When asked to summarize articles or answer questions given a passage, large language models (LLMs) can hallucinate details and respond with unsubstantiated answers that are inaccurate with respect to the input context. This paper describes a simple approach for detecting such **contextual hallucinations**. We hypothesize that contextual hallucinations are related to the extent to which an LLM attends to information in the provided context versus its own generations. Based on this intuition, we propose a simple hallucination detection model whose input features are given by the ratio of attention weights on the context versus newly generated tokens (for each attention head). We find that a linear classifier based on these _lookback ratio_ features is as effective as a richer detector that utilizes the entire hidden states of an LLM or a text-based entailment model. The lookback ratio-based detector—**Lookback Lens**—is found to transfer across tasks and even models, allowing a detector that is trained on a 7B model to be applied (without retraining) to a larger 13B model. We further apply this detector to mitigate contextual hallucinations, and find that a simple classifier-guided decoding approach is able to reduce the amount of hallucination, for example by 9.6% in the XSum summarization task.

Query rewriting is a crucial technique for passage retrieval in open-domain conversational question answering (CQA). It decontexualizes conversational queries into self-contained questions suitable for off-the-shelf retrievers. Existing methods attempt to incorporate retriever’s preference during the training of rewriting models. However, these approaches typically rely on extensive annotations such as in-domain rewrites and/or relevant passage labels, limiting the models’ generalization and adaptation capabilities. In this paper, we introduce AdaQR (Adaptive Query Rewriting), a framework for training query rewriting models with limited rewrite annotations from seed datasets and completely no passage label. Our approach begins by fine-tuning compact large language models using only 10% of rewrite annotations from the seed dataset training split. The models are then utilized to self-sample rewrite candidates for each query instance, further eliminating the expense for human labeling or larger language model prompting often adopted in curating preference data. A novel approach is then proposed to assess retriever’s preference for these candidates with the probability of answers conditioned on the conversational query by marginalizing the Top-K passages. This serves as the reward for optimizing the rewriter further using Direct Preference Optimization (DPO), a process free of rewrite and retrieval annotations. Experimental results on four open-domain CQA datasets demonstrate that AdaQR not only enhances the in-domain capabilities of the rewriter with limited annotation requirement, but also adapts effectively to out-of-domain datasets.

Passage retrieval is a crucial component of modern open-domain question answering (QA) systems, providing information for downstream QA components to generate accurate and transparent answers. In this study we focus on passage re-ranking, proposing a simple yet effective method, Joint Passage Re-ranking (JPR), that optimizes the mutual information between query and passage distributions, integrating both cross-encoders and generative models in the re-ranking process. Experimental results demonstrate that JPR outperforms conventional re-rankers and language model scorers in both open-domain QA retrieval settings and diverse retrieval benchmarks under zero-shot settings.

How can we perform computations over natural language representations to solve tasks that require symbolic and numeric reasoning? We propose natural language embedded programs (NLEP) as a unifying framework for addressing math/symbolic reasoning, natural language understanding, and instruction following tasks. Our approach prompts a language model to generate full Python programs that define functions over data structures which contain natural language representations of structured knowledge. A Python interpreter then executes the generated code and prints the output. Despite using a task-general prompt, we find that this approach can improve upon strong baselines across a range of different tasks including math and symbolic reasoning, text classification, question answering, and instruction following. We found that the generated programs are interpretable since they outline the exact reasoning process followed by the program interpreter.

Recent works have demonstrated the effectiveness of self-alignment in which a large language model is aligned to follow general instructions using instructional data generated from the model itself starting from a handful of human-written seeds. Instead of general alignment, in this work, we focus on self-alignment for expert domain specialization (e.g., biomedicine, finance). As a preliminary, we quantitively show the marginal effect that generic instruction-following training has on downstream expert domains’ performance. To remedy this, we propose self-specialization - allowing for effective model specialization while achieving cross-task generalization by leveraging only a few labeled seeds. Self-specialization offers a data- and parameter-efficient way of “carving out” an expert model out of a generalist pre-trained LLM. Exploring a variety of popular open large models as a base for specialization, our experimental results in both biomedical and financial domains show that our self-specialized models outperform their base models by a large margin, and even larger models that are generally instruction-tuned or that have been adapted to the target domain by other means.

Large language models (LLMs), even when specifically trained to process long input contexts, struggle to capture relevant information located in the middle of their input. This phenomenon has been known as the lost-in-the-middle problem. In this work, we make three contributions. First, we set out to understand the factors that cause this phenomenon. In doing so, we establish a connection between lost-in-the-middle to LLMs’ intrinsic attention bias: LLMs exhibit an U-shaped attention bias where the tokens at the beginning and at the end of its input receive higher attention, regardless of their relevance. Second, we mitigate this positional bias through a calibration mechanism, found-in-the-middle, that allows the model to attend to contexts faithfully according to their relevance, even though when they are in the middle. Third, we show found-in-the-middle not only achieves better performance in locating relevant information within a long context, but also eventually leads to improved retrieval-augmented generation (RAG) performance across various tasks, outperforming existing methods by up to 10 percentage point. These findings open up future directions in understanding LLM attention bias and its potential consequences.

