Recently, there has been increasing interest in using Large Language Models (LLMs) to construct complex multi-agent systems to perform tasks such as compiling literature reviews, drafting consumer reports, and planning vacations. Many tools and libraries exist for helping create such systems, however none support *recursive* multi-agent systems—where the models themselves flexibly decide when to delegate tasks and how to organize their delegation structure. In this work, we introduce ReDel: a toolkit for recursive multi-agent systems that supports custom tool-use, delegation schemes, event-based logging, and interactive replay in an easy-to-use web interface. We show that, using ReDel, we are able to achieve significant performance gains on agentic benchmarks and easily identify potential areas of improvements through the visualization and debugging tools. Our code, documentation, and PyPI package are open-source at https://github.com/zhudotexe/redel, and free to use under the MIT license.
Planning in a text-based environment continues to be a significant challenge for AI systems. Recent approaches have utilized language models to predict planning domain definitions (e.g., PDDL) but have only been evaluated in closed-domain simulated environments. To address this, we present Proc2PDDL, the first dataset containing open-domain procedural texts paired with expert-annotated PDDL representations. Using this dataset, we evaluate the task of predicting domain actions (parameters, preconditions, and effects). We experiment with various large language models (LLMs) and prompting mechanisms, including a novel instruction inspired by the zone of proximal development (ZPD), which reconstructs the task as incremental basic skills. Our results demonstrate that Proc2PDDL is highly challenging for end-to-end LLMs, with GPT-3.5’s success rate close to 0% and GPT-4o’s 38%. With ZPD instructions, GPT-4o’s success rate increases to 45%, outperforming regular chain-of-thought prompting’s 34%. Our analysis systematically examines both syntactic and semantic errors, providing insights into the strengths and weaknesses of language models in generating domain-specific programs.
The goal of text style transfer is to transform the style of texts while preserving their original meaning, often with only a few examples of the target style. Existing style transfer methods generally rely on the few-shot capabilities of large language models or on complex controllable text generation approaches that are inefficient and underperform on fluency metrics. We introduce TinyStyler, a lightweight but effective approach, which leverages a small language model (800M params) and pre-trained authorship embeddings to perform efficient, few-shot text style transfer. We evaluate on the challenging task of authorship style transfer and find TinyStyler outperforms strong approaches such as GPT-4. We also evaluate TinyStyler’s ability to perform text attribute style transfer (formal ↔ informal) with automatic and human evaluations and find that the approach outperforms recent controllable text generation methods.
The proliferation of inflammatory or misleading “fake” news content has become increasingly common in recent years. Simultaneously, it has become easier than ever to use AI tools to generate photorealistic images depicting any scene imaginable. Combining these two—AI-generated fake news content—is particularly potent and dangerous. To combat the spread of AI-generated fake news, we propose the MiRAGeNews Dataset, a dataset of 12,500 high-quality real and AI-generated image-caption pairs from state-of-the-art generators. We find that our dataset poses a significant challenge to humans (60% F-1) and state-of-the-art multi-modal LLMs (< 24% F-1). Using our dataset we train a multi-modal detector (MiRAGe) that improves by +5.1% F-1 over state-of-the-art baselines on image-caption pairs from out-of-domain image generators and news publishers. We release our code and data to aid future work on detecting AI-generated content.
Do the Spratly Islands belong to China, the Philippines, or Vietnam? A pretrained large language model (LLM) may answer differently if asked in the languages of each claimant country: Chinese, Tagalog, or Vietnamese. This contrasts with a multilingual human, who would likely answer consistently. In this paper, we show that LLMs recall certain geographical knowledge inconsistently when queried in different languages—a phenomenon we term geopolitical bias. As a targeted case study, we consider territorial disputes, an inherently controversial and multilingual task. We introduce BorderLines, a dataset of territorial disputes which covers 251 territories, each associated with a set of multiple-choice questions in the languages of each claimant country (49 languages in total). We also propose a suite of evaluation metrics to precisely quantify bias and consistency in responses across different languages. We then evaluate various multilingual LLMs on our dataset and metrics to probe their internal knowledge and use the proposed metrics to discover numerous inconsistencies in how these models respond in different languages. Finally, we explore several prompt modification strategies, aiming to either amplify or mitigate geopolitical bias, which highlights how brittle LLMs are and how they tailor their responses depending on cues from the interaction context. Our code and data are available at https://github.com/manestay/borderlines.
Large language models excel at creative generation but continue to struggle with the issues of hallucination and bias. While retrieval-augmented generation (RAG) provides a framework for grounding LLMs’ responses in accurate and up-to-date information, it still raises the question of bias: which sources should be selected for inclusion in the context? And how should their importance be weighted? In this paper, we study the challenge of cross-lingual RAG and present a dataset to investigate the robustness of existing systems at answering queries about geopolitical disputes, which exist at the intersection of linguistic, cultural, and political boundaries. Our dataset is sourced from Wikipedia pages containing information relevant to the given queries and we investigate the impact of including additional context, as well as the composition of this context in terms of language and source, on an LLM’s response. Our results show that existing RAG systems continue to be challenged by cross-lingual use cases and suffer from a lack of consistency when they are provided with competing information in multiple languages. We present case studies to illustrate these issues and outline steps for future research to address these challenges.
We study how differences in persuasive language across Wikipedia articles, written in either English and Russian, can uncover each culture’s distinct perspective on different subjects. We develop a large language model (LLM) powered system to identify instances of persuasive language in multilingual texts. Instead of directly prompting LLMs to detect persuasion, which is subjective and difficult, we propose to reframe the task to instead ask high-level questions (HLQs) which capture different persuasive aspects. Importantly, these HLQs are authored by LLMs themselves. LLMs over-generate a large set of HLQs, which are subsequently filtered to a small set aligned with human labels for the original task. We then apply our approach to a large-scale, bilingual dataset of Wikipedia articles (88K total), using a two-stage identify-then-extract prompting strategy to find instances of persuasion. We quantify the amount of persuasion per article, and explore the differences in persuasion through several experiments on the paired articles. Notably, we generate rankings of articles by persuasion in both languages. These rankings match our intuitions on the culturally-salient subjects; Russian Wikipedia highlights subjects on Ukraine, while English Wikipedia highlights the Middle East. Grouping subjects into larger topics, we find politically-related events contain more persuasion than others. We further demonstrate that HLQs obtain similar performance when posed in either English or Russian. Our methodology enables cross-lingual, cross-cultural understanding at scale, and we release our code, prompts, and data.
