Alignment serves as an important step to steer large language models (LLMs) towards human preferences. In this paper, we propose an automatic way to construct contrastive data for LLM, using preference pairs from multiple models of varying strengths (e.g., InstructGPT, ChatGPT and GPT-4). We compare the contrastive techniques of SLiC and DPO to SFT baselines and find that DPO provides a step-function improvement even after continuing SFT saturates. We also explore a data curriculum learning scheme for contrastive post-training, which starts by learning from “easier” pairs and transitioning to “harder” ones, which further improves alignment. Finally, we scale up our experiments to train with more data and larger models like Orca. Remarkably, our automatic contrastive post-training further improves the performance of Orca, already a state-of-the-art instruction learning model tuned with GPT-4 outputs, to outperform ChatGPT.

Effectively training language models on longinputs poses many technical challenges. As acost consideration, languages models are pre-trained on a fixed sequence length before beingadapted to longer sequences. We explore var-ious methods for adapting models to longerinputs by training on segmented sequences andan interpolation-based method for extendingabsolute positional embeddings. We developa training procedure to extend the input con-text size of pretrained models with no architec-tural changes and no additional memory coststhan training on the original input lengths. Bysub-sampling segments from long inputs whilemaintaining their original position the model isable to learn new positional interactions. Ourmethod benefits both models trained with abso-lute positional embeddings, by extending theirinput contexts, as well as popular relative posi-tional embedding methods showing a reducedperplexity on sequences longer than they weretrained on. We demonstrate our method canextend input contexts by a factor of 4× whileimproving perplexity.

Large language models (LLMs) such as GPT-3 and GPT-4 are powerful but their weights are often publicly unavailable and their immense sizes make the models difficult to be tuned with common hardware. As a result, effectively tuning these models with large-scale supervised data can be challenging. As an alternative, In-Context Learning (ICL) can only use a small number of supervised examples due to context length limits. In this paper, we propose Super In-Context Learning (SuperICL) which allows black-box LLMs to work with locally fine-tuned smaller models, resulting in superior performance on supervised tasks. Our experiments demonstrate that SuperICL can improve performance beyond state-of-the-art fine-tuned models while addressing the instability problem of in-context learning.

We consider the task of building a dialogue system that can motivate users to adopt positive lifestyle changes, Motivational Interviewing (MI). Addressing such a task requires a system that could infer how to motivate the user effectively. We propose DIIR, a framework that is capable of learning and applying conversation strategies in the form of natural language inductive rules from expert demonstrations. Automatic and human evaluation on instruction-following large language models show natural language strategies descriptions discovered by DIIR can improve active listening skills, reduce unsolicited advice, and promote more collaborative and less authoritative conversations, outperforming in-context demonstrations that are over 50 times longer.

Do current large language models (LLMs) better solve graph reasoning and generation tasks with parameter updates? In this paper, we propose InstructGraph, a framework that empowers LLMs with the abilities of graph reasoning and generation by instruction tuning and preference alignment. Specifically, we first propose a structured format verbalizer to unify all graph data into a universal code-like format, which can simply represent the graph without any external graph-specific encoders. Furthermore, a graph instruction tuning stage is introduced to guide LLMs in solving graph reasoning and generation tasks. Finally, we identify potential hallucination problems in graph tasks and sample negative instances for preference alignment, the target of which is to enhance the output’s reliability of the model. Extensive experiments across multiple graph-centric tasks exhibit that InstructGraph can achieve the best performance and outperform GPT-4 and LLaMA2 by more than 13% and 38%, respectively.

Synthetic users are cost-effective proxies for real users in the evaluation of conversational recommender systems. Large language models show promise in simulating human-like behavior, raising the question of their ability to represent a diverse population of users. We introduce a new protocol to measure the degree to which language models can accurately emulate human behavior in conversational recommendation. This protocol is comprised of five tasks, each designed to evaluate a key property that a synthetic user should exhibit: choosing which items to talk about, expressing binary preferences, expressing open-ended preferences, requesting recommendations, and giving feedback. Through evaluation of baseline simulators, we demonstrate these tasks effectively reveal deviations of language models from human behavior, and offer insights on how to reduce the deviations with model selection and prompting strategies.