This paper introduces Robust Spin (R-Spin), a data-efficient domain-specific self-supervision method for speaker and noise-invariant speech representations by learning discrete acoustic units with speaker-invariant clustering (Spin). R-Spin resolves Spin’s issues and enhances content representations by learning to predict acoustic pieces. R-Spin offers a 12X reduction in computational resources compared to previous state-of-the-art methods while outperforming them in severely distorted speech scenarios. This paper provides detailed analyses to show how discrete units contribute to speech encoder training and improving robustness in diverse acoustic environments.

In this work, we explore a useful but often neglected methodology for robustness analysis of text generation evaluation metrics: stress tests with synthetic data. Basically, we design and synthesize a wide range of potential errors and check whether they result in a commensurate drop in the metric scores. We examine a range of recently proposed evaluation metrics based on pretrained language models, for the tasks of open-ended generation, translation, and summarization. Our experiments reveal interesting insensitivities, biases, or even loopholes in existing metrics. For example, we find that BERTScore is confused by truncation errors in summarization, and MAUVE (built on top of GPT-2) is insensitive to errors at the beginning or middle of generations. Further, we investigate the reasons behind these blind spots and suggest practical workarounds for a more reliable evaluation of text generation. We have released our code and data at https://github.com/cloudygoose/blindspot_nlg.

Entailment has been recognized as an important metric for evaluating natural language understanding (NLU) models, and recent studies have found that entailment pretraining benefits weakly supervised fine-tuning. In this work, we design a prompting strategy that formulates a number of different NLU tasks as contextual entailment. This approach improves the zero-shot adaptation of pretrained entailment models. Secondly, we notice that self-training entailment-based models with unlabeled data can significantly improve the adaptation performance on downstream tasks. To achieve more stable improvement, we propose the Simple Pseudo-Label Editing (SimPLE) algorithm for better pseudo-labeling quality in self-training. We also found that both pretrained entailment-based models and the self-trained models are robust against adversarial evaluation data. Experiments on binary and multi-class classification tasks show that SimPLE leads to more robust self-training results, indicating that the self-trained entailment models are more efficient and trustworthy than large language models on language understanding tasks.

We propose EAR, a query Expansion And Reranking approach for improving passage retrieval, with the application to open-domain question answering. EAR first applies a query expansion model to generate a diverse set of queries, and then uses a query reranker to select the ones that could lead to better retrieval results. Motivated by the observation that the best query expansion often is not picked by greedy decoding, EAR trains its reranker to predict the rank orders of the gold passages when issuing the expanded queries to a given retriever. By connecting better the query expansion model and retriever, EAR significantly enhances a traditional sparse retrieval method, BM25. Empirically, EAR improves top-5/20 accuracy by 3-8 and 5-10 points in in-domain and out-of-domain settings, respectively, when compared to a vanilla query expansion model, GAR, and a dense retrieval model, DPR.

Large language models (LLMs) have been significantly improved by instruction fine-tuning, but still lack transparency and the ability to utilize up-to-date knowledge and information. In this work, we propose search-augmented instruction learning (SAIL), which grounds the language generation and instruction following abilities on complex search results generated by in-house and external search engines. With an instruction tuning corpus, we collect search results for each training case from different search APIs and domains, and construct a new search-grounded training set containing (instruction, grounding information, response) triplets. We then fine-tune the LLaMA-7B model on the constructed training set. Since the collected results contain unrelated and disputing languages, the model needs to learn to ground on trustworthy search results, filter out distracting passages, and generate the target response. The search result-denoising process entails explicit trustworthy information selection and multi-hop reasoning, since the retrieved passages might be informative but not contain the instruction-following answer. Experiments show that the fine-tuned SAIL-7B model has a strong instruction-following ability, and it performs significantly better on transparency-sensitive tasks, including open-ended question answering and fact checking.

Due to their similarity-based learning objectives, pretrained sentence encoders often internalize stereotypical assumptions that reflect the social biases that exist within their training corpora. In this paper, we describe several kinds of stereotypes concerning different communities that are present in popular sentence representation models, including pretrained next sentence prediction and contrastive sentence representation models. We compare such models to textual entailment models that learn language logic for a variety of downstream language understanding tasks. By comparing strong pretrained models based on text similarity with textual entailment learning, we conclude that the explicit logic learning with textual entailment can significantly reduce bias and improve the recognition of social communities, without an explicit de-biasing process.

