing logical reasoning problems intotheorems within Lean, we can solve them by proving or disproving the corresponding theorems. This method reduces the risk of logical inconsistencies with the help of Lean’s symbolic solver. It also enhances our ability to treat complex reasoning tasks using Lean’s extensive library of theorem proofs. Our method achieves state-of-the-art performance on the FOLIO dataset and achieves performance near this level on ProofWriter. Notably, these results were accomplished by fine-tuning on fewer than 100 in-domain samples for each dataset

Space program agencies execute complex satellite operations that need to be supported by the technical knowledge contained in their extensive information systems. Knowledge Base (KB) databases are an effective way of storing and accessing such information to scale. In this work we present a system, developed for the European Space Agency, that can answer complex natural language queries, to support engineers in accessing the information contained in a KB that models the orbital space debris environment. Our system is based on a pipeline which first generates a program sketch from a natural language question, then specializes the sketch into a concrete query program with mentions of entities, attributes and relations, and finally executes the program against the database. This pipeline decomposition approach enables us to train the system by leveraging out-of-domain data and semi-synthetic data generated by GPT-3, thus reducing overfitting and shortcut learning even with limited amount of in-domain training data.

Large Language Models (LLMs) have successfully been applied to code generation tasks, raising the question of how well these models understand programming. Typical programming languages have invariances and equivariances in their semantics that human programmers intuitively understand and exploit, such as the (near) invariance to the renaming of identifiers. We show that LLMs not only fail to properly generate correct Python code when default function names are swapped, but some of them even become more confident in their incorrect predictions as the model size increases, an instance of the recently discovered phenomenon of Inverse Scaling, which runs contrary to the commonly observed trend of increasing prediction quality with increasing model size. Our findings indicate that, despite their astonishing typical-case performance, LLMs still lack a deep, abstract understanding of the content they manipulate, making them unsuitable for tasks that statistically deviate from their training data, and that mere scaling is not enough to achieve such capability.

This paper introduces PMIndiaSum, a multilingual and massively parallel summarization corpus focused on languages in India. Our corpus provides a training and testing ground for four language families, 14 languages, and the largest to date with 196 language pairs. We detail our construction workflow including data acquisition, processing, and quality assurance. Furthermore, we publish benchmarks for monolingual, cross-lingual, and multilingual summarization by fine-tuning, prompting, as well as translate-and-summarize. Experimental results confirm the crucial role of our data in aiding summarization between Indian languages. Our dataset is publicly available and can be freely modified and re-distributed.

We present an analysis tool based on joint matrix factorization for comparing latent representations of multilingual and monolingual models. An alternative to probing, this tool allows us to analyze multiple sets of representations in a joint manner. Using this tool, we study to what extent and how morphosyntactic features are reflected in the representations learned by multilingual pre-trained models. We conduct a large-scale empirical study of over 33 languages and 17 morphosyntactic categories. Our findings demonstrate variations in the encoding of morphosyntactic information across upper and lower layers, with category-specific differences influenced by language properties. Hierarchical clustering of the factorization outputs yields a tree structure that is related to phylogenetic trees manually crafted by linguists. Moreover, we find the factorization outputs exhibit strong associations with performance observed across different cross-lingual tasks. We release our code to facilitate future research.

Hallucinations pose a significant challenge to the reliability of neural models for abstractive summarisation. While automatically generated summaries may be fluent, they often lack faithfulness to the original document. This issue becomes even more pronounced in low-resource languages, where summarisation requires cross-lingual transfer. With the existing faithful metrics focusing on English, even measuring the extent of this phenomenon in cross-lingual settings is hard. To address this, we first develop a novel metric, mFACT, evaluating the faithfulness of non-English summaries, leveraging translation-based transfer from multiple English faithfulness metrics. Through extensive experiments in multiple languages, we demonstrate that mFACT is best suited to detect hallucinations compared to alternative metrics. With mFACT, we assess a broad range of multilingual large language models, and find that they all tend to hallucinate often in languages different from English. We then propose a simple but effective method to reduce hallucinations in cross-lingual transfer, which weighs the loss of each training example by its faithfulness score. This method drastically increases both performance and faithfulness according to both automatic and human evaluation when compared to strong baselines for cross-lingual transfer such as MAD-X. Our code and dataset are available at https://github.com/yfqiu-nlp/mfact-summ.

We present the Granular AMR Parsing Evaluation Suite (GrAPES), a challenge set for Abstract Meaning Representation (AMR) parsing with accompanying evaluation metrics. AMR parsers now obtain high scores on the standard AMR evaluation metric Smatch, close to or even above reported inter-annotator agreement. But that does not mean that AMR parsing is solved; in fact, human evaluation in previous work indicates that current parsers still quite frequently make errors on node labels or graph structure that substantially distort sentence meaning. Here, we provide an evaluation suite that tests AMR parsers on a range of phenomena of practical, technical, and linguistic interest. Our 36 categories range from seen and unseen labels, to structural generalization, to coreference. GrAPES reveals in depth the abilities and shortcomings of current AMR parsers.