Large language models (LLMs) often generate biased outputs containing offensive, toxic, or stereotypical text. Existing LLM alignment methods such as reinforcement learning from human feedback (RLHF) alleviate biases primarily based on reward signals from current model outputs without considering the source of biases. In this work, to explore how biases are formed, we revisit LLMs’ text generation from a causal perspective. We identify pretraining data and input prompts, which contain semantic correlations of textual phrases, as two confounders between LLMs and model outputs causing biases. Inspired by our causal view, we leverage the reward model in RL alignment as an instrumental variable to perform causal intervention on LLMs. Utilizing the reward difference between an initial LLM and intervened LLM as interventional feedback to guide RL finetuning, we propose Causality-Aware Alignment (CAA) for LLM debiasing. Experiments on two text generation tasks with three different alignment objectives demonstrate the advantages of our method in aligning LLMs to generate less biased and safer outputs.

Large language models (LLMs) often require task-relevant knowledge to augment their internal knowledge through prompts. However, simply injecting external knowledge into prompts does not guarantee that LLMs can identify and use relevant information in the prompts to conduct chain-of-thought reasoning, especially when the LLM’s internal knowledge is derived from biased information on the pretraining data. In this paper, we propose a novel causal view to formally explain the internal knowledge bias of LLMs via a Structural Causal Model (SCM). We review the chain-of-thought (CoT) prompting from a causal perspective and discover that the biased information from pretrained models can impair LLMs’ reasoning abilities. When the CoT reasoning paths are misled by irrelevant information from prompts and are logically incorrect, simply editing factual information is insufficient to reach the correct answer. To estimate the confounding effect on CoT reasoning in LLMs, we use external knowledge as an instrumental variable. We further introduce CoT as a mediator to conduct front-door adjustment and generate logically correct CoTs where the spurious correlation between LLMs’ pretrained knowledge and task queries is reduced. With extensive experiments, we validate that our approach enables more accurate CoT reasoning and enhances LLM generation on knowledge-intensive tasks.

While recent works have been considerably improving the quality of the natural language explanations (NLEs) generated by a model to justify its predictions, there is very limited research in detecting and alleviating inconsistencies among generated NLEs. In this work, we leverage external knowledge bases to significantly improve on an existing adversarial attack for detecting inconsistent NLEs. We apply our attack to high-performing NLE models and show that models with higher NLE quality do not necessarily generate fewer inconsistencies. Moreover, we propose an off-the-shelf mitigation method to alleviate inconsistencies by grounding the model into external background knowledge. Our method decreases the inconsistencies of previous high-performing NLE models as detected by our attack.

Effectively scaling large Transformer models is a main driver of recent advances in natural language processing. Dynamic neural networks, as an emerging research direction, are capable of scaling up neural networks with sub-linear increases in computation and time by dynamically adjusting their computational path based on the input. Dynamic neural networks could be a promising solution to the growing parameter numbers of pretrained language models, allowing both model pretraining with trillions of parameters and faster inference on mobile devices. In this survey, we summarize the progress of three types of dynamic neural networks in NLP: skimming, mixture of experts, and early exit. We also highlight current challenges in dynamic neural networks and directions for future research.

Pre-training models with large crawled corpora can lead to issues such as toxicity and bias, as well as copyright and privacy concerns. A promising way of alleviating such concerns is to conduct pre-training with synthetic tasks and data, since no real-world information is ingested by the model. Our goal in this paper is to understand the factors that contribute to the effectiveness of pre-training models when using synthetic resources, particularly in the context of neural machine translation. We propose several novel approaches to pre-training translation models that involve different levels of lexical and structural knowledge, including: 1) generating obfuscated data from a large parallel corpus 2) concatenating phrase pairs extracted from a small word-aligned corpus, and 3) generating synthetic parallel data without real human language corpora. Our experiments on multiple language pairs reveal that pre-training benefits can be realized even with high levels of obfuscation or purely synthetic parallel data. We hope the findings from our comprehensive empirical analysis will shed light on understanding what matters for NMT pre-training, as well as pave the way for the development of more efficient and less toxic models.

Engaging with discussion of TV shows online often requires individuals to refrain from consuming show-related content for extended periods to avoid spoilers. While existing research on spoiler detection shows promising results in safeguarding viewers from general spoilers, it fails to address the issue of users abstaining from show-related content during their watch. This is primarily because the definition of a spoiler varies depending on the viewer’s progress in the show, and conventional spoiler detection methods lack the granularity to capture this complexity. To tackle this challenge, we propose the task of spoiler matching, which involves assigning an episode number to a spoiler given a specific TV show. We frame this task as semantic text matching and introduce a dataset comprised of comments and episode summaries to evaluate model performance. Given the length of each example, our dataset can also serve as a benchmark for long-range language models.