End-to-end neural Natural Language Processing (NLP) models are notoriously difficult to understand. This has given rise to numerous efforts towards model explainability in recent years. One desideratum of model explanation is faithfulness, that is, an explanation should accurately represent the reasoning process behind the model’s prediction. In this survey, we review over 110 model explanation methods in NLP through the lens of faithfulness. We first discuss the definition and evaluation of faithfulness, as well as its significance for explainability. We then introduce recent advances in faithful explanation, grouping existing approaches into five categories: similarity-based methods, analysis of model-internal structures, backpropagation-based methods, counterfactual intervention, and self-explanatory models. For each category, we synthesize its representative studies, strengths, and weaknesses. Finally, we summarize their common virtues and remaining challenges, and reflect on future work directions towards faithful explainability in NLP.
Planning in textual environments have been shown to be a long-standing challenge even for current models. A recent, promising line of work uses LLMs to generate a formal representation of the environment that can be solved by a symbolic planner. However, existing methods rely on a fully-observed environment where all entity states are initially known, so a one-off representation can be constructed, leading to a complete plan. In contrast, we tackle partially-observed environments where there is initially no sufficient information to plan for the end-goal. We propose PDDLEGO that iteratively construct a planning representation that can lead to a partial plan for a given sub-goal. By accomplishing the sub-goal, more information is acquired to augment the representation, eventually achieving the end-goal. We show that plans produced by few-shot PDDLEGO are 43% more efficient than generating plans end-to-end on the Coin Collector simulation, with strong performance (98%) on the more complex Cooking World simulation where end-to-end LLMs fail to generate coherent plans (4%).
Much texts describe a changing world (e.g., procedures, stories, newswires), and understanding them requires tracking how entities change. An earlier dataset, OpenPI, provided crowdsourced annotations of entity state changes in text. However, a major limitation was that those annotations were free-form and did not identify salient changes, hampering model evaluation. To overcome these limitations, we present an improved dataset, OpenPI2.0, where entities and attributes are fully canonicalized and additional entity salience annotations are added. On our fairer evaluation setting, we find that current state-of-the-art language models are far from competent. We also show that using state changes of salient entities as a chain-of-thought prompt, downstream performance is improved on tasks such as question answering and classical planning, outperforming the setting involving all related entities indiscriminately. We offer OpenPI2.0 for the continued development of models that can understand the dynamics of entities in text.
Bistable images, also known as ambiguous or reversible images, present visual stimuli that can be seen in two distinct interpretations, though not simultaneously, by the observer. In this study, we conduct the most extensive examination of vision-language models using bistable images to date. We manually gathered a dataset of 29 bistable images, along with their associated labels, and subjected them to 121 different manipulations in brightness, resolution, tint, and rotation. We evaluated twelve different models in both classification and generative tasks across six model architectures. Our findings reveal that, with the exception of models from the Idefics family and LLaVA1.5-13b, there is a pronounced preference for one interpretation over another among the models, and minimal variance under image manipulations, with few exceptions on image rotations. Additionally, we compared the models’ preferences with humans, noting that the models do not exhibit the same continuity biases as humans and often diverge from human initial interpretations. We also investigated the influence of variations in prompts and the use of synonymous labels, discovering that these factors significantly affect model interpretations more than image manipulations showing a higher influence of the language priors on bistable image interpretations compared to image-text training data. All code and data is open sourced.
Large language models (LLMs) have become a dominant and important tool for NLP researchers in a wide range of tasks. Today, many researchers use LLMs in synthetic data generation, task evaluation, fine-tuning, distillation, and other model-in-the-loop research workflows. However, challenges arise when using these models that stem from their scale, their closed source nature, and the lack of standardized tooling for these new and emerging workflows. The rapid rise to prominence of these models and these unique challenges has had immediate adverse impacts on open science and on the reproducibility of work that uses them. In this ACL 2024 theme track paper, we introduce DataDreamer, an open source Python library that allows researchers to write simple code to implement powerful LLM workflows. DataDreamer also helps researchers adhere to best practices that we propose to encourage open science and reproducibility. The library and documentation are available at: https://github.com/datadreamer-dev/DataDreamer.
Many commercial and open-source models claim to detect machine-generated text with extremely high accuracy (99% or more). However, very few of these detectors are evaluated on shared benchmark datasets and even when they are, the datasets used for evaluation are insufficiently challenging—lacking variations in sampling strategy, adversarial attacks, and open-source generative models. In this work we present RAID: the largest and most challenging benchmark dataset for machine-generated text detection. RAID includes over 6 million generations spanning 11 models, 8 domains, 11 adversarial attacks and 4 decoding strategies. Using RAID, we evaluate the out-of-domain and adversarial robustness of 8 open- and 4 closed-source detectors and find that current detectors are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models. We release our data along with a leaderboard to encourage future research.
One type of question that is commonly found in day-to-day scenarios is “fan-out” questions, complex multi-hop, multi-document reasoning questions that require finding information about a large number of entities. However, there exist few resources to evaluate this type of question-answering capability among large language models. To evaluate complex reasoning in LLMs more fully, we present FanOutQA, a high-quality dataset of fan-out question-answer pairs and human-annotated decompositions with English Wikipedia as the knowledge base. We formulate three benchmark settings across our dataset and benchmark 7 LLMs, including GPT-4, LLaMA 2, Claude-2.1, and Mixtral-8x7B, finding that contemporary models still have room to improve reasoning over inter-document dependencies in a long context. We provide our dataset, along with open-source tools to run models to encourage evaluation.
Script learning studies how daily events unfold. It enables machines to reason about narratives with implicit information. Previous works mainly consider a script as a linear sequence of events while ignoring the potential branches that arise due to people’s circumstantial choices. We hence propose Choice-75, the first benchmark that challenges intelligent systems to make decisions given descriptive scenarios, containing 75 scripts and more than 600 scenarios. We also present preliminary results with current large language models (LLM). Although they demonstrate overall decent performances, there is still notable headroom in hard scenarios.
Dungeons & Dragons (D&D) is a tabletop roleplaying game with complex natural language interactions between players and hidden state information. Recent work has shown that large language models (LLMs) that have access to state information can generate higher quality game turns than LLMs that use dialog history alone. However, previous work used game state information that was heuristically created and was not a true gold standard game state. We present FIREBALL, a large dataset containing nearly 25,000 unique sessions from real D&D gameplay on Discord with true game state info. We recorded game play sessions of players who used the Avrae bot, which was developed to aid people in playing D&D online, capturing language, game commands and underlying game state information. We demonstrate that FIREBALL can improve natural language generation (NLG) by using Avrae state information, improving both automated metrics and human judgments of quality. Additionally, we show that LLMs can generate executable Avrae commands, particularly after finetuning.