We describe a simple and effective method (Spectral Attribute removaL; SAL) to remove private or guarded information from neural representations. Our method uses matrix decomposition to project the input representations into directions with reduced covariance with the guarded information rather than maximal covariance as factorization methods normally use. We begin with linear information removal and proceed to generalize our algorithm to the case of nonlinear information removal using kernels. Our experiments demonstrate that our algorithm retains better main task performance after removing the guarded information compared to previous work. In addition, our experiments demonstrate that we need a relatively small amount of guarded attribute data to remove information about these attributes, which lowers the exposure to sensitive data and is more suitable for low-resource scenarios.

We introduce a task consisting in matching a proof to a given mathematical statement. The task fits well within current research on Mathematical Information Retrieval and, more generally, mathematical article analysis (Mathematical Sciences, 2014). We present a dataset for the task (the MATcH dataset) consisting of over 180k statement-proof pairs extracted from modern mathematical research articles. We find this dataset highly representative of our task, as it consists of relatively new findings useful to mathematicians. We propose a bilinear similarity model and two decoding methods to match statements to proofs effectively. While the first decoding method matches a proof to a statement without being aware of other statements or proofs, the second method treats the task as a global matching problem. Through a symbol replacement procedure, we analyze the “insights” that pre-trained language models have in such mathematical article analysis and show that while these models perform well on this task with the best performing mean reciprocal rank of 73.7, they follow a relatively shallow symbolic analysis and matching to achieve that performance.

We present the Assignment-Maximization Spectral Attribute removaL (AMSAL) algorithm, which erases information from neural representations when the information to be erased is implicit rather than directly being aligned to each input example. Our algorithm works by alternating between two steps. In one, it finds an assignment of the input representations to the information to be erased, and in the other, it creates projections of both the input representations and the information to be erased into a joint latent space. We test our algorithm on an extensive array of datasets, including a Twitter dataset with multiple guarded attributes, the BiasBios dataset, and the BiasBench benchmark. The latter benchmark includes four datasets with various types of protected attributes. Our results demonstrate that bias can often be removed in our setup. We also discuss the limitations of our approach when there is a strong entanglement between the main task and the information to be erased.1

We investigate how different domains are encoded in modern neural network architectures. We analyze the relationship between natural language domains, model size, and the amount of training data used. The primary analysis tool we develop is based on subpopulation analysis with Singular Vector Canonical Correlation Analysis (SVCCA), which we apply to Transformer-based language models (LMs). We compare the latent representations of such a language model at its different layers from a pair of models: a model trained on multiple domains (an experimental model) and a model trained on a single domain (a control model). Through our method, we find that increasing the model capacity impacts how domain information is stored in upper and lower layers differently. In addition, we show that larger experimental models simultaneously embed domain-specific information as if they were conjoined control models. These findings are confirmed qualitatively, demonstrating the validity of our method.

We propose a sentence-incremental neural coreference resolution system which incrementally builds clusters after marking mention boundaries in a shift-reduce method. The system is aimed at bridging two recent approaches at coreference resolution: (1) state-of-the-art non-incremental models that incur quadratic complexity in document length with high computational cost, and (2) memory network-based models which operate incrementally but do not generalize beyond pronouns. For comparison, we simulate an incremental setting by constraining non-incremental systems to form partial coreference chains before observing new sentences. In this setting, our system outperforms comparable state-of-the-art methods by 2 F1 on OntoNotes and 6.8 F1 on the CODI-CRAC 2021 corpus. In a conventional coreference setup, our system achieves 76.3 F1 on OntoNotes and 45.5 F1 on CODI-CRAC 2021, which is comparable to state-of-the-art baselines. We also analyze variations of our system and show that the degree of incrementality in the encoder has a surprisingly large effect on the resulting performance.

Sequential abstractive neural summarizers often do not use the underlying structure in the input article or dependencies between the input sentences. This structure is essential to integrate and consolidate information from different parts of the text. To address this shortcoming, we propose a hierarchy-aware graph neural network (HierGNN) which captures such dependencies through three main steps: 1) learning a hierarchical document structure through a latent structure tree learned by a sparse matrix-tree computation; 2) propagating sentence information over this structure using a novel message-passing node propagation mechanism to identify salient information; 3) using graph-level attention to concentrate the decoder on salient information. Experiments confirm HierGNN improves strong sequence models such as BART, with a 0.55 and 0.75 margin in average ROUGE-1/2/L for CNN/DM and XSum. Further human evaluation demonstrates that summaries produced by our model are more relevant and less redundant than the baselines, into which HierGNN is incorporated. We also find HierGNN synthesizes summaries by fusing multiple source sentences more, rather than compressing a single source sentence, and that it processes long inputs more effectively.