Chat models, such as ChatGPT, have shown impressive capabilities and have been rapidly adopted across numerous domains. However, these models are only accessible through a restricted API, creating barriers for new research and progress in the field. We propose a pipeline that can automatically generate a high-quality multi-turn chat corpus by leveraging ChatGPT to engage in a conversation with itself. Subsequently, we employ parameter-efficient tuning to enhance LLaMA, an open-source large language model. The resulting model, named Baize, demonstrates good performance in multi-turn dialogues with guardrails that minimize potential risks. Additionally, we propose a new technique called Self-Distill with Feedback, to further improve the performance of the Baize models with feedback from ChatGPT.

Curated datasets for healthcare are often limited due to the need of human annotations from experts. In this paper, we present MedEval, a multi-level, multi-task, and multi-domain medical benchmark to facilitate the development of language models for healthcare. MedEval is comprehensive and consists of data from several healthcare systems and spans 35 human body regions from 8 examination modalities. With 22,779 collected sentences and 21,228 reports, we provide expert annotations at multiple levels, offering a granular potential usage of the data and supporting a wide range of tasks. Moreover, we systematically evaluated 10 generic and domain-specific language models under zero-shot and finetuning settings, from domain-adapted baselines in healthcare to general-purposed state-of-the-art large language models (e.g., ChatGPT). Our evaluations reveal varying effectiveness of the two categories of language models across different tasks, from which we notice the importance of instruction tuning for few-shot usage of large language models. Our investigation paves the way toward benchmarking language models for healthcare and provides valuable insights into the strengths and limitations of adopting large language models in medical domains, informing their practical applications and future advancements.

Debiasing methods in NLP models traditionally focus on isolating information related to a sensitive attribute (e.g., gender or race). We instead argue that a favorable debiasing method should use sensitive information ‘fairly,’ with explanations, rather than blindly eliminating it. This fair balance is often subjective and can be challenging to achieve algorithmically. We explore two interactive setups with a frozen predictive model and show that users able to provide feedback can achieve a better and fairer balance between task performance and bias mitigation. In one setup, users, by interacting with test examples, further decreased bias in the explanations (5-8%) while maintaining the same prediction accuracy. In the other setup, human feedback was able to disentangle associated bias and predictive information from the input leading to superior bias mitigation and improved task performance (4-5%) simultaneously.

There has recently been growing interest in the automatic generation of cooking recipes that satisfy some form of dietary restrictions, thanks in part to the availability of online recipe data. Prior studies have used pre-trained language models, or relied on small paired recipe data (e.g., a recipe paired with a similar one that satisfies a dietary constraint). However, pre-trained language models generate inconsistent or incoherent recipes, and paired datasets are not available at scale. We address these deficiencies with RecipeCrit, a hierarchical denoising auto-encoder that edits recipes given ingredient-level critiques. The model is trained for recipe completion to learn semantic relationships within recipes. Our work’s main innovation is our unsupervised critiquing module that allows users to edit recipes by interacting with the predicted ingredients; the system iteratively rewrites recipes to satisfy users’ feedback. Experiments onthe Recipe1M recipe dataset show that our model can more effectively edit recipes compared to strong language-modeling baselines, creating recipes that satisfy user constraints and are more correct, serendipitous, coherent, and relevant as measured by human judges.

Masked language modeling (MLM) plays a key role in pretraining large language models. But the MLM objective is often dominated by high-frequency words that are sub-optimal for learning factual knowledge. In this work, we propose an approach for influencing MLM pretraining in a way that can improve language model performance on a variety of knowledge-intensive tasks. We force the language model to prioritize informative words in a fully unsupervised way. Experiments demonstrate that the proposed approach can significantly improve the performance of pretrained language models on tasks such as factual recall, question answering, sentiment analysis, and natural language inference in a closed-book setting.