Large Language Models (LLMs) are so powerful that they sometimes learn correlations between labels and features that are irrelevant to the task, leading to poor generalization on out-of-distribution data. We propose explanation-based finetuning as a general approach to mitigate LLMs’ reliance on spurious correlations. Unlike standard finetuning where the model only predicts the answer given the input, we finetune the model to additionally generate a free-text explanation supporting its answer. To evaluate our method, we finetune the model on artificially constructed training sets containing different types of spurious cues, and test it on a test set without these cues. Compared to standard finetuning, our method makes GPT-3 (davinci) remarkably more robust against spurious cues in terms of accuracy drop across four classification tasks: ComVE (+1.2), CREAK (+9.1), e-SNLI (+15.4), and SBIC (+6.5). The efficacy generalizes across multiple model families and scales, with greater gains for larger models. Finally, our method also works well with explanations generated by the model, implying its applicability to more datasets without human-written explanations.
Event schemas are a form of world knowledge about the typical progression of events. Recent methods for event schema induction use information extraction systems to construct a large number of event graph instances from documents, and then learn to generalize the schema from such instances. In contrast, we propose to treat event schemas as a form of commonsense knowledge that can be derived from large language models (LLMs). This new paradigm greatly simplifies the schema induction process and allows us to handle both hierarchical relations and temporal relations between events in a straightforward way. Since event schemas have complex graph structures, we design an incremental prompting and verification method IncPrompt to break down the construction of a complex event graph into three stages: event skeleton construction, event expansion, and event-event relation verification. Compared to directly using LLMs to generate a linearized graph, IncSchema can generate large and complex schemas with 7.2% F1 improvement in temporal relations and 31.0% F1 improvement in hierarchical relations. In addition, compared to the previous state-of-the-art closed-domain schema induction model, human assessors were able to cover ~10% more events when translating the schemas into coherent stories and rated our schemas 1.3 points higher (on a 5-point scale) in terms of readability.
We propose a novel task, G4C, to study teacher-student natural language interactions in a goal-driven and grounded environment. Dungeons and Dragons (D&D), a role-playing game, provides an ideal setting to investigate such interactions. Here, the Dungeon Master (DM), i.e., the teacher, guides the actions of several players—students, each with their own personas and abilities—to achieve shared goals grounded in a fantasy world. Our approach is to decompose and model these interactions into (1) the DM’s intent to guide players toward a given goal; (2) the DM’s guidance utterance to the players expressing this intent; and (3) a theory-of-mind (ToM) model that anticipates the players’ reaction to the guidance one turn into the future. We develop a novel reinforcement learning (RL) method for training a DM that generates guidance for players by rewarding utterances where the intent matches the ToM-anticipated player actions. Human and automated evaluations show that a DM trained to explicitly model intents and incorporate ToM of the players using RL generates better-quality guidance that is 3x more likely to fulfill the DM’s intent than a vanilla natural language generation (NLG) approach.
Schema induction builds a graph representation explaining how events unfold in a scenario. Existing approaches have been based on information retrieval (IR) and information extraction (IE), often with limited human curation. We demonstrate a human-in-the-loop schema induction system powered by GPT-3. We first describe the different modules of our system, including prompting to generate schematic elements, manual edit of those elements, and conversion of those into a schema graph. By qualitatively comparing our system to previous ones, we show that our system not only transfers to new domains more easily than previous approaches, but also reduces efforts of human curation thanks to our interactive interface.
Entities and events are crucial to natural language reasoning and common in procedural texts. Existing work has focused either exclusively on entity state tracking (e.g., whether a pan is hot) or on event reasoning (e.g., whether one would burn themselves by touching the pan), while these two tasks are often causally related. We propose CREPE, the first benchmark on causal reasoning of event plausibility and entity states. We show that most language models, including GPT-3, perform close to chance at .35 F1, lagging far behind human at .87 F1. We boost model performance to .59 F1 by creatively representing events as programming languages while prompting language models pretrained on code. By injecting the causal relations between entities and events as intermediate reasoning steps in our representation, we further boost the performance to .67 F1. Our findings indicate not only the challenge that CREPE brings for language models, but also the efficacy of code-like prompting combined with chain-of-thought prompting for multihop event reasoning.
Story generation and understanding—as with all NLG/NLU tasks—has seen a surge in neurosymbolic work. Researchers have recognized that, while large language models (LLMs) have tremendous utility, they can be augmented with symbolic means to be even better and to make up for many flaws that neural networks have. However, symbolic methods are extremely costly in terms of the amount of time and expertise needed to create them. In this work, we capitalize on state-of-the-art Code-LLMs, such as Codex, to bootstrap the use of symbolic methods for tracking the state of stories and aiding in story understanding. We show that our CoRRPUS system and abstracted prompting procedures can beat current state-of-the-art structured LLM techniques on pre-existing story understanding tasks (bAbI Task 2 and Re³) with minimal hand engineering. This work highlights the usefulness of code-based symbolic representations for enabling LLMs to better perform story reasoning tasks.
Existing question answering (QA) systems owe much of their success to large, high-quality training data. Such annotation efforts are costly, and the difficulty compounds in the cross-lingual setting. Therefore, prior cross-lingual QA work has focused on releasing evaluation datasets, and then applying zero-shot methods as baselines. This work proposes a synthetic data generation method for cross-lingual QA which leverages indirect supervision from existing parallel corpora. Our method termed PAXQA (Projecting annotations for cross-lingual (x) QA) decomposes cross-lingual QA into two stages. First, we apply a question generation (QG) model to the English side. Second, we apply annotation projection to translate both the questions and answers. To better translate questions, we propose a novel use of lexically-constrained machine translation, in which constrained entities are extracted from the parallel bitexts. We apply PAXQA to generate cross-lingual QA examples in 4 languages (662K examples total), and perform human evaluation on a subset to create validation and test splits. We then show that models fine-tuned on these datasets outperform prior synthetic data generation models over several extractive QA datasets. The largest performance gains are for directions with non-English questions and English contexts. Ablation studies show that our dataset generation method is relatively robust to noise from automatic word alignments, showing the sufficient quality of our generations. To facilitate follow-up work, we release our code and datasets at https://github.com/manestay/paxqa.