Collecting and constructing human-annotated corpora for training conversational question-answering (CQA) models has recently been shown to be inefficient and costly. To solve this problem, previous works have proposed training QA models with automatically generated QA data. In this work, we extend earlier studies on QA synthesis, and propose an efficient QA data generation algorithm under conversational settings. Our model recognizes potential dialogue topics, generates corresponding questions, and extracts answers from grounding passages. To improve the quality of generated QAs and downstream self-training of CQA models, we propose dropout and agreement-based QA selection methods. We conduct experiments on both data augmentation and domain adaptation settings. Experiments on the QuAC and Doc2Dial tasks show that the proposed method can significantly improve the quality of generated QA data, and also improves the accuracy of self-trained CQA models based on the constructed training corpora.

Existing work on controlled text generation (CTG) assumes a control interface of categorical attributes. In this work, we propose a natural language (NL) interface, where we craft a PCFG to embed the control attributes into natural language commands, and propose variants of existing CTG models that take commands as input. In our experiments, we design tailored setups to test the model’s generalization abilities. We find our PCFG-based command generation approach is effective for handling unseen commands compared to fix-set templates. Further, our proposed NL models can effectively generalize to unseen attributes (a new ability enabled by the NL interface), as well as unseen attribute combinations. Interestingly, in model comparisons, the simple conditional generation approach, enhanced with our proposed NL interface, is shown to be a strong baseline in those challenging settings.

Pretrained language models have significantly improved the performance of downstream language understanding tasks, including extractive question answering, by providing high-quality contextualized word embeddings. However, training question answering models still requires large amounts of annotated data for specific domains. In this work, we propose a cooperative self-training framework, RGX, for automatically generating more non-trivial question-answer pairs to improve model performance. RGX is built upon a masked answer extraction task with an interactive learning environment containing an answer entity Recognizer, a question Generator, and an answer eXtractor. Given a passage with a masked entity, the generator generates a question around the entity, and the extractor is trained to extract the masked entity with the generated question and raw texts. The framework allows the training of question generation and answering models on any text corpora without annotation. We further leverage a self-training technique to improve the performance of both question generation and answer extraction models. Experiment results show that RGX outperforms the state-of-the-art (SOTA) pretrained language models and transfer learning approaches on standard question-answering benchmarks, and yields the new SOTA performance under given model size and transfer learning settings.

We propose DiffCSE, an unsupervised contrastive learning framework for learning sentence embeddings. DiffCSE learns sentence embeddings that are sensitive to the difference between the original sentence and an edited sentence, where the edited sentence is obtained by stochastically masking out the original sentence and then sampling from a masked language model. We show that DiffSCE is an instance of equivariant contrastive learning, which generalizes contrastive learning and learns representations that are insensitive to certain types of augmentations and sensitive to other “harmful” types of augmentations. Our experiments show that DiffCSE achieves state-of-the-art results among unsupervised sentence representation learning methods, outperforming unsupervised SimCSE by 2.3 absolute points on semantic textual similarity tasks.

In contrast to recent advances focusing on high-level representation learning across modalities, in this work we present a self-supervised learning framework that is able to learn a representation that captures finer levels of granularity across different modalities such as concepts or events represented by visual objects or spoken words. Our framework relies on a discretized embedding space created via vector quantization that is shared across different modalities. Beyond the shared embedding space, we propose a Cross-Modal Code Matching objective that forces the representations from different views (modalities) to have a similar distribution over the discrete embedding space such that cross-modal objects/actions localization can be performed without direct supervision. We show that the proposed discretized multi-modal fine-grained representation (e.g., pixel/word/frame) can complement high-level summary representations (e.g., video/sentence/waveform) for improved performance on cross-modal retrieval tasks. We also observe that the discretized representation uses individual clusters to represent the same semantic concept across modalities.

The finetuning of pretrained transformer-based language generation models are typically conducted in an end-to-end manner, where the model learns to attend to relevant parts of the input by itself. However, there does not exist a mechanism to directly control the model’s focus. This work aims to develop a control mechanism by which a user can select spans of context as “highlights” for the model to focus on, and generate relevant output. To achieve this goal, we augment a pretrained model with trainable “focus vectors” that are directly applied to the model’s embeddings, while the model itself is kept fixed. These vectors, trained on automatic annotations derived from attribution methods, act as indicators for context importance. We test our approach on two core generation tasks: dialogue response generation and abstractive summarization. We also collect evaluation data where the highlight-generation pairs are annotated by humans. Our experiments show that the trained focus vectors are effective in steering the model to generate outputs that are relevant to user-selected highlights.