Prompt-based learning (i.e., prompting) is an emerging paradigm for exploiting knowledge learned by a pretrained language model. In this paper, we propose Automatic Multi-Label Prompting (AMuLaP), a simple yet effective method to automatically select label mappings for few-shot text classification with prompting. Our method exploits one-to-many label mappings and a statistics-based algorithm to select label mappings given a prompt template. Our experiments demonstrate that AMuLaP achieves competitive performance on the GLUE benchmark without human effort or external resources.

A limitation of current neural dialog models is that they tend to suffer from a lack of specificity and informativeness in generated responses, primarily due to dependence on training data that covers a limited variety of scenarios and conveys limited knowledge. One way to alleviate this issue is to extract relevant knowledge from external sources at decoding time and incorporate it into the dialog response. In this paper, we propose a post-hoc knowledge-injection technique where we first retrieve a diverse set of relevant knowledge snippets conditioned on both the dialog history and an initial response from an existing dialog model. We construct multiple candidate responses, individually injecting each retrieved snippet into the initial response using a gradient-based decoding method, and then select the final response with an unsupervised ranking step. Our experiments in goal-oriented and knowledge-grounded dialog settings demonstrate that human annotators judge the outputs from the proposed method to be more engaging and informative compared to responses from prior dialog systems. We further show that knowledge-augmentation promotes success in achieving conversational goals in both experimental settings.

High-quality phrase representations are essential to finding topics and related terms in documents (a.k.a. topic mining). Existing phrase representation learning methods either simply combine unigram representations in a context-free manner or rely on extensive annotations to learn context-aware knowledge. In this paper, we propose UCTopic, a novel unsupervised contrastive learning framework for context-aware phrase representations and topic mining. UCTopic is pretrained in a large scale to distinguish if the contexts of two phrase mentions have the same semantics. The key to the pretraining is positive pair construction from our phrase-oriented assumptions. However, we find traditional in-batch negatives cause performance decay when finetuning on a dataset with small topic numbers. Hence, we propose cluster-assisted contrastive learning (CCL) which largely reduces noisy negatives by selecting negatives from clusters and further improves phrase representations for topics accordingly. UCTopic outperforms the state-of-the-art phrase representation model by 38.2% NMI in average on four entity clustering tasks. Comprehensive evaluation on topic mining shows that UCTopic can extract coherent and diverse topical phrases.

We present Knowledge Distillation with Meta Learning (MetaDistil), a simple yet effective alternative to traditional knowledge distillation (KD) methods where the teacher model is fixed during training. We show the teacher network can learn to better transfer knowledge to the student network (i.e., learning to teach) with the feedback from the performance of the distilled student network in a meta learning framework. Moreover, we introduce a pilot update mechanism to improve the alignment between the inner-learner and meta-learner in meta learning algorithms that focus on an improved inner-learner. Experiments on various benchmarks show that MetaDistil can yield significant improvements compared with traditional KD algorithms and is less sensitive to the choice of different student capacity and hyperparameters, facilitating the use of KD on different tasks and models.

In this paper, we propose SkipBERT to accelerate BERT inference by skipping the computation of shallow layers. To achieve this, our approach encodes small text chunks into independent representations, which are then materialized to approximate the shallow representation of BERT. Since the use of such approximation is inexpensive compared with transformer calculations, we leverage it to replace the shallow layers of BERT to skip their runtime overhead. With off-the-shelf early exit mechanisms, we also skip redundant computation from the highest few layers to further improve inference efficiency. Results on GLUE show that our approach can reduce latency by 65% without sacrificing performance. By using only two-layer transformer calculations, we can still maintain 95% accuracy of BERT.

Intermediate-task transfer can benefit a wide range of NLP tasks with properly selected source datasets. However, it is computationally infeasible to experiment with all intermediate transfer combinations, making choosing a useful source task a challenging problem. In this paper, we anticipate that task-specific parameters updated in parameter-efficient tuning methods are likely to encode task-specific information. Therefore, such parameters can be predictive for inter-task transferability. Thus, we propose to exploit these efficiently tuned parameters as off-the-shelf task embeddings for the efficient selection of source datasets for intermediate-task transfer. We experiment with 11 text classification tasks and 11 question answering tasks. Experimental results show that our approach consistently outperforms existing inter-task transferability prediction methods while being conceptually simple and computationally efficient. Our analysis also reveals that the ability of efficiently tuned parameters on transferability prediction is disentangled with their in-task performance. This allows us to use parameters from early checkpoints as task embeddings to further improve efficiency.