Style representation learning builds content-independent representations of author style in text. To date, no large dataset of texts with stylometric annotations on a wide range of style dimensions has been compiled, perhaps because the linguistic expertise to perform such annotation would be prohibitively expensive. Therefore, current style representation approaches make use of unsupervised neural methods to disentangle style from content to create style vectors. These approaches, however, result in uninterpretable representations, complicating their usage in downstream applications like authorship attribution where auditing and explainability is critical. In this work, we use prompting to perform stylometry on a large number of texts to generate a synthetic stylometry dataset. We use this synthetic data to then train human-interpretable style representations we call LISA embeddings. We release our synthetic dataset (StyleGenome) and our interpretable style embedding model (LISA) as resources.
In this paper, we propose Learn, a unified, easy-to-use tool to apply question generation and selection in classrooms. The tool lets instructors and TAs create assignments that can write and re-write themselves. Given existing course materials, for example a reference textbook, Learn can generate questions, select the highest quality questions, show the questions to students, adapt question difficulty to student knowledge, and generate new questions based on how effectively old questions help students learn. The modular, composable nature of the tools for handling each sub-task allow instructors to use only the parts of the tool necessary to the course, allowing for integration in a large number of courses with varied teaching styles. We also report on the adoption of the tool in classes at the University of Pennsylvania with over 1000 students. Learn is publicly released at https://learn.withmartian.com.
Language model applications are becoming increasingly popular and complex, often including features like tool usage and retrieval augmentation. However, existing frameworks for such applications are often opinionated, deciding for developers how their prompts ought to be formatted and imposing limitations on customizability and reproducibility. To solve this we present Kani: a lightweight, flexible, and model-agnostic open-source framework for building language model applications. Kani helps developers implement a variety of complex features by supporting the core building blocks of chat interaction: model interfacing, chat management, and robust function calling. All Kani core functions are easily overridable and well documented to empower developers to customize functionality for their own needs. Kani thus serves as a useful tool for researchers, hobbyists, and industry professionals alike to accelerate their development while retaining interoperability and fine-grained control.
Recent work has shown that prompting language models with code-like representations of natural language leads to performance improvements on structured reasoning tasks. However, such tasks comprise only a small subset of all natural language tasks. In our work, we seek to answer whether or not code-prompting is the preferred way of interacting with language models in general. We compare code and text prompts across three popular GPT models (davinci, code-davinci-002, and text-davinci-002) on a broader selection of tasks (e.g., QA, sentiment, summarization) and find that with few exceptions, code prompts do not consistently outperform text prompts. Furthermore, we show that the style of code prompt has a large effect on performance for some (but not all) tasks and that fine-tuning on text instructions leads to better relative performance of code prompts.
The representation space of pretrained Language Models (LMs) encodes rich information about words and their relationships (e.g., similarity, hypernymy, polysemy) as well as abstract semantic notions (e.g., intensity). In this paper, we demonstrate that lexical stylistic notions such as complexity, formality, and figurativeness, can also be identified in this space. We show that it is possible to derive a vector representation for each of these stylistic notions from only a small number of seed pairs. Using these vectors, we can characterize new texts in terms of these dimensions by performing simple calculations in the corresponding embedding space. We conduct experiments on five datasets and find that static embeddings encode these features more accurately at the level of words and phrases, whereas contextualized LMs perform better on sentences. The lower performance of contextualized representations at the word level is partially attributable to the anisotropy of their vector space, which can be corrected to some extent using techniques like standardization.
Large language models can solve reasoning tasks (like math problems) more effectively when they are allowed to generate rationales. However, a good tutoring system should not just generate solutions, but should also generate explanations and should be able to correct and guide students. We show that providing a code scratchpad improves performance on each tutoring step with a gradeschool mathematics dataset. On these tutoring tasks, GPT-3 models provided with a code scratchpad significantly outperform those given only a language scratchpad (77.7% vs 48.7% cumulative accuracy).
We address the problem of generating high-quality question-answer pairs for educational materials. Previous work on this problem showed that using summaries as input improves the quality of question generation (QG) over original textbook text and that human-written summaries result in higher quality QG than automatic summaries. In this paper, a) we show that advances in Large Language Models (LLMs) are not yet sufficient to generate quality summaries for QG and b) we introduce a new methodology for enhancing bullet point student notes into fully fledged summaries and find that our methodology yields higher quality QG. We conducted a large-scale human annotation study of generated question-answer pairs for the evaluation of our methodology. In order to aid in future research, we release a new dataset of 9.2K human annotations of generated questions.
We introduce a novel technique for automatically summarizing lecture videos using large language models such as GPT-3 and we present a user study investigating the effects on the studying experience when automatic summaries are added to lecture videos. We test students under different conditions and find that the students who are shown a summary next to a lecture video perform better on quizzes designed to test the course materials than the students who have access only to the video or the summary. Our findings suggest that adding automatic summaries to lecture videos enhances the learning experience. Qualitatively, students preferred summaries when studying under time constraints.
We propose a two-stage approach for training a single NMT model to translate unseen languages both to and from English. For the first stage, we initialize an encoder-decoder model to pretrained XLM-R and RoBERTa weights, then perform multilingual fine-tuning on parallel data in 40 languages to English. We find this model can generalize to zero-shot translations on unseen languages. For the second stage, we leverage this generalization ability to generate synthetic parallel data from monolingual datasets, then bidirectionally train with successive rounds of back-translation. Our approach, which we EcXTra (uE/unglish-uc/uentric Crosslingual (uX/u) uTra/unsfer), is conceptually simple, only using a standard cross-entropy objective throughout. It is also data-driven, sequentially leveraging auxiliary parallel data and monolingual data. We evaluate unsupervised NMT results for 7 low-resource languages, and find that each round of back-translation training further refines bidirectional performance. Our final single EcXTra-trained model achieves competitive translation performance in all translation directions, notably establishing a new state-of-the-art for English-to-Kazakh (22.9 10.4 BLEU). Our code is available at [this URL](https://github.com/manestay/EcXTra).
Recursive noun phrases (NPs) have interesting semantic properties. For example, “my favorite new movie” is not necessarily my favorite movie, whereas “my new favorite movie” is. This is common sense to humans, yet it is unknown whether language models have such knowledge. We introduce the Recursive Noun Phrase Challenge (RNPC), a dataset of three textual inference tasks involving textual entailment and event plausibility comparison, precisely targeting the understanding of recursive NPs. When evaluated on RNPC, state-of-the-art Transformer models only perform around chance. Still, we show that such knowledge is learnable with appropriate data. We further probe the models for relevant linguistic features that can be learned from our tasks, including modifier semantic category and modifier scope. Finally, models trained on RNPC achieve strong zero-shot performance on an extrinsic Harm Detection evaluation task, showing the usefulness of the understanding of recursive NPs in downstream applications.