Neuropsychological exams are commonly used to diagnose various kinds of cognitive impairment. They typically involve a trained examiner who conducts a series of cognitive tests with a subject. In recent years, there has been growing interest in developing machine learning methods to extract speech and language biomarkers from exam recordings to provide automated input for cognitive assessment. Inspired by recent findings suggesting that the examiner’s language can influence cognitive impairment classifications, in this paper, we study the influence of the examiner on automatic dementia identification decisions in real-world neuropsychological exams. To mitigate the influence of the examiner, we propose a systematic three-stage pipeline for detecting dementia from exam recordings. In the first stage, we perform audio-based speaker diarization (i.e., estimating who spoke when?) by incorporating speaker discriminative features. In the second stage, we employ text-based language models to identify the role of the speaker (i.e., examiner or subject). Finally, in the third stage, we employ text- and audio-based models to detect cognitive impairment from hypothesized subject segments. Our studies suggest that incorporating audio-based diarization followed by text-based role identification helps mitigate the influences from the examiner’s segments. Further, we found that the text and audio modalities complement each other, and the performance improves when we use both modalities. We also perform several carefully designed experimental studies to assess the performance of each stage.

We present Speak, a toolkit that allows researchers to crowdsource speech audio recordings using Amazon Mechanical Turk (MTurk). Speak allows MTurk workers to submit speech recordings in response to a task prompt and stimulus (e.g. image, text excerpt, audio file) defined by researchers, a functionality that is not natively offered by MTurk at the time of writing this paper. Importantly, the toolkit employs numerous measures to ensure that speech recordings collected are of adequate quality, in order to avoid accepting unusable data and prevent abuse/fraud. Speak has demonstrated utility, having collected over 600,000 recordings to date. The toolkit is open-source and available for download.

In this paper we present the first model for directly synthesizing fluent, natural-sounding spoken audio captions for images that does not require natural language text as an intermediate representation or source of supervision. Instead, we connect the image captioning module and the speech synthesis module with a set of discrete, sub-word speech units that are discovered with a self-supervised visual grounding task. We conduct experiments on the Flickr8k spoken caption dataset in addition to a novel corpus of spoken audio captions collected for the popular MSCOCO dataset, demonstrating that our generated captions also capture diverse visual semantics of the images they describe. We investigate several different intermediate speech representations, and empirically find that the representation must satisfy several important properties to serve as drop-in replacements for text.

Automatic detection of toxic language plays an essential role in protecting social media users, especially minority groups, from verbal abuse. However, biases toward some attributes, including gender, race, and dialect, exist in most training datasets for toxicity detection. The biases make the learned models unfair and can even exacerbate the marginalization of people. Considering that current debiasing methods for general natural language understanding tasks cannot effectively mitigate the biases in the toxicity detectors, we propose to use invariant rationalization (InvRat), a game-theoretic framework consisting of a rationale generator and a predictor, to rule out the spurious correlation of certain syntactic patterns (e.g., identity mentions, dialect) to toxicity labels. We empirically show that our method yields lower false positive rate in both lexical and dialectal attributes than previous debiasing methods.

Online users today are exposed to misleading and propagandistic news articles and media posts on a daily basis. To counter thus, a number of approaches have been designed aiming to achieve a healthier and safer online news and media consumption. Automatic systems are able to support humans in detecting such content; yet, a major impediment to their broad adoption is that besides being accurate, the decisions of such systems need also to be interpretable in order to be trusted and widely adopted by users. Since misleading and propagandistic content influences readers through the use of a number of deception techniques, we propose to detect and to show the use of such techniques as a way to offer interpretability. In particular, we define qualitatively descriptive features and we analyze their suitability for detecting deception techniques. We further show that our interpretable features can be easily combined with pre-trained language models, yielding state-of-the-art results.

In this work, we study how the finetuning stage in the pretrain-finetune framework changes the behavior of a pretrained neural language generator. We focus on the transformer encoder-decoder model for the open-domain dialogue response generation task. Our major finding is that after standard finetuning, the model forgets some of the important language generation skills acquired during large-scale pretraining. We demonstrate the forgetting phenomenon through a set of detailed behavior analysis from the perspectives of knowledge transfer, context sensitivity, and function space projection. As a preliminary attempt to alleviate the forgetting problem, we propose an intuitive finetuning strategy named “mix-review”. We find that mix-review effectively regularizes the finetuning process, and the forgetting problem is alleviated to some extent. Finally, we discuss interesting behavior of the resulting dialogue model and its implications.