Masked language modeling is widely used for pretraining large language models for natural language understanding (NLU). However, random masking is suboptimal, allocating an equal masking rate for all tokens. In this paper, we propose InforMask, a new unsupervised masking strategy for training masked language models. InforMask exploits Pointwise Mutual Information (PMI) to select the most informative tokens to mask. We further propose two optimizations for InforMask to improve its efficiency. With a one-off preprocessing step, InforMask outperforms random masking and previously proposed masking strategies on the factual recall benchmark LAMA and the question answering benchmark SQuAD v1 and v2.

The large population of home cooks with dietary restrictions is under-served by existing cooking resources and recipe generation models. To help them, we propose the task of controllable recipe editing: adapt a base recipe to satisfy a user-specified dietary constraint. This task is challenging, and cannot be adequately solved with human-written ingredient substitution rules or existing end-to-end recipe generation models. We tackle this problem with SHARE: a System for Hierarchical Assistive Recipe Editing, which performs simultaneous ingredient substitution before generating natural-language steps using the edited ingredients. By decoupling ingredient and step editing, our step generator can explicitly integrate the available ingredients. Experiments on the novel RecipePairs dataset—83K pairs of similar recipes where each recipe satisfies one of seven dietary constraints—demonstrate that SHARE produces convincing, coherent recipes that are appropriate for a target dietary constraint. We further show through human evaluations and real-world cooking trials that recipes edited by SHARE can be easily followed by home cooks to create appealing dishes.

In this paper, we propose LaPraDoR, a pretrained dual-tower dense retriever that does not require any supervised data for training. Specifically, we first present Iterative Contrastive Learning (ICoL) that iteratively trains the query and document encoders with a cache mechanism. ICoL not only enlarges the number of negative instances but also keeps representations of cached examples in the same hidden space. We then propose Lexicon-Enhanced Dense Retrieval (LEDR) as a simple yet effective way to enhance dense retrieval with lexical matching. We evaluate LaPraDoR on the recently proposed BEIR benchmark, including 18 datasets of 9 zero-shot text retrieval tasks. Experimental results show that LaPraDoR achieves state-of-the-art performance compared with supervised dense retrieval models, and further analysis reveals the effectiveness of our training strategy and objectives. Compared to re-ranking, our lexicon-enhanced approach can be run in milliseconds (22.5x faster) while achieving superior performance.

Understanding human language often necessitates understanding entities and their place in a taxonomy of knowledge—their types.Previous methods to learn entity types rely on training classifiers on datasets with coarse, noisy, and incomplete labels. We introduce a method to instill fine-grained type knowledge in language models with text-to-text pre-training on type-centric questions leveraging knowledge base documents and knowledge graphs.We create the WikiWiki dataset: entities and passages from 10M Wikipedia articles linked to the Wikidata knowledge graph with 41K types.Models trained on WikiWiki achieve state-of-the-art performance in zero-shot dialog state tracking benchmarks, accurately infer entity types in Wikipedia articles, and can discover new types deemed useful by human judges.

Recent work on reducing bias in NLP models usually focuses on protecting or isolating information related to a sensitive attribute (like gender or race). However, when sensitive information is semantically entangled with the task information of the input, e.g., gender information is predictive for a profession, a fair trade-off between task performance and bias mitigation is difficult to achieve. Existing approaches perform this trade-off by eliminating bias information from the latent space, lacking control over how much bias is necessarily required to be removed. We argue that a favorable debiasing method should use sensitive information ‘fairly’, rather than blindly eliminating it (Caliskan et al., 2017; Sun et al., 2019; Bogen et al., 2020). In this work, we provide a novel debiasing algorithm by adjustingthe predictive model’s belief to (1) ignore the sensitive information if it is not useful for the task; (2) use sensitive information minimally as necessary for the prediction (while also incurring a penalty). Experimental results on two text classification tasks (influenced by gender) and an open-ended generation task (influenced by race) indicate that our model achieves a desirable trade-off between debiasing and task performance along with producing debiased rationales as evidence.