We introduce RESIN-11, a new schema-guided event extraction&prediction framework that can be applied to a large variety of newsworthy scenarios. The framework consists of two parts: (1) an open-domain end-to-end multimedia multilingual information extraction system with weak-supervision and zero-shot learningbased techniques. (2) schema matching and schema-guided event prediction based on our curated schema library. We build a demo website based on our dockerized system and schema library publicly available for installation (https://github.com/RESIN-KAIROS/RESIN-11). We also include a video demonstrating the system.
Procedures are inherently hierarchical. To “make videos”, one may need to “purchase a camera”, which in turn may require one to “set a budget”. While such hierarchical knowledge is critical for reasoning about complex procedures, most existing work has treated procedures as shallow structures without modeling the parent-child relation. In this work, we attempt to construct an open-domain hierarchical knowledge-base (KB) of procedures based on wikiHow, a website containing more than 110k instructional articles, each documenting the steps to carry out a complex procedure. To this end, we develop a simple and efficient method that links steps (e.g., “purchase a camera”) in an article to other articles with similar goals (e.g., “how to choose a camera”), recursively constructing the KB. Our method significantly outperforms several strong baselines according to automatic evaluation, human judgment, and application to downstream tasks such as instructional video retrieval.
We find that existing language modeling datasets contain many near-duplicate examples and long repetitive substrings. As a result, over 1% of the unprompted output of language models trained on these datasets is copied verbatim from the training data. We develop two tools that allow us to deduplicate training datasets—for example removing from C4 a single 61 word English sentence that is repeated over 60,000 times. Deduplication allows us to train models that emit memorized text ten times less frequently and require fewer training steps to achieve the same or better accuracy. We can also reduce train-test overlap, which affects over 4% of the validation set of standard datasets, thus allowing for more accurate evaluation. Code for deduplication is released at https://github.com/google-research/deduplicate-text-datasets.
In this paper, we leverage large language models (LLMs) to perform zero-shot text style transfer. We present a prompting method that we call augmented zero-shot learning, which frames style transfer as a sentence rewriting task and requires only a natural language instruction, without model fine-tuning or exemplars in the target style. Augmented zero-shot learning is simple and demonstrates promising results not just on standard style transfer tasks such as sentiment, but also on arbitrary transformations such as ‘make this melodramatic’ or ‘insert a metaphor.’
This paper provides an overview and update on the Linguistic Data Consortium’s (LDC) NIEUW (Novel Incentives and Workflows) project supported by the National Science Foundation and part of LDC’s larger goal of improving the cost, variety, scale, and quality of language resources available for education, research, and technology development. NIEUW leverages the power of novel incentives to elicit linguistic data and annotations from a wide variety of contributors including citizen scientists, game players, and language students and professionals. In order to align appropriate incentives with the various contributors, LDC has created three distinct web portals to bring together researchers and other language professionals with participants best suited to their project needs. These portals include LanguageARC designed for citizen scientists, Machina Pro Linguistica designed for students and language professionals, and LingoBoingo designed for game players. The design, interface, and underlying tools for each web portal were developed to appeal to the different incentives and motivations of their respective target audiences.
AI researchers have posited Dungeons and Dragons (D&D) as a challenge problem to test systems on various language-related capabilities. In this paper, we frame D&D specifically as a dialogue system challenge, where the tasks are to both generate the next conversational turn in the game and predict the state of the game given the dialogue history. We create a gameplay dataset consisting of nearly 900 games, with a total of 7,000 players, 800,000 dialogue turns, 500,000 dice rolls, and 58 million words. We automatically annotate the data with partial state information about the game play. We train a large language model (LM) to generate the next game turn, conditioning it on different information. The LM can respond as a particular character or as the player who runs the game—i.e., the Dungeon Master (DM). It is trained to produce dialogue that is either in-character (roleplaying in the fictional world) or out-of-character (discussing rules or strategy). We perform a human evaluation to determine what factors make the generated output plausible and interesting. We further perform an automatic evaluation to determine how well the model can predict the game state given the history and examine how well tracking the game state improves its ability to produce plausible conversational output.
Entity linking, the task of linking potentially ambiguous mentions in texts to corresponding knowledge-base entities, is an important component for language understanding. We address two challenge in entity linking: how to leverage wider contexts surrounding a mention, and how to deal with limited training data. We propose a fully unsupervised model called SumMC that first generates a guided summary of the contexts conditioning on the mention, and then casts the task to a multiple-choice problem where the model chooses an entity from a list of candidates. In addition to evaluating our model on existing datasets that focus on named entities, we create a new dataset that links noun phrases from WikiHow to Wikidata. We show that our SumMC model achieves state-of-the-art unsupervised performance on our new dataset and on exiting datasets.
We conduct a feasibility study into the applicability of answer-agnostic question generation models to textbook passages. We show that a significant portion of errors in such systems arise from asking irrelevant or un-interpretable questions and that such errors can be ameliorated by providing summarized input. We find that giving these models human-written summaries instead of the original text results in a significant increase in acceptability of generated questions (33% → 83%) as determined by expert annotators. We also find that, in the absence of human-written summaries, automatic summarization can serve as a good middle ground.
The task of inserting text into a specified position in a passage, known as fill in the blank (FitB), is useful for a variety of applications where writers interact with a natural language generation (NLG) system to craft text. While previous work has tackled this problem with models trained specifically to do fill in the blank, a more useful model is one that can effectively perform _both_ FitB and continuation tasks. In this work, we evaluate the feasibility of using a single model to do both tasks. We show that models pre-trained with a FitB-style objective are capable of both tasks, while models pre-trained for continuation are not. Finally, we show how these models can be easily finetuned to allow for fine-grained control over the length and word choice of the generation.
Neural language models encode rich knowledge about entities and their relationships which can be extracted from their representations using probing. Common properties of nouns (e.g., red strawberries, small ant) are, however, more challenging to extract compared to other types of knowledge because they are rarely explicitly stated in texts.We hypothesize this to mainly be the case for perceptual properties which are obvious to the participants in the communication. We propose to extract these properties from images and use them in an ensemble model, in order to complement the information that is extracted from language models. We consider perceptual properties to be more concrete than abstract properties (e.g., interesting, flawless). We propose to use the adjectives’ concreteness score as a lever to calibrate the contribution of each source (text vs. images). We evaluate our ensemble model in a ranking task where the actual properties of a noun need to be ranked higher than other non-relevant properties. Our results show that the proposed combination of text and images greatly improves noun property prediction compared to powerful text-based language models.