Exposure bias has been regarded as a central problem for auto-regressive language models (LM). It claims that teacher forcing would cause the test-time generation to be incrementally distorted due to the training-generation discrepancy. Although a lot of algorithms have been proposed to avoid teacher forcing and therefore alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is. In this work, we focus on the task of open-ended language generation, propose metrics to quantify the impact of exposure bias in the aspects of quality, diversity, and consistency. Our key intuition is that if we feed ground-truth data prefixes (instead of prefixes generated by the model itself) into the model and ask it to continue the generation, the performance should become much better because the training-generation discrepancy in the prefix is removed. Both automatic and human evaluations are conducted in our experiments. On the contrary to the popular belief in exposure bias, we find that the the distortion induced by the prefix discrepancy is limited, and does not seem to be incremental during the generation. Moreover, our analysis reveals an interesting self-recovery ability of the LM, which we hypothesize to be countering the harmful effects from exposure bias.

In this work, we propose a novel goal-oriented dialog task, automatic symptom detection. We build a system that can interact with patients through dialog to detect and collect clinical symptoms automatically, which can save a doctor’s time interviewing the patient. Given a set of explicit symptoms provided by the patient to initiate a dialog for diagnosing, the system is trained to collect implicit symptoms by asking questions, in order to collect more information for making an accurate diagnosis. After getting the reply from the patient for each question, the system also decides whether current information is enough for a human doctor to make a diagnosis. To achieve this goal, we propose two neural models and a training pipeline for the multi-step reasoning task. We also build a knowledge graph as additional inputs to further improve model performance. Experiments show that our model significantly outperforms the baseline by 4%, discovering 67% of implicit symptoms on average with a limited number of questions.

This work studies the widely adopted ancestral sampling algorithms for auto-regressive language models. We use the quality-diversity (Q-D) trade-off to investigate three popular sampling methods (top-k, nucleus and tempered sampling). We focus on the task of open-ended language generation, and first show that the existing sampling algorithms have similar performance. By carefully inspecting the transformations defined by different sampling algorithms, we identify three key properties that are shared among them: entropy reduction, order preservation, and slope preservation. To validate the importance of the identified properties, we design two sets of new sampling methods: one set in which each algorithm satisfies all three properties, and one set in which each algorithm violates at least one of the properties. We compare their performance with existing algorithms, and find that violating the identified properties could lead to drastic performance degradation, as measured by the Q-D trade-off. On the other hand, we find that the set of sampling algorithms that satisfy these properties performs on par with the existing sampling algorithms.

Although deep learning models have brought tremendous advancements to the field of open-domain dialogue response generation, recent research results have revealed that the trained models have undesirable generation behaviors, such as malicious responses and generic (boring) responses. In this work, we propose a framework named “Negative Training” to minimize such behaviors. Given a trained model, the framework will first find generated samples that exhibit the undesirable behavior, and then use them to feed negative training signals for fine-tuning the model. Our experiments show that negative training can significantly reduce the hit rate of malicious responses, or discourage frequent responses and improve response diversity.

Training objectives based on predictive coding have recently been shown to be very effective at learning meaningful representations from unlabeled speech. One example is Autoregressive Predictive Coding (Chung et al., 2019), which trains an autoregressive RNN to generate an unseen future frame given a context such as recent past frames. The basic hypothesis of these approaches is that hidden states that can accurately predict future frames are a useful representation for many downstream tasks. In this paper we extend this hypothesis and aim to enrich the information encoded in the hidden states by training the model to make more accurate future predictions. We propose an auxiliary objective that serves as a regularization to improve generalization of the future frame prediction task. Experimental results on phonetic classification, speech recognition, and speech translation not only support the hypothesis, but also demonstrate the effectiveness of our approach in learning representations that contain richer phonetic content.

Predicting the political bias and the factuality of reporting of entire news outlets are critical elements of media profiling, which is an understudied but an increasingly important research direction. The present level of proliferation of fake, biased, and propagandistic content online has made it impossible to fact-check every single suspicious claim, either manually or automatically. Thus, it has been proposed to profile entire news outlets and to look for those that are likely to publish fake or biased content. This makes it possible to detect likely “fake news” the moment they are published, by simply checking the reliability of their source. From a practical perspective, political bias and factuality of reporting have a linguistic aspect but also a social context. Here, we study the impact of both, namely (i) what was written (i.e., what was published by the target medium, and how it describes itself in Twitter) vs. (ii) who reads it (i.e., analyzing the target medium’s audience on social media). We further study (iii) what was written about the target medium (in Wikipedia). The evaluation results show that what was written matters most, and we further show that putting all information sources together yields huge improvements over the current state-of-the-art.

This paper investigates contextual word representation models from the lens of similarity analysis. Given a collection of trained models, we measure the similarity of their internal representations and attention. Critically, these models come from vastly different architectures. We use existing and novel similarity measures that aim to gauge the level of localization of information in the deep models, and facilitate the investigation of which design factors affect model similarity, without requiring any external linguistic annotation. The analysis reveals that models within the same family are more similar to one another, as may be expected. Surprisingly, different architectures have rather similar representations, but different individual neurons. We also observed differences in information localization in lower and higher layers and found that higher layers are more affected by fine-tuning on downstream tasks.