We study the problem of building entity tagging systems by using a few rules as weak supervision. Previous methods mostly focus on disambiguating entity types based on contexts and expert-provided rules, while assuming entity spans are given. In this work, we propose a novel method TALLOR that bootstraps high-quality logical rules to train a neural tagger in a fully automated manner. Specifically, we introduce compound rules that are composed from simple rules to increase the precision of boundary detection and generate more diverse pseudo labels. We further design a dynamic label selection strategy to ensure pseudo label quality and therefore avoid overfitting the neural tagger. Experiments on three datasets demonstrate that our method outperforms other weakly supervised methods and even rivals a state-of-the-art distantly supervised tagger with a lexicon of over 2,000 terms when starting from only 20 simple rules. Our method can serve as a tool for rapidly building taggers in emerging domains and tasks. Case studies show that learned rules can potentially explain the predicted entities.

Humans often refer to personal narratives, life experiences, and events to make a conversation more engaging and rich. While persona-grounded dialog models are able to generate responses that follow a given persona, they often miss out on stating detailed experiences or events related to a persona, often leaving conversations shallow and dull. In this work, we equip dialog models with ‘background stories’ related to a persona by leveraging fictional narratives from existing story datasets (e.g. ROCStories). Since current dialog datasets do not contain such narratives as responses, we perform an unsupervised adaptation of a retrieved story for generating a dialog response using a gradient-based rewriting technique. Our proposed method encourages the generated response to be fluent (i.e., highly likely) with the dialog history, minimally different from the retrieved story to preserve event ordering and consistent with the original persona. We demonstrate that our method can generate responses that are more diverse, and are rated more engaging and human-like by human evaluators, compared to outputs from existing dialog models.

Cant is important for understanding advertising, comedies and dog-whistle politics. However, computational research on cant is hindered by a lack of available datasets. In this paper, we propose a large and diverse Chinese dataset for creating and understanding cant from a computational linguistics perspective. We formulate a task for cant understanding and provide both quantitative and qualitative analysis for tested word embedding similarity and pretrained language models. Experiments suggest that such a task requires deep language understanding, common sense, and world knowledge and thus can be a good testbed for pretrained language models and help models perform better on other tasks.

The ability to generate clarification questions i.e., questions that identify useful missing information in a given context, is important in reducing ambiguity. Humans use previous experience with similar contexts to form a global view and compare it to the given context to ascertain what is missing and what is useful in the context. Inspired by this, we propose a model for clarification question generation where we first identify what is missing by taking a difference between the global and the local view and then train a model to identify what is useful and generate a question about it. Our model outperforms several baselines as judged by both automatic metrics and humans.

Dialog State Tracking (DST), an integral part of modern dialog systems, aims to track user preferences and constraints (slots) in task-oriented dialogs. In real-world settings with constantly changing services, DST systems must generalize to new domains and unseen slot types. Existing methods for DST do not generalize well to new slot names and many require known ontologies of slot types and values for inference. We introduce a novel ontology-free framework that supports natural language queries for unseen constraints and slots in multi-domain task-oriented dialogs. Our approach is based on generative question-answering using a conditional language model pre-trained on substantive English sentences. Our model improves joint goal accuracy in zero-shot domain adaptation settings by up to 9% (absolute) over the previous state-of-the-art on the MultiWOZ 2.1 dataset.

Radiology report generation aims at generating descriptive text from radiology images automatically, which may present an opportunity to improve radiology reporting and interpretation. A typical setting consists of training encoder-decoder models on image-report pairs with a cross entropy loss, which struggles to generate informative sentences for clinical diagnoses since normal findings dominate the datasets. To tackle this challenge and encourage more clinically-accurate text outputs, we propose a novel weakly supervised contrastive loss for medical report generation. Experimental results demonstrate that our method benefits from contrasting target reports with incorrect but semantically-close ones. It outperforms previous work on both clinical correctness and text generation metrics for two public benchmarks.

Written language carries explicit and implicit biases that can distract from meaningful signals. For example, letters of reference may describe male and female candidates differently, or their writing style may indirectly reveal demographic characteristics. At best, such biases distract from the meaningful content of the text; at worst they can lead to unfair outcomes. We investigate the challenge of re-generating input sentences to ‘neutralize’ sensitive attributes while maintaining the semantic meaning of the original text (e.g. is the candidate qualified?). We propose a gradient-based rewriting framework, Detect and Perturb to Neutralize (DEPEN), that first detects sensitive components and masks them for regeneration, then perturbs the generation model at decoding time under a neutralizing constraint that pushes the (predicted) distribution of sensitive attributes towards a uniform distribution. Our experiments in two different scenarios show that DEPEN can regenerate fluent alternatives that are neutral in the sensitive attribute while maintaining the semantics of other attributes.