We argue that disentangling content selection from the budget used to cover salient content improves the performance and applicability of abstractive summarizers. Our method, FactorSum, does this disentanglement by factorizing summarization into two steps through an energy function: (1) generation of abstractive summary views covering salient information in subsets of the input document (document views); (2) combination of these views into a final summary, following a budget and content guidance. This guidance may come from different sources, including from an advisor model such as BART or BigBird, or in oracle mode – from the reference. This factorization achieves significantly higher ROUGE scores on multiple benchmarks for long document summarization, namely PubMed, arXiv, and GovReport. Most notably, our model is effective for domain adaptation. When trained only on PubMed samples, it achieves a 46.29 ROUGE-1 score on arXiv, outperforming PEGASUS trained in domain by a large margin. Our experimental results indicate that the performance gains are due to more flexible budget adaptation and processing of shorter contexts provided by partial document views.

We introduce a method for unsupervised parsing that relies on bootstrapping classifiers to identify if a node dominates a specific span in a sentence. There are two types of classifiers, an inside classifier that acts on a span, and an outside classifier that acts on everything outside of a given span. Through self-training and co-training with the two classifiers, we show that the interplay between them helps improve the accuracy of both, and as a result, effectively parse. A seed bootstrapping technique prepares the data to train these classifiers. Our analyses further validate that such an approach in conjunction with weak supervision using prior branching knowledge of a known language (left/right-branching) and minimal heuristics injects strong inductive bias into the parser, achieving 63.1 F₁ on the English (PTB) test set. In addition, we show the effectiveness of our architecture by evaluating on treebanks for Chinese (CTB) and Japanese (KTB) and achieve new state-of-the-art results.

We propose neural models to generate text from formal meaning representations based on Discourse Representation Structures (DRSs). DRSs are document-level representations which encode rich semantic detail pertaining to rhetorical relations, presupposition, and co-reference within and across sentences. We formalize the task of neural DRS-to-text generation and provide modeling solutions for the problems of condition ordering and variable naming which render generation from DRSs non-trivial. Our generator relies on a novel sibling treeLSTM model which is able to accurately represent DRS structures and is more generally suited to trees with wide branches. We achieve competitive performance (59.48 BLEU) on the GMB benchmark against several strong baselines.

We consider the task of crosslingual semantic parsing in the style of Discourse Representation Theory (DRT) where knowledge from annotated corpora in a resource-rich language is transferred via bitext to guide learning in other languages. We introduce 𝕌niversal Discourse Representation Theory (𝕌DRT), a variant of DRT that explicitly anchors semantic representations to tokens in the linguistic input. We develop a semantic parsing framework based on the Transformer architecture and utilize it to obtain semantic resources in multiple languages following two learning schemes. The many-to-one approach translates non-English text to English, and then runs a relatively accurate English parser on the translated text, while the one-to-many approach translates gold standard English to non-English text and trains multiple parsers (one per language) on the translations. Experimental results on the Parallel Meaning Bank show that our proposal outperforms strong baselines by a wide margin and can be used to construct (silver-standard) meaning banks for 99 languages.

An open-domain knowledge graph (KG) has entities as nodes and natural language relations as edges, and is constructed by extracting (subject, relation, object) triples from text. The task of open-domain link prediction is to infer missing relations in the KG. Previous work has used standard link prediction for the task. Since triples are extracted from text, we can ground them in the larger textual context in which they were originally found. However, standard link prediction methods only rely on the KG structure and ignore the textual context that each triple was extracted from. In this paper, we introduce the new task of open-domain contextual link prediction which has access to both the textual context and the KG structure to perform link prediction. We build a dataset for the task and propose a model for it. Our experiments show that context is crucial in predicting missing relations. We also demonstrate the utility of contextual link prediction in discovering context-independent entailments between relations, in the form of entailment graphs (EG), in which the nodes are the relations. The reverse holds too: context-independent EGs assist in predicting relations in context.

Abstract Meaning Representations (AMR) are a broad-coverage semantic formalism which represents sentence meaning as a directed acyclic graph. To train most AMR parsers, one needs to segment the graph into subgraphs and align each such subgraph to a word in a sentence; this is normally done at preprocessing, relying on hand-crafted rules. In contrast, we treat both alignment and segmentation as latent variables in our model and induce them as part of end-to-end training. As marginalizing over the structured latent variables is infeasible, we use the variational autoencoding framework. To ensure end-to-end differentiable optimization, we introduce a differentiable relaxation of the segmentation and alignment problems. We observe that inducing segmentation yields substantial gains over using a ‘greedy’ segmentation heuristic. The performance of our method also approaches that of a model that relies on the segmentation rules of Lyu and Titov (2018), which were hand-crafted to handle individual AMR constructions.

We describe two approaches to single-root dependency parsing that yield significant speed ups in such parsing. One approach has been previously used in dependency parsers in practice, but remains undocumented in the parsing literature, and is considered a heuristic. We show that this approach actually finds the optimal dependency tree. The second approach relies on simple reweighting of the inference graph being input to the dependency parser and has an optimal running time. Here, we again show that this approach is fully correct and identifies the highest-scoring parse tree. Our experiments demonstrate a manyfold speed up compared to a previous graph-based state-of-the-art parser without any loss in accuracy or optimality.