Despite the recent success of deep neural networks in natural language processing and other spheres of artificial intelligence, their interpretability remains a challenge. We analyze the representations learned by neural machine translation (NMT) models at various levels of granularity and evaluate their quality through relevant extrinsic properties. In particular, we seek answers to the following questions: (i) How accurately is word structure captured within the learned representations, which is an important aspect in translating morphologically rich languages? (ii) Do the representations capture long-range dependencies, and effectively handle syntactically divergent languages? (iii) Do the representations capture lexical semantics? We conduct a thorough investigation along several parameters: (i) Which layers in the architecture capture each of these linguistic phenomena; (ii) How does the choice of translation unit (word, character, or subword unit) impact the linguistic properties captured by the underlying representations? (iii) Do the encoder and decoder learn differently and independently? (iv) Do the representations learned by multilingual NMT models capture the same amount of linguistic information as their bilingual counterparts? Our data-driven, quantitative evaluation illuminates important aspects in NMT models and their ability to capture various linguistic phenomena. We show that deep NMT models trained in an end-to-end fashion, without being provided any direct supervision during the training process, learn a non-trivial amount of linguistic information. Notable findings include the following observations: (i) Word morphology and part-of-speech information are captured at the lower layers of the model; (ii) In contrast, lexical semantics or non-local syntactic and semantic dependencies are better represented at the higher layers of the model; (iii) Representations learned using characters are more informed about word-morphology compared to those learned using subword units; and (iv) Representations learned by multilingual models are richer compared to bilingual models.

We explore the task of predicting the leading political ideology or bias of news articles. First, we collect and release a large dataset of 34,737 articles that were manually annotated for political ideology –left, center, or right–, which is well-balanced across both topics and media. We further use a challenging experimental setup where the test examples come from media that were not seen during training, which prevents the model from learning to detect the source of the target news article instead of predicting its political ideology. From a modeling perspective, we propose an adversarial media adaptation, as well as a specially adapted triplet loss. We further add background information about the source, and we show that it is quite helpful for improving article-level prediction. Our experimental results show very sizable improvements over using state-of-the-art pre-trained Transformers in this challenging setup.

We describe our submission to SemEval-2019 Task 4 on Hyperpartisan News Detection. We rely on a variety of engineered features originally used to detect propaganda. This is based on the assumption that biased messages are propagandistic and promote a particular political cause or viewpoint. In particular, we trained a logistic regression model with features ranging from simple bag of words to vocabulary richness and text readability. Our system achieved 72.9% accuracy on the manually annotated testset, and 60.8% on the test data that was obtained with distant supervision. Additional experiments showed that significant performance gains can be achieved with better feature pre-processing.

Common language models typically predict the next word given the context. In this work, we propose a method that improves language modeling by learning to align the given context and the following phrase. The model does not require any linguistic annotation of phrase segmentation. Instead, we define syntactic heights and phrase segmentation rules, enabling the model to automatically induce phrases, recognize their task-specific heads, and generate phrase embeddings in an unsupervised learning manner. Our method can easily be applied to language models with different network architectures since an independent module is used for phrase induction and context-phrase alignment, and no change is required in the underlying language modeling network. Experiments have shown that our model outperformed several strong baseline models on different data sets. We achieved a new state-of-the-art performance of 17.4 perplexity on the Wikitext-103 dataset. Additionally, visualizing the outputs of the phrase induction module showed that our model is able to learn approximate phrase-level structural knowledge without any annotation.

The field of natural language processing has seen impressive progress in recent years, with neural network models replacing many of the traditional systems. A plethora of new models have been proposed, many of which are thought to be opaque compared to their feature-rich counterparts. This has led researchers to analyze, interpret, and evaluate neural networks in novel and more fine-grained ways. In this survey paper, we review analysis methods in neural language processing, categorize them according to prominent research trends, highlight existing limitations, and point to potential directions for future work.

In the context of fake news, bias, and propaganda, we study two important but relatively under-explored problems: (i) trustworthiness estimation (on a 3-point scale) and (ii) political ideology detection (left/right bias on a 7-point scale) of entire news outlets, as opposed to evaluating individual articles. In particular, we propose a multi-task ordinal regression framework that models the two problems jointly. This is motivated by the observation that hyper-partisanship is often linked to low trustworthiness, e.g., appealing to emotions rather than sticking to the facts, while center media tend to be generally more impartial and trustworthy. We further use several auxiliary tasks, modeling centrality, hyper-partisanship, as well as left-vs.-right bias on a coarse-grained scale. The evaluation results show sizable performance gains by the joint models over models that target the problems in isolation.