Recent studies on compression of pretrained language models (e.g., BERT) usually use preserved accuracy as the metric for evaluation. In this paper, we propose two new metrics, label loyalty and probability loyalty that measure how closely a compressed model (i.e., student) mimics the original model (i.e., teacher). We also explore the effect of compression with regard to robustness under adversarial attacks. We benchmark quantization, pruning, knowledge distillation and progressive module replacing with loyalty and robustness. By combining multiple compression techniques, we provide a practical strategy to achieve better accuracy, loyalty and robustness.

Automatic medical image report generation has drawn growing attention due to its potential to alleviate radiologists’ workload. Existing work on report generation often trains encoder-decoder networks to generate complete reports. However, such models are affected by data bias (e.g. label imbalance) and face common issues inherent in text generation models (e.g. repetition). In this work, we focus on reporting abnormal findings on radiology images; instead of training on complete radiology reports, we propose a method to identify abnormal findings from the reports in addition to grouping them with unsupervised clustering and minimal rules. We formulate the task as cross-modal retrieval and propose Conditional Visual-Semantic Embeddings to align images and fine-grained abnormal findings in a joint embedding space. We demonstrate that our method is able to retrieve abnormal findings and outperforms existing generation models on both clinical correctness and text generation metrics.

In this work, we perform the first large-scale analysis of discourse in media dialog and its impact on generative modeling of dialog turns, with a focus on interrogative patterns and use of external knowledge. Discourse analysis can help us understand modes of persuasion, entertainment, and information elicitation in such settings, but has been limited to manual review of small corpora. We introduce **Interview**—a large-scale (105K conversations) media dialog dataset collected from news interview transcripts—which allows us to investigate such patterns at scale. We present a dialog model that leverages external knowledge as well as dialog acts via auxiliary losses and demonstrate that our model quantitatively and qualitatively outperforms strong discourse-agnostic baselines for dialog modeling—generating more specific and topical responses in interview-style conversations.

Existing persona-grounded dialog models often fail to capture simple implications of given persona descriptions, something which humans are able to do seamlessly. For example, state-of-the-art models cannot infer that interest in hiking might imply love for nature or longing for a break. In this paper, we propose to expand available persona sentences using existing commonsense knowledge bases and paraphrasing resources to imbue dialog models with access to an expanded and richer set of persona descriptions. Additionally, we introduce fine-grained grounding on personas by encouraging the model to make a discrete choice among persona sentences while synthesizing a dialog response. Since such a choice is not observed in the data, we model it using a discrete latent random variable and use variational learning to sample from hundreds of persona expansions. Our model outperforms competitive baselines on the Persona-Chat dataset in terms of dialog quality and diversity while achieving persona-consistent and controllable dialog generation.

This paper presents computational approaches for automatically detecting critical plot twists in reviews of media products. First, we created a large-scale book review dataset that includes fine-grained spoiler annotations at the sentence-level, as well as book and (anonymized) user information. Second, we carefully analyzed this dataset, and found that: spoiler language tends to be book-specific; spoiler distributions vary greatly across books and review authors; and spoiler sentences tend to jointly appear in the latter part of reviews. Third, inspired by these findings, we developed an end-to-end neural network architecture to detect spoiler sentences in review corpora. Quantitative and qualitative results demonstrate that the proposed method substantially outperforms existing baselines.

Open-domain question answering remains a challenging task as it requires models that are capable of understanding questions and answers, collecting useful information, and reasoning over evidence. Previous work typically formulates this task as a reading comprehension or entailment problem given evidence retrieved from search engines. However, existing techniques struggle to retrieve indirectly related evidence when no directly related evidence is provided, especially for complex questions where it is hard to parse precisely what the question asks. In this paper we propose a retriever-reader model that learns to attend on essential terms during the question answering process. We build (1) an essential term selector which first identifies the most important words in a question, then reformulates the query and searches for related evidence; and (2) an enhanced reader that distinguishes between essential terms and distracting words to predict the answer. We evaluate our model on multiple open-domain QA datasets, notably achieving the level of the state-of-the-art on the AI2 Reasoning Challenge (ARC) dataset.