While popular Television (TV) shows are airing, some users interested in these shows publish social media posts about the show. Analyzing social media posts related to a TV show can be beneficial for gaining insights about what happened during scenes of the show. This is a challenging task partly because a significant number of social media posts associated with a TV show or event may not clearly describe what happened during the event. In this work, we propose a method to predict social media posts (associated with scenes of a TV show) that are indicative of what transpired during the scenes of the show. We evaluate our method on social media (Twitter) posts associated with an episode of a popular TV show, Game of Thrones. We show that for each of the identified scenes, with high AUC’s, our method can predict posts that are indicative of what happened in a scene from those that are not-indicative. Based on Twitters policy, we will make the Tweeter ID’s of the Twitter posts used for this work publicly available.
Neural Machine Translation (NMT) models have been observed to produce poor translations when there are few/no parallel sentences to train the models. In the absence of parallel data, several approaches have turned to the use of images to learn translations. Since images of words, e.g., horse may be unchanged across languages, translations can be identified via images associated with words in different languages that have a high degree of visual similarity. However, translating via images has been shown to improve upon text-only models only marginally. To better understand when images are useful for translation, we study image translatability of words, which we define as the translatability of words via images, by measuring intra- and inter-cluster similarities of image representations of words that are translations of each other. We find that images of words are not always invariant across languages, and that language pairs with shared culture, meaning having either a common language family, ethnicity or religion, have improved image translatability (i.e., have more similar images for similar words) compared to its converse, regardless of their geographic proximity. In addition, in line with previous works that show images help more in translating concrete words, we found that concrete words have improved image translatability compared to abstract ones.
We present a new information extraction system that can automatically construct temporal event graphs from a collection of news documents from multiple sources, multiple languages (English and Spanish for our experiment), and multiple data modalities (speech, text, image and video). The system advances state-of-the-art from two aspects: (1) extending from sentence-level event extraction to cross-document cross-lingual cross-media event extraction, coreference resolution and temporal event tracking; (2) using human curated event schema library to match and enhance the extraction output. We have made the dockerlized system publicly available for research purpose at GitHub, with a demo video.
The knowledge of scripts, common chains of events in stereotypical scenarios, is a valuable asset for task-oriented natural language understanding systems. We propose the Goal-Oriented Script Construction task, where a model produces a sequence of steps to accomplish a given goal. We pilot our task on the first multilingual script learning dataset supporting 18 languages collected from wikiHow, a website containing half a million how-to articles. For baselines, we consider both a generation-based approach using a language model and a retrieval-based approach by first retrieving the relevant steps from a large candidate pool and then ordering them. We show that our task is practical, feasible but challenging for state-of-the-art Transformer models, and that our methods can be readily deployed for various other datasets and domains with decent zero-shot performance.
While day-to-day questions come with a variety of answer types, the current question-answering (QA) literature has failed to adequately address the answer diversity of questions. To this end, we present GooAQ, a large-scale dataset with a variety of answer types. This dataset contains over 5 million questions and 3 million answers collected from Google. GooAQ questions are collected semi-automatically from the Google search engine using its autocomplete feature. This results in naturalistic questions of practical interest that are nonetheless short and expressed using simple language. GooAQ answers are mined from Google’s responses to our collected questions, specifically from the answer boxes in the search results. This yields a rich space of answer types, containing both textual answers (short and long) as well as more structured ones such as collections. We benchmark T5 models on GooAQ and observe that: (a) in line with recent work, LM’s strong performance on GooAQ’s short-answer questions heavily benefit from annotated data; however, (b) their quality in generating coherent and accurate responses for questions requiring long responses (such as ‘how’ and ‘why’ questions) is less reliant on observing annotated data and mainly supported by their pre-training. We release GooAQ to facilitate further research on improving QA with diverse response types.
We present a simple but effective approach for leveraging Wikipedia for neural machine translation as well as cross-lingual tasks of image captioning and dependency parsing without using any direct supervision from external parallel data or supervised models in the target language. We show that first sentences and titles of linked Wikipedia pages, as well as cross-lingual image captions, are strong signals for a seed parallel data to extract bilingual dictionaries and cross-lingual word embeddings for mining parallel text from Wikipedia. Our final model achieves high BLEU scores that are close to or sometimes higher than strong supervised baselines in low-resource languages; e.g. supervised BLEU of 4.0 versus 12.1 from our model in English-to-Kazakh. Moreover, we tailor our wikily translation models to unsupervised image captioning, and cross-lingual dependency parser transfer. In image captioning, we train a multi-tasking machine translation and image captioning pipeline for Arabic and English from which the Arabic training data is a wikily translation of the English captioning data. Our captioning results on Arabic are slightly better than that of its supervised model. In dependency parsing, we translate a large amount of monolingual text, and use it as an artificial training data in an annotation projection framework. We show that our model outperforms recent work on cross-lingual transfer of dependency parsers.
Understanding what sequence of steps are needed to complete a goal can help artificial intelligence systems reason about human activities. Past work in NLP has examined the task of goal-step inference for text. We introduce the visual analogue. We propose the Visual Goal-Step Inference (VGSI) task, where a model is given a textual goal and must choose which of four images represents a plausible step towards that goal. With a new dataset harvested from wikiHow consisting of 772,277 images representing human actions, we show that our task is challenging for state-of-the-art multimodal models. Moreover, the multimodal representation learned from our data can be effectively transferred to other datasets like HowTo100m, increasing the VGSI accuracy by 15 - 20%. Our task will facilitate multimodal reasoning about procedural events.
An important task in NLP applications such as sentence simplification is the ability to take a long, complex sentence and split it into shorter sentences, rephrasing as necessary. We introduce a novel dataset and a new model for this ‘split and rephrase’ task. Our BiSECT training data consists of 1 million long English sentences paired with shorter, meaning-equivalent English sentences. We obtain these by extracting 1-2 sentence alignments in bilingual parallel corpora and then using machine translation to convert both sides of the corpus into the same language. BiSECT contains higher quality training examples than the previous Split and Rephrase corpora, with sentence splits that require more significant modifications. We categorize examples in our corpus and use these categories in a novel model that allows us to target specific regions of the input sentence to be split and edited. Moreover, we show that models trained on BiSECT can perform a wider variety of split operations and improve upon previous state-of-the-art approaches in automatic and human evaluations.