Approaching named entities transliteration as a Neural Machine Translation (NMT) problem is common practice. While many have applied various NMT techniques to enhance machine transliteration models, few focus on the linguistic features particular to the relevant languages. In this paper, we investigate the effect of incorporating phonetic features for English-to-Chinese transliteration under the multi-task learning (MTL) setting—where we define a phonetic auxiliary task aimed to improve the generalization performance of the main transliteration task. In addition to our system, we also release a new English-to-Chinese dataset and propose a novel evaluation metric which considers multiple possible transliterations given a source name. Our results show that the multi-task model achieves similar performance as the previous state of the art with a model of a much smaller size.

Deep reinforcement learning is a promising approach to training a dialog manager, but current methods struggle with the large state and action spaces of multi-domain dialog systems. Building upon Deep Q-learning from Demonstrations (DQfD), an algorithm that scores highly in difficult Atari games, we leverage dialog data to guide the agent to successfully respond to a user’s requests. We make progressively fewer assumptions about the data needed, using labeled, reduced-labeled, and even unlabeled data to train expert demonstrators. We introduce Reinforced Fine-tune Learning, an extension to DQfD, enabling us to overcome the domain gap between the datasets and the environment. Experiments in a challenging multi-domain dialog system framework validate our approaches, and get high success rates even when trained on out-of-domain data.

Discourse representation structures (DRSs) are scoped semantic representations for texts of arbitrary length. Evaluating the accuracy of predicted DRSs plays a key role in developing semantic parsers and improving their performance. DRSs are typically visualized as boxes which are not straightforward to process automatically. Counter transforms DRSs to clauses and measures clause overlap by searching for variable mappings between two DRSs. However, this metric is computationally costly (with respect to memory and CPU time) and does not scale with longer texts. We introduce Dscorer, an efficient new metric which converts box-style DRSs to graphs and then measures the overlap of n-grams. Experiments show that Dscorer computes accuracy scores that are correlated with Counter at a fraction of the time.

Machine reading is an ambitious goal in NLP that subsumes a wide range of text understanding capabilities. Within this broad framework, we address the task of machine reading the time of historical events, compile datasets for the task, and develop a model for tackling it. Given a brief textual description of an event, we show that good performance can be achieved by extracting relevant sentences from Wikipedia, and applying a combination of task-specific and general-purpose feature embeddings for the classification. Furthermore, we establish a link between the historical event ordering task and the event focus time task from the information retrieval literature, showing they also provide a challenging test case for machine reading algorithms.

The goal of homomorphic encryption is to encrypt data such that another party can operate on it without being explicitly exposed to the content of the original data. We introduce an idea for a privacy-preserving transformation on natural language data, inspired by homomorphic encryption. Our primary tool is obfuscation, relying on the properties of natural language. Specifically, a given English text is obfuscated using a neural model that aims to preserve the syntactic relationships of the original sentence so that the obfuscated sentence can be parsed instead of the original one. The model works at the word level, and learns to obfuscate each word separately by changing it into a new word that has a similar syntactic role. The text obfuscated by our model leads to better performance on three syntactic parsers (two dependency and one constituency parsers) in comparison to an upper-bound random substitution baseline. More specifically, the results demonstrate that as more terms are obfuscated (by their part of speech), the substitution upper bound significantly degrades, while the neural model maintains a relatively high performing parser. All of this is done without much sacrifice of privacy compared to the random substitution upper bound. We also further analyze the results, and discover that the substituted words have similar syntactic properties, but different semantic content, compared to the original words.

Semiring parsing is an elegant framework for describing parsers by using semiring weighted logic programs. In this paper we present a generalization of this concept: latent-variable semiring parsing. With our framework, any semiring weighted logic program can be latentified by transforming weights from scalar values of a semiring to rank-n arrays, or tensors, of semiring values, allowing the modelling of latent-variable models within the semiring parsing framework. Semiring is too strong a notion when dealing with tensors, and we have to resort to a weaker structure: a partial semiring. We prove that this generalization preserves all the desired properties of the original semiring framework while strictly increasing its expressiveness.