We present FAKTA which is a unified framework that integrates various components of a fact-checking process: document retrieval from media sources with various types of reliability, stance detection of documents with respect to given claims, evidence extraction, and linguistic analysis. FAKTA predicts the factuality of given claims and provides evidence at the document and sentence level to explain its predictions.

We study cross-lingual stance detection, which aims to leverage labeled data in one language to identify the relative perspective (or stance) of a given document with respect to a claim in a different target language. In particular, we introduce a novel contrastive language adaptation approach applied to memory networks, which ensures accurate alignment of stances in the source and target languages, and can effectively deal with the challenge of limited labeled data in the target language. The evaluation results on public benchmark datasets and comparison against current state-of-the-art approaches demonstrate the effectiveness of our approach.

We introduce Tanbih, a news aggregator with intelligent analysis tools to help readers understanding what’s behind a news story. Our system displays news grouped into events and generates media profiles that show the general factuality of reporting, the degree of propagandistic content, hyper-partisanship, leading political ideology, general frame of reporting, and stance with respect to various claims and topics of a news outlet. In addition, we automatically analyse each article to detect whether it is propagandistic and to determine its stance with respect to a number of controversial topics.

We present a multi-task learning model that leverages large amount of textual information from existing datasets to improve stance prediction. In particular, we utilize multiple NLP tasks under both unsupervised and supervised settings for the target stance prediction task. Our model obtains state-of-the-art performance on a public benchmark dataset, Fake News Challenge, outperforming current approaches by a wide margin.

We present an effective end-to-end memory network model that jointly (i) predicts whether a given document can be considered as relevant evidence for a given claim, and (ii) extracts snippets of evidence that can be used to reason about the factuality of the target claim. Our model combines the advantages of convolutional and recurrent neural networks as part of a memory network. We further introduce a similarity matrix at the inference level of the memory network in order to extract snippets of evidence for input claims more accurately. Our experiments on a public benchmark dataset, FakeNewsChallenge, demonstrate the effectiveness of our approach.

Although transfer learning has been shown to be successful for tasks like object and speech recognition, its applicability to question answering (QA) has yet to be well-studied. In this paper, we conduct extensive experiments to investigate the transferability of knowledge learned from a source QA dataset to a target dataset using two QA models. The performance of both models on a TOEFL listening comprehension test (Tseng et al., 2016) and MCTest (Richardson et al., 2013) is significantly improved via a simple transfer learning technique from MovieQA (Tapaswi et al., 2016). In particular, one of the models achieves the state-of-the-art on all target datasets; for the TOEFL listening comprehension test, it outperforms the previous best model by 7%. Finally, we show that transfer learning is helpful even in unsupervised scenarios when correct answers for target QA dataset examples are not available.

A reasonable approach for fact checking a claim involves retrieving potentially relevant documents from different sources (e.g., news websites, social media, etc.), determining the stance of each document with respect to the claim, and finally making a prediction about the claim’s factuality by aggregating the strength of the stances, while taking the reliability of the source into account. Moreover, a fact checking system should be able to explain its decision by providing relevant extracts (rationales) from the documents. Yet, this setup is not directly supported by existing datasets, which treat fact checking, document retrieval, source credibility, stance detection and rationale extraction as independent tasks. In this paper, we support the interdependencies between these tasks as annotations in the same corpus. We implement this setup on an Arabic fact checking corpus, the first of its kind.

We propose a process for investigating the extent to which sentence representations arising from neural machine translation (NMT) systems encode distinct semantic phenomena. We use these representations as features to train a natural language inference (NLI) classifier based on datasets recast from existing semantic annotations. In applying this process to a representative NMT system, we find its encoder appears most suited to supporting inferences at the syntax-semantics interface, as compared to anaphora resolution requiring world knowledge. We conclude with a discussion on the merits and potential deficiencies of the existing process, and how it may be improved and extended as a broader framework for evaluating semantic coverage

Neuropsychological examinations are an important screening tool for the presence of cognitive conditions (e.g. Alzheimer’s, Parkinson’s Disease), and require a trained tester to conduct the exam through spoken interactions with the subject. While audio is relatively easy to record, it remains a challenge to automatically diarize (who spoke when?), decode (what did they say?), and assess a subject’s cognitive health. This paper demonstrates a method to determine the cognitive health (impaired or not) of 92 subjects, from audio that was diarized using an automatic speech recognition system trained on TED talks and on the structured language used by testers and subjects. Using leave-one-out cross validation and logistic regression modeling we show that even with noisily decoded data (81% WER) we can still perform accurate enough diarization (0.02% confusion rate) to determine the cognitive state of a subject (0.76 AUC).