Acquiring training data for natural language processing systems can be expensive and time-consuming. Given a few training examples crafted by experts, large corpora can be mined for thousands of semantically similar examples that provide useful variability to improve model generalization. We present TopGuNN, a fast contextualized k-NN retrieval system that can efficiently index and search over contextual embeddings generated from large corpora. TopGuNN is demonstrated for a training data augmentation use case over the Gigaword corpus. Using approximate k-NN and an efficient architecture, TopGuNN performs queries over an embedding space of 4.63TB (approximately 1.5B embeddings) in less than a day.
Many people live-tweet televised events like Presidential debates and popular TV-shows and discuss people or characters in the event. Naturally, many tweets make pronominal reference to these people/characters. We propose an algorithm for resolving personal pronouns that make reference to people involved in an event, in tweet streams collected during the event.
Modern task-oriented dialog systems need to reliably understand users’ intents. Intent detection is even more challenging when moving to new domains or new languages, since there is little annotated data. To address this challenge, we present a suite of pretrained intent detection models which can predict a broad range of intended goals from many actions because they are trained on wikiHow, a comprehensive instructional website. Our models achieve state-of-the-art results on the Snips dataset, the Schema-Guided Dialogue dataset, and all 3 languages of the Facebook multilingual dialog datasets. Our models also demonstrate strong zero- and few-shot performance, reaching over 75% accuracy using only 100 training examples in all datasets.
Recent advancements in neural language modelling make it possible to rapidly generate vast amounts of human-sounding text. The capabilities of humans and automatic discriminators to detect machine-generated text have been a large source of research interest, but humans and machines rely on different cues to make their decisions. Here, we perform careful benchmarking and analysis of three popular sampling-based decoding strategies—top-_k_, nucleus sampling, and untruncated random sampling—and show that improvements in decoding methods have primarily optimized for fooling humans. This comes at the expense of introducing statistical abnormalities that make detection easy for automatic systems. We also show that though both human and automatic detector performance improve with longer excerpt length, even multi-sentence excerpts can fool expert human raters over 30% of the time. Our findings reveal the importance of using both human and automatic detectors to assess the humanness of text generation systems.
We propose a sentence-level language model which selects the next sentence in a story from a finite set of fluent alternatives. Since it does not need to model fluency, the sentence-level language model can focus on longer range dependencies, which are crucial for multi-sentence coherence. Rather than dealing with individual words, our method treats the story so far as a list of pre-trained sentence embeddings and predicts an embedding for the next sentence, which is more efficient than predicting word embeddings. Notably this allows us to consider a large number of candidates for the next sentence during training. We demonstrate the effectiveness of our approach with state-of-the-art accuracy on the unsupervised Story Cloze task and with promising results on larger-scale next sentence prediction tasks.
We propose a suite of reasoning tasks on two types of relations between procedural events: goal-step relations (“learn poses” is a step in the larger goal of “doing yoga”) and step-step temporal relations (“buy a yoga mat” typically precedes “learn poses”). We introduce a dataset targeting these two relations based on wikiHow, a website of instructional how-to articles. Our human-validated test set serves as a reliable benchmark for common-sense inference, with a gap of about 10% to 20% between the performance of state-of-the-art transformer models and human performance. Our automatically-generated training set allows models to effectively transfer to out-of-domain tasks requiring knowledge of procedural events, with greatly improved performances on SWAG, Snips, and Story Cloze Test in zero- and few-shot settings.
In recent years, large neural networks for natural language generation (NLG) have made leaps and bounds in their ability to generate fluent text. However, the tasks of evaluating quality differences between NLG systems and understanding how humans perceive the generated text remain both crucial and difficult. In this system demonstration, we present Real or Fake Text (RoFT), a website that tackles both of these challenges by inviting users to try their hand at detecting machine-generated text in a variety of domains. We introduce a novel evaluation task based on detecting the boundary at which a text passage that starts off human-written transitions to being machine-generated. We show preliminary results of using RoFT to evaluate detection of machine-generated news articles.
While conditional language models have greatly improved in their ability to output high quality natural language, many NLP applications benefit from being able to generate a diverse set of candidate sequences. Diverse decoding strategies aim to, within a given-sized candidate list, cover as much of the space of high-quality outputs as possible, leading to improvements for tasks that rerank and combine candidate outputs. Standard decoding methods, such as beam search, optimize for generating high likelihood sequences rather than diverse ones, though recent work has focused on increasing diversity in these methods. In this work, we perform an extensive survey of decoding-time strategies for generating diverse outputs from a conditional language model. In addition, we present a novel method where we over-sample candidates, then use clustering to remove similar sequences, thus achieving high diversity without sacrificing quality.
This work presents PerspectroScope, a web-based system which lets users query a discussion-worthy natural language claim, and extract and visualize various perspectives in support or against the claim, along with evidence supporting each perspective. The system thus lets users explore various perspectives that could touch upon aspects of the issue at hand. The system is built as a combination of retrieval engines and learned textual-entailment-like classifiers built using a few recent developments in natural language understanding. To make the system more adaptive, expand its coverage, and improve its decisions over time, our platform employs various mechanisms to get corrections from the users. PerspectroScope is available at github.com/CogComp/perspectroscope Web demo link: http://orwell.seas.upenn.edu:4002/ Link to demo video: https://www.youtube.com/watch?v=MXBTR1Sp3Bs
Many natural language processing tasks require discriminating the particular meaning of a word in context, but building corpora for developing sense-aware models can be a challenge. We present a large resource of example usages for words having a particular meaning, called Paraphrase-Sense-Tagged Sentences (PSTS). Built on the premise that a word’s paraphrases instantiate its fine-grained meanings (i.e., bug has different meanings corresponding to its paraphrases fly and microbe) the resource contains up to 10,000 sentences for each of 3 million target-paraphrase pairs where the target word takes on the meaning of the paraphrase. We describe an automatic method based on bilingual pivoting used to enumerate sentences for PSTS, and present two models for ranking PSTS sentences based on their quality. Finally, we demonstrate the utility of PSTS by using it to build a dataset for the task of hypernym prediction in context. Training a model on this automatically generated dataset produces accuracy that is competitive with a model trained on smaller datasets crafted with some manual effort.