Abstract Meaning Representation (AMR) parsing aims at converting sentences into AMR representations. These are graphs and not trees because AMR supports reentrancies (nodes with more than one parent). Following previous findings on the importance of reen- trancies for AMR, we empirically find and discuss several linguistic phenomena respon- sible for reentrancies in AMR, some of which have not received attention before. We cate- gorize the types of errors AMR parsers make with respect to reentrancies. Furthermore, we find that correcting these errors provides an in- crease of up to 5% Smatch in parsing perfor- mance and 20% in reentrancy prediction

It is well-known that abstractive summaries are subject to hallucination—including material that is not supported by the original text. While summaries can be made hallucination-free by limiting them to general phrases, such summaries would fail to be very informative. Alternatively, one can try to avoid hallucinations by verifying that any specific entities in the summary appear in the original text in a similar context. This is the approach taken by our system, Herman. The system learns to recognize and verify quantity entities (dates, numbers, sums of money, etc.) in a beam-worth of abstractive summaries produced by state-of-the-art models, in order to up-rank those summaries whose quantity terms are supported by the original text. Experimental results demonstrate that the ROUGE scores of such up-ranked summaries have a higher Precision than summaries that have not been up-ranked, without a comparable loss in Recall, resulting in higher F1. Preliminary human evaluation of up-ranked vs. original summaries shows people’s preference for the former.

AMR-to-text generation is used to transduce Abstract Meaning Representation structures (AMR) into text. A key challenge in this task is to efficiently learn effective graph representations. Previously, Graph Convolution Networks (GCNs) were used to encode input AMRs, however, vanilla GCNs are not able to capture non-local information and additionally, they follow a local (first-order) information aggregation scheme. To account for these issues, larger and deeper GCN models are required to capture more complex interactions. In this paper, we introduce a dynamic fusion mechanism, proposing Lightweight Dynamic Graph Convolutional Networks (LDGCNs) that capture richer non-local interactions by synthesizing higher order information from the input graphs. We further develop two novel parameter saving strategies based on the group graph convolutions and weight tied convolutions to reduce memory usage and model complexity. With the help of these strategies, we are able to train a model with fewer parameters while maintaining the model capacity. Experiments demonstrate that LDGCNs outperform state-of-the-art models on two benchmark datasets for AMR-to-text generation with significantly fewer parameters.

Complex reasoning over text requires understanding and chaining together free-form predicates and logical connectives. Prior work has largely tried to do this either symbolically or with black-box transformers. We present a middle ground between these two extremes: a compositional model reminiscent of neural module networks that can perform chained logical reasoning. This model first finds relevant sentences in the context and then chains them together using neural modules. Our model gives significant performance improvements (up to 29% relative error reduction when combined with a reranker) on ROPES, a recently-introduced complex reasoning dataset.

Minimalist Grammars (Stabler, 1997) are a computationally oriented, and rigorous formalisation of many aspects of Chomsky’s (1995) Minimalist Program. This paper presents the first ever application of this formalism to the task of realistic wide-coverage parsing. The parser uses a linguistically expressive yet highly constrained grammar, together with an adaptation of the A* search algorithm currently used in CCG parsing (Lewis and Steedman, 2014; Lewis et al., 2016), with supertag probabilities provided by a bi-LSTM neural network supertagger trained on MGbank, a corpus of MG derivation trees. We report on some promising initial experimental results for overall dependency recovery as well as on the recovery of certain unbounded long distance dependencies. Finally, although like other MG parsers, ours has a high order polynomial worst case time complexity, we show that in practice its expected time complexity is cubic in the length of the sentence. The parser is publicly available.

Link prediction and entailment graph induction are often treated as different problems. In this paper, we show that these two problems are actually complementary. We train a link prediction model on a knowledge graph of assertions extracted from raw text. We propose an entailment score that exploits the new facts discovered by the link prediction model, and then form entailment graphs between relations. We further use the learned entailments to predict improved link prediction scores. Our results show that the two tasks can benefit from each other. The new entailment score outperforms prior state-of-the-art results on a standard entialment dataset and the new link prediction scores show improvements over the raw link prediction scores.

We introduce a novel semantic parsing task based on Discourse Representation Theory (DRT; Kamp and Reyle 1993). Our model operates over Discourse Representation Tree Structures which we formally define for sentences and documents. We present a general framework for parsing discourse structures of arbitrary length and granularity. We achieve this with a neural model equipped with a supervised hierarchical attention mechanism and a linguistically-motivated copy strategy. Experimental results on sentence- and document-level benchmarks show that our model outperforms competitive baselines by a wide margin.

Lexicalized parsing models are based on the assumptions that (i) constituents are organized around a lexical head and (ii) bilexical statistics are crucial to solve ambiguities. In this paper, we introduce an unlexicalized transition-based parser for discontinuous constituency structures, based on a structure-label transition system and a bi-LSTM scoring system. We compare it with lexicalized parsing models in order to address the question of lexicalization in the context of discontinuous constituency parsing. Our experiments show that unlexicalized models systematically achieve higher results than lexicalized models, and provide additional empirical evidence that lexicalization is not necessary to achieve strong parsing results. Our best unlexicalized model sets a new state of the art on English and German discontinuous constituency treebanks. We further provide a per-phenomenon analysis of its errors on discontinuous constituents.