Several recent works have considered the problem of generating reviews (or ‘tips’) as a form of explanation as to why a recommendation might match a customer’s interests. While promising, we demonstrate that existing approaches struggle (in terms of both quality and content) to generate justifications that are relevant to users’ decision-making process. We seek to introduce new datasets and methods to address the recommendation justification task. In terms of data, we first propose an ‘extractive’ approach to identify review segments which justify users’ intentions; this approach is then used to distantly label massive review corpora and construct large-scale personalized recommendation justification datasets. In terms of generation, we are able to design two personalized generation models with this data: (1) a reference-based Seq2Seq model with aspect-planning which can generate justifications covering different aspects, and (2) an aspect-conditional masked language model which can generate diverse justifications based on templates extracted from justification histories. We conduct experiments on two real-world datasets which show that our model is capable of generating convincing and diverse justifications.

Existing approaches to dialogue state tracking rely on pre-defined ontologies consisting of a set of all possible slot types and values. Though such approaches exhibit promising performance on single-domain benchmarks, they suffer from computational complexity that increases proportionally to the number of pre-defined slots that need tracking. This issue becomes more severe when it comes to multi-domain dialogues which include larger numbers of slots. In this paper, we investigate how to approach DST using a generation framework without the pre-defined ontology list. Given each turn of user utterance and system response, we directly generate a sequence of belief states by applying a hierarchical encoder-decoder structure. In this way, the computational complexity of our model will be a constant regardless of the number of pre-defined slots. Experiments on both the multi-domain and the single domain dialogue state tracking dataset show that our model not only scales easily with the increasing number of pre-defined domains and slots but also reaches the state-of-the-art performance.

Existing approaches to recipe generation are unable to create recipes for users with culinary preferences but incomplete knowledge of ingredients in specific dishes. We propose a new task of personalized recipe generation to help these users: expanding a name and incomplete ingredient details into complete natural-text instructions aligned with the user’s historical preferences. We attend on technique- and recipe-level representations of a user’s previously consumed recipes, fusing these ‘user-aware’ representations in an attention fusion layer to control recipe text generation. Experiments on a new dataset of 180K recipes and 700K interactions show our model’s ability to generate plausible and personalized recipes compared to non-personalized baselines.

Stories generated with neural language models have shown promise in grammatical and stylistic consistency. However, the generated stories are still lacking in common sense reasoning, e.g., they often contain sentences deprived of world knowledge. We propose a simple multi-task learning scheme to achieve quantitatively better common sense reasoning in language models by leveraging auxiliary training signals from datasets designed to provide common sense grounding. When combined with our two-stage fine-tuning pipeline, our method achieves improved common sense reasoning and state-of-the-art perplexity on the WritingPrompts (Fan et al., 2018) story generation dataset.

In this paper, we focus on the problem of building assistive systems that can help users to write reviews. We cast this problem using an encoder-decoder framework that generates personalized reviews by expanding short phrases (e.g. review summaries, product titles) provided as input to the system. We incorporate aspect-level information via an aspect encoder that learns aspect-aware user and item representations. An attention fusion layer is applied to control generation by attending on the outputs of multiple encoders. Experimental results show that our model successfully learns representations capable of generating coherent and diverse reviews. In addition, the learned aspect-aware representations discover those aspects that users are more inclined to discuss and bias the generated text toward their personalized aspect preferences.

Traditional approaches to recommendation focus on learning from large volumes of historical feedback to estimate simple numerical quantities (Will a user click on a product? Make a purchase? etc.). Natural language approaches that model information like product reviews have proved to be incredibly useful in improving the performance of such methods, as reviews provide valuable auxiliary information that can be used to better estimate latent user preferences and item properties. In this paper, rather than using reviews as an inputs to a recommender system, we focus on generating reviews as the model’s output. This requires us to efficiently model text (at the character level) to capture the preferences of the user, the properties of the item being consumed, and the interaction between them (i.e., the user’s preference). We show that this can model can be used to (a) generate plausible reviews and estimate nuanced reactions; (b) provide personalized rankings of existing reviews; and (c) recommend existing products more effectively.

Unlike Entity Disambiguation in web search results, Opinion Disambiguation is a relatively unexplored topic. RevOpiD shared task at IJCNLP-2107 aimed to attract attention towards this research problem. In this paper, we summarize the first run of this task and introduce a new dataset that we have annotated for the purpose of evaluating Opinion Mining, Summarization and Disambiguation methods.