We present the results and the findings of the Second VarDial Evaluation Campaign on Natural Language Processing (NLP) for Similar Languages, Varieties and Dialects. The campaign was organized as part of the fifth edition of the VarDial workshop, collocated with COLING’2018. This year, the campaign included five shared tasks, including two task re-runs – Arabic Dialect Identification (ADI) and German Dialect Identification (GDI) –, and three new tasks – Morphosyntactic Tagging of Tweets (MTT), Discriminating between Dutch and Flemish in Subtitles (DFS), and Indo-Aryan Language Identification (ILI). A total of 24 teams submitted runs across the five shared tasks, and contributed 22 system description papers, which were included in the VarDial workshop proceedings and are referred to in this report.

We present a study on predicting the factuality of reporting and bias of news media. While previous work has focused on studying the veracity of claims or documents, here we are interested in characterizing entire news media. This is an under-studied, but arguably important research problem, both in its own right and as a prior for fact-checking systems. We experiment with a large list of news websites and with a rich set of features derived from (i) a sample of articles from the target news media, (ii) its Wikipedia page, (iii) its Twitter account, (iv) the structure of its URL, and (v) information about the Web traffic it attracts. The experimental results show sizable performance gains over the baseline, and reveal the importance of each feature type.

While neural machine translation (NMT) models provide improved translation quality in an elegant framework, it is less clear what they learn about language. Recent work has started evaluating the quality of vector representations learned by NMT models on morphological and syntactic tasks. In this paper, we investigate the representations learned at different layers of NMT encoders. We train NMT systems on parallel data and use the models to extract features for training a classifier on two tasks: part-of-speech and semantic tagging. We then measure the performance of the classifier as a proxy to the quality of the original NMT model for the given task. Our quantitative analysis yields interesting insights regarding representation learning in NMT models. For instance, we find that higher layers are better at learning semantics while lower layers tend to be better for part-of-speech tagging. We also observe little effect of the target language on source-side representations, especially in higher quality models.

Given a collection of images and spoken audio captions, we present a method for discovering word-like acoustic units in the continuous speech signal and grounding them to semantically relevant image regions. For example, our model is able to detect spoken instances of the word ‘lighthouse’ within an utterance and associate them with image regions containing lighthouses. We do not use any form of conventional automatic speech recognition, nor do we use any text transcriptions or conventional linguistic annotations. Our model effectively implements a form of spoken language acquisition, in which the computer learns not only to recognize word categories by sound, but also to enrich the words it learns with semantics by grounding them in images.

Neural machine translation (MT) models obtain state-of-the-art performance while maintaining a simple, end-to-end architecture. However, little is known about what these models learn about source and target languages during the training process. In this work, we analyze the representations learned by neural MT models at various levels of granularity and empirically evaluate the quality of the representations for learning morphology through extrinsic part-of-speech and morphological tagging tasks. We conduct a thorough investigation along several parameters: word-based vs. character-based representations, depth of the encoding layer, the identity of the target language, and encoder vs. decoder representations. Our data-driven, quantitative evaluation sheds light on important aspects in the neural MT system and its ability to capture word structure.

Discriminating between closely-related language varieties is considered a challenging and important task. This paper describes our submission to the DSL 2016 shared-task, which included two sub-tasks: one on discriminating similar languages and one on identifying Arabic dialects. We developed a character-level neural network for this task. Given a sequence of characters, our model embeds each character in vector space, runs the sequence through multiple convolutions with different filter widths, and pools the convolutional representations to obtain a hidden vector representation of the text that is used for predicting the language or dialect. We primarily focused on the Arabic dialect identification task and obtained an F1 score of 0.4834, ranking 6th out of 18 participants. We also analyze errors made by our system on the Arabic data in some detail, and point to challenges such an approach is faced with.

In real-world data, e.g., from Web forums, text is often contaminated with redundant or irrelevant content, which leads to introducing noise in machine learning algorithms. In this paper, we apply Long Short-Term Memory networks with an attention mechanism, which can select important parts of text for the task of similar question retrieval from community Question Answering (cQA) forums. In particular, we use the attention weights for both selecting entire sentences and their subparts, i.e., word/chunk, from shallow syntactic trees. More interestingly, we apply tree kernels to the filtered text representations, thus exploiting the implicit features of the subtree space for learning question reranking. Our results show that the attention-based pruning allows for achieving the top position in the cQA challenge of SemEval 2016, with a relatively large gap from the other participants while greatly decreasing running time.

We present a model of unsupervised phonological lexicon discovery—the problem of simultaneously learning phoneme-like and word-like units from acoustic input. Our model builds on earlier models of unsupervised phone-like unit discovery from acoustic data (Lee and Glass, 2012), and unsupervised symbolic lexicon discovery using the Adaptor Grammar framework (Johnson et al., 2006), integrating these earlier approaches using a probabilistic model of phonological variation. We show that the model is competitive with state-of-the-art spoken term discovery systems, and present analyses exploring the model’s behavior and the kinds of linguistic structures it learns.