We introduce a novel transition system for discontinuous constituency parsing. Instead of storing subtrees in a stack –i.e. a data structure with linear-time sequential access– the proposed system uses a set of parsing items, with constant-time random access. This change makes it possible to construct any discontinuous constituency tree in exactly 4n–2 transitions for a sentence of length n. At each parsing step, the parser considers every item in the set to be combined with a focus item and to construct a new constituent in a bottom-up fashion. The parsing strategy is based on the assumption that most syntactic structures can be parsed incrementally and that the set –the memory of the parser– remains reasonably small on average. Moreover, we introduce a provably correct dynamic oracle for the new transition system, and present the first experiments in discontinuous constituency parsing using a dynamic oracle. Our parser obtains state-of-the-art results on three English and German discontinuous treebanks.

AMR-to-text generation is a problem recently introduced to the NLP community, in which the goal is to generate sentences from Abstract Meaning Representation (AMR) graphs. Sequence-to-sequence models can be used to this end by converting the AMR graphs to strings. Approaching the problem while working directly with graphs requires the use of graph-to-sequence models that encode the AMR graph into a vector representation. Such encoding has been shown to be beneficial in the past, and unlike sequential encoding, it allows us to explicitly capture reentrant structures in the AMR graphs. We investigate the extent to which reentrancies (nodes with multiple parents) have an impact on AMR-to-text generation by comparing graph encoders to tree encoders, where reentrancies are not preserved. We show that improvements in the treatment of reentrancies and long-range dependencies contribute to higher overall scores for graph encoders. Our best model achieves 24.40 BLEU on LDC2015E86, outperforming the state of the art by 1.1 points and 24.54 BLEU on LDC2017T10, outperforming the state of the art by 1.24 points.

We present a new neural model for text summarization that first extracts sentences from a document and then compresses them. The pro-posed model offers a balance that sidesteps thedifficulties in abstractive methods while gener-ating more concise summaries than extractivemethods. In addition, our model dynamically determines the length of the output summary based on the gold summaries it observes during training and does not require length constraints typical to extractive summarization. The model achieves state-of-the-art results on the CNN/DailyMail and Newsroom datasets, improving over current extractive and abstractive methods. Human evaluations demonstratethat our model generates concise and informa-tive summaries. We also make available a new dataset of oracle compressive summaries derived automatically from the CNN/DailyMailreference summaries.

Modern state-of-the-art Semantic Role Labeling (SRL) methods rely on expressive sentence encoders (e.g., multi-layer LSTMs) but tend to model only local (if any) interactions between individual argument labeling decisions. This contrasts with earlier work and also with the intuition that the labels of individual arguments are strongly interdependent. We model interactions between argument labeling decisions through iterative refinement. Starting with an output produced by a factorized model, we iteratively refine it using a refinement network. Instead of modeling arbitrary interactions among roles and words, we encode prior knowledge about the SRL problem by designing a restricted network architecture capturing non-local interactions. This modeling choice prevents overfitting and results in an effective model, outperforming strong factorized baseline models on all 7 CoNLL-2009 languages, and achieving state-of-the-art results on 5 of them, including English.

Multi-view learning algorithms are powerful representation learning tools, often exploited in the context of multimodal problems. However, for problems requiring inference at the token-level of a sequence (that is, a separate prediction must be made for every time step), it is often the case that single-view systems are used, or that more than one views are fused in a simple manner. We describe an incremental neural architecture paired with a novel training objective for incremental inference. The network operates on multi-view data. We demonstrate the effectiveness of our approach on the problem of predicting perpetrators in crime drama series, for which our model significantly outperforms previous work and strong baselines. Moreover, we introduce two tasks, crime case and speaker type tagging, that contribute to movie understanding and demonstrate the effectiveness of our model on them.

Neural language models are usually trained using Maximum-Likelihood Estimation (MLE). The corresponding objective function for MLE is derived from the Kullback-Leibler (KL) divergence between the empirical probability distribution representing the data and the parametric probability distribution output by the model. However, the word frequency discrepancies in natural language make performance extremely uneven: while the perplexity is usually very low for frequent words, it is especially difficult to predict rare words. In this paper, we experiment with several families (alpha, beta and gamma) of power divergences, generalized from the KL divergence, for learning language models with an objective different than standard MLE. Intuitively, these divergences should affect the way the probability mass is spread during learning, notably by prioritizing performances on high or low-frequency words. In addition, we implement and experiment with various sampling-based objectives, where the computation of the output layer is only done on a small subset of the vocabulary. They are derived as power generalizations of a softmax approximated via Importance Sampling, and Noise Contrastive Estimation, for accelerated learning. Our experiments on the Penn Treebank and Wikitext-2 show that these power divergences can indeed be used to prioritize learning on the frequent or rare words, and lead to general performance improvements in the case of sampling-based learning.

We describe the systems we developed for Discourse Representation Structure (DRS) parsing as part of the IWCS-2019 Shared Task of DRS Parsing.1 Our systems are based on sequence-to-sequence modeling. To implement our model, we use the open-source neural machine translation system implemented in PyTorch, OpenNMT-py. We experimented with a variety of encoder-decoder models based on recurrent neural networks and the Transformer model. We conduct experiments on the standard benchmark of the Parallel Meaning Bank (PMB 2.2). Our best system achieves a score of 84.8% F1 in the DRS parsing shared task.

We propose a formal model for translating unranked syntactic trees, such as dependency trees, into semantic graphs. These tree-to-graph transducers can serve as a formal basis of transition systems for semantic parsing which recently have been shown to perform very well, yet hitherto lack formalization. Our model features “extended” rules and an arc-factored normal form, comes with an efficient translation algorithm, and can be equipped with weights in a straightforward manner.

Abstract Meaning Representation (AMR) parsing aims at abstracting away from the syntactic realization of a sentence, and denote only its meaning in a canonical form. As such, it is ideal for paraphrase detection, a problem in which one is required to specify whether two sentences have the same meaning. We show that naïve use of AMR in paraphrase detection is not necessarily useful, and turn to describe a technique based on latent semantic analysis in combination with AMR parsing that significantly advances state-of-the-art results in paraphrase detection for the Microsoft Research Paraphrase Corpus. Our best results in the transductive setting are 86.6% for accuracy and 90.0% for F₁ measure.

Abstract Meaning Representation (AMR) research has mostly focused on English. We show that it is possible to use AMR annotations for English as a semantic representation for sentences written in other languages. We exploit an AMR parser for English and parallel corpora to learn AMR parsers for Italian, Spanish, German and Chinese. Qualitative analysis show that the new parsers overcome structural differences between the languages. We further propose a method to evaluate the parsers that does not require gold standard data in the target languages. This method highly correlates with the gold standard evaluation, obtaining a Pearson correlation coefficient of 0.95.

Single document summarization is the task of producing a shorter version of a document while preserving its principal information content. In this paper we conceptualize extractive summarization as a sentence ranking task and propose a novel training algorithm which globally optimizes the ROUGE evaluation metric through a reinforcement learning objective. We use our algorithm to train a neural summarization model on the CNN and DailyMail datasets and demonstrate experimentally that it outperforms state-of-the-art extractive and abstractive systems when evaluated automatically and by humans.

We introduce an open-domain neural semantic parser which generates formal meaning representations in the style of Discourse Representation Theory (DRT; Kamp and Reyle 1993). We propose a method which transforms Discourse Representation Structures (DRSs) to trees and develop a structure-aware model which decomposes the decoding process into three stages: basic DRS structure prediction, condition prediction (i.e., predicates and relations), and referent prediction (i.e., variables). Experimental results on the Groningen Meaning Bank (GMB) show that our model outperforms competitive baselines by a wide margin.

Stock movement prediction is a challenging problem: the market is highly stochastic, and we make temporally-dependent predictions from chaotic data. We treat these three complexities and present a novel deep generative model jointly exploiting text and price signals for this task. Unlike the case with discriminative or topic modeling, our model introduces recurrent, continuous latent variables for a better treatment of stochasticity, and uses neural variational inference to address the intractable posterior inference. We also provide a hybrid objective with temporal auxiliary to flexibly capture predictive dependencies. We demonstrate the state-of-the-art performance of our proposed model on a new stock movement prediction dataset which we collected.

Document modeling is essential to a variety of natural language understanding tasks. We propose to use external information to improve document modeling for problems that can be framed as sentence extraction. We develop a framework composed of a hierarchical document encoder and an attention-based extractor with attention over external information. We evaluate our model on extractive document summarization (where the external information is image captions and the title of the document) and answer selection (where the external information is a question). We show that our model consistently outperforms strong baselines, in terms of both informativeness and fluency (for CNN document summarization) and achieves state-of-the-art results for answer selection on WikiQA and NewsQA.

We show that the general problem of string transduction can be reduced to the problem of sequence labeling. While character deletion and insertions are allowed in string transduction, they do not exist in sequence labeling. We show how to overcome this difference. Our approach can be used with any sequence labeling algorithm and it works best for problems in which string transduction imposes a strong notion of locality (no long range dependencies). We experiment with spelling correction for social media, OCR correction, and morphological inflection, and we see that it behaves better than seq2seq models and yields state-of-the-art results in several cases.

In this paper we argue that crime drama exemplified in television programs such as CSI: Crime Scene Investigation is an ideal testbed for approximating real-world natural language understanding and the complex inferences associated with it. We propose to treat crime drama as a new inference task, capitalizing on the fact that each episode poses the same basic question (i.e., who committed the crime) and naturally provides the answer when the perpetrator is revealed. We develop a new dataset based on CSI episodes, formalize perpetrator identification as a sequence labeling problem, and develop an LSTM-based model which learns from multi-modal data. Experimental results show that an incremental inference strategy is key to making accurate guesses as well as learning from representations fusing textual, visual, and acoustic input.

This paper presents a new method for learning typed entailment graphs from text. We extract predicate-argument structures from multiple-source news corpora, and compute local distributional similarity scores to learn entailments between predicates with typed arguments (e.g., person contracted disease). Previous work has used transitivity constraints to improve local decisions, but these constraints are intractable on large graphs. We instead propose a scalable method that learns globally consistent similarity scores based on new soft constraints that consider both the structures across typed entailment graphs and inside each graph. Learning takes only a few hours to run over 100K predicates and our results show large improvements over local similarity scores on two entailment data sets. We further show improvements over paraphrases and entailments from the Paraphrase Database, and prior state-of-the-art entailment graphs. We show that the entailment graphs improve performance in a downstream task.

This article deals with adversarial attacks towards deep learning systems for Natural Language Processing (NLP), in the context of privacy protection. We study a specific type of attack: an attacker eavesdrops on the hidden representations of a neural text classifier and tries to recover information about the input text. Such scenario may arise in situations when the computation of a neural network is shared across multiple devices, e.g. some hidden representation is computed by a user’s device and sent to a cloud-based model. We measure the privacy of a hidden representation by the ability of an attacker to predict accurately specific private information from it and characterize the tradeoff between the privacy and the utility of neural representations. Finally, we propose several defense methods based on modified training objectives and show that they improve the privacy of neural representations.

We introduce “extreme summarization”, a new single-document summarization task which does not favor extractive strategies and calls for an abstractive modeling approach. The idea is to create a short, one-sentence news summary answering the question “What is the article about?”. We collect a real-world, large-scale dataset for this task by harvesting online articles from the British Broadcasting Corporation (BBC). We propose a novel abstractive model which is conditioned on the article’s topics and based entirely on convolutional neural networks. We demonstrate experimentally that this architecture captures long-range dependencies in a document and recognizes pertinent content, outperforming an oracle extractive system and state-of-the-art abstractive approaches when evaluated automatically and by humans.

Clustering news across languages enables efficient media monitoring by aggregating articles from multilingual sources into coherent stories. Doing so in an online setting allows scalable processing of massive news streams. To this end, we describe a novel method for clustering an incoming stream of multilingual documents into monolingual and crosslingual clusters. Unlike typical clustering approaches that report results on datasets with a small and known number of labels, we tackle the problem of discovering an ever growing number of cluster labels in an online fashion, using real news datasets in multiple languages. In our formulation, the monolingual clusters group together documents while the crosslingual clusters group together monolingual clusters, one per language that appears in the stream. Our method is simple to implement, computationally efficient and produces state-of-the-art results on datasets in German, English and Spanish.

We propose a new sentence simplification task (Split-and-Rephrase) where the aim is to split a complex sentence into a meaning preserving sequence of shorter sentences. Like sentence simplification, splitting-and-rephrasing has the potential of benefiting both natural language processing and societal applications. Because shorter sentences are generally better processed by NLP systems, it could be used as a preprocessing step which facilitates and improves the performance of parsers, semantic role labellers and machine translation systems. It should also be of use for people with reading disabilities because it allows the conversion of longer sentences into shorter ones. This paper makes two contributions towards this new task. First, we create and make available a benchmark consisting of 1,066,115 tuples mapping a single complex sentence to a sequence of sentences expressing the same meaning. Second, we propose five models (vanilla sequence-to-sequence to semantically-motivated models) to understand the difficulty of the proposed task.

Abstract Meaning Representation (AMR) is a semantic representation for natural language that embeds annotations related to traditional tasks such as named entity recognition, semantic role labeling, word sense disambiguation and co-reference resolution. We describe a transition-based parser for AMR that parses sentences left-to-right, in linear time. We further propose a test-suite that assesses specific subtasks that are helpful in comparing AMR parsers, and show that our parser is competitive with the state of the art on the LDC2015E86 dataset and that it outperforms state-of-the-art parsers for recovering named entities and handling polarity.

We present the first prototype of the SUMMA Platform: an integrated platform for multilingual media monitoring. The platform contains a rich suite of low-level and high-level natural language processing technologies: automatic speech recognition of broadcast media, machine translation, automated tagging and classification of named entities, semantic parsing to detect relationships between entities, and automatic construction / augmentation of factual knowledge bases. Implemented on the Docker platform, it can easily be deployed, customised, and scaled to large volumes of incoming media streams.

Canonical correlation analysis (CCA) is a method for reducing the dimension of data represented using two views. It has been previously used to derive word embeddings, where one view indicates a word, and the other view indicates its context. We describe a way to incorporate prior knowledge into CCA, give a theoretical justification for it, and test it by deriving word embeddings and evaluating them on a myriad of datasets.

Adaptor grammars are a flexible, powerful formalism for defining nonparametric, unsupervised models of grammar productions. This flexibility comes at the cost of expensive inference. We address the difficulty of inference through an online algorithm which uses a hybrid of Markov chain Monte Carlo and variational inference. We show that this inference strategy improves scalability without sacrificing performance on unsupervised word segmentation and topic modeling tasks.