One key consequence of the information revolution is a significant increase and a contamination of our information supply. The practice of fact checking won’t suffice to eliminate the biases in text data we observe, as the degree of factuality alone does not determine whether biases exist in the spectrum of opinions visible to us. To better understand controversial issues, one needs to view them from a diverse yet comprehensive set of perspectives. For example, there are many ways to respond to a claim such as “animals should have lawful rights”, and these responses form a spectrum of perspectives, each with a stance relative to this claim and, ideally, with evidence supporting it. Inherently, this is a natural language understanding task, and we propose to address it as such. Specifically, we propose the task of substantiated perspective discovery where, given a claim, a system is expected to discover a diverse set of well-corroborated perspectives that take a stance with respect to the claim. Each perspective should be substantiated by evidence paragraphs which summarize pertinent results and facts. We construct PERSPECTRUM, a dataset of claims, perspectives and evidence, making use of online debate websites to create the initial data collection, and augmenting it using search engines in order to expand and diversify our dataset. We use crowd-sourcing to filter out noise and ensure high-quality data. Our dataset contains 1k claims, accompanied with pools of 10k and 8k perspective sentences and evidence paragraphs, respectively. We provide a thorough analysis of the dataset to highlight key underlying language understanding challenges, and show that human baselines across multiple subtasks far outperform ma-chine baselines built upon state-of-the-art NLP techniques. This poses a challenge and opportunity for the NLP community to address.
Sentence simplification is the task of rewriting texts so they are easier to understand. Recent research has applied sequence-to-sequence (Seq2Seq) models to this task, focusing largely on training-time improvements via reinforcement learning and memory augmentation. One of the main problems with applying generic Seq2Seq models for simplification is that these models tend to copy directly from the original sentence, resulting in outputs that are relatively long and complex. We aim to alleviate this issue through the use of two main techniques. First, we incorporate content word complexities, as predicted with a leveled word complexity model, into our loss function during training. Second, we generate a large set of diverse candidate simplifications at test time, and rerank these to promote fluency, adequacy, and simplicity. Here, we measure simplicity through a novel sentence complexity model. These extensions allow our models to perform competitively with state-of-the-art systems while generating simpler sentences. We report standard automatic and human evaluation metrics.
Open-domain dialog systems (i.e. chatbots) are difficult to evaluate. The current best practice for analyzing and comparing these dialog systems is the use of human judgments. However, the lack of standardization in evaluation procedures, and the fact that model parameters and code are rarely published hinder systematic human evaluation experiments. We introduce a unified framework for human evaluation of chatbots that augments existing tools and provides a web-based hub for researchers to share and compare their dialog systems. Researchers can submit their trained models to the ChatEval web interface and obtain comparisons with baselines and prior work. The evaluation code is open-source to ensure standardization and transparency. In addition, we introduce open-source baseline models and evaluation datasets. ChatEval can be found at https://chateval.org.
Word embedding representations provide good estimates of word meaning and give state-of-the art performance in semantic tasks. Embedding approaches differ as to whether and how they account for the context surrounding a word. We present a comparison of different word and context representations on the task of proposing substitutes for a target word in context (lexical substitution). We also experiment with tuning contextualized word embeddings on a dataset of sense-specific instances for each target word. We show that powerful contextualized word representations, which give high performance in several semantics-related tasks, deal less well with the subtle in-context similarity relationships needed for substitution. This is better handled by models trained with this objective in mind, where the inter-dependence between word and context representations is explicitly modeled during training.
Story infilling involves predicting words to go into a missing span from a story. This challenging task has the potential to transform interactive tools for creative writing. However, state-of-the-art conditional language models have trouble balancing fluency and coherence with novelty and diversity. We address this limitation with a hierarchical model which first selects a set of rare words and then generates text conditioned on that set. By relegating the high entropy task of picking rare words to a word-sampling model, the second-stage model conditioned on those words can achieve high fluency and coherence by searching for likely sentences, without sacrificing diversity.
Pre-scheduled events, such as TV shows and sports games, usually garner considerable attention from the public. Twitter captures large volumes of discussions and messages related to these events, in real-time. Twitter streams related to pre-scheduled events are characterized by the following: (1) spikes in the volume of published tweets reflect the highlights of the event and (2) some of the published tweets make reference to the characters involved in the event, in the context in which they are currently portrayed in a subevent. In this paper, we take advantage of these characteristics to identify the highlights of pre-scheduled events from tweet streams and we demonstrate a method to summarize these highlights. We evaluate our algorithm on tweets collected around 2 episodes of a popular TV show, Game of Thrones, Season 7.
Lexical simplification involves identifying complex words or phrases that need to be simplified, and recommending simpler meaning-preserving substitutes that can be more easily understood. We propose a complex word identification (CWI) model that exploits both lexical and contextual features, and a simplification mechanism which relies on a word-embedding lexical substitution model to replace the detected complex words with simpler paraphrases. We compare our CWI and lexical simplification models to several baselines, and evaluate the performance of our simplification system against human judgments. The results show that our models are able to detect complex words with higher accuracy than other commonly used methods, and propose good simplification substitutes in context. They also highlight the limited contribution of context features for CWI, which nonetheless improve simplification compared to context-unaware models.
Building a taxonomy from the ground up involves several sub-tasks: selecting terms to include, predicting semantic relations between terms, and selecting a subset of relational instances to keep, given constraints on the taxonomy graph. Methods for this final step – taxonomic organization – vary both in terms of the constraints they impose, and whether they enable discovery of synonymous terms. It is hard to isolate the impact of these factors on the quality of the resulting taxonomy because organization methods are rarely compared directly. In this paper, we present a head-to-head comparison of six taxonomic organization algorithms that vary with respect to their structural and transitivity constraints, and treatment of synonymy. We find that while transitive algorithms out-perform their non-transitive counterparts, the top-performing transitive algorithm is prohibitively slow for taxonomies with as few as 50 entities. We propose a simple modification to a non-transitive optimum branching algorithm to explicitly incorporate synonymy, resulting in a method that is substantially faster than the best transitive algorithm while giving complementary performance.
We propose a variant of a well-known machine translation (MT) evaluation metric, HyTER (Dreyer and Marcu, 2012), which exploits reference translations enriched with meaning equivalent expressions. The original HyTER metric relied on hand-crafted paraphrase networks which restricted its applicability to new data. We test, for the first time, HyTER with automatically built paraphrase lattices. We show that although the metric obtains good results on small and carefully curated data with both manually and automatically selected substitutes, it achieves medium performance on much larger and noisier datasets, demonstrating the limits of the metric for tuning and evaluation of current MT systems.
We conduct the most comprehensive study to date into translating words via images. To facilitate research on the task, we introduce a large-scale multilingual corpus of images, each labeled with the word it represents. Past datasets have been limited to only a few high-resource languages and unrealistically easy translation settings. In contrast, we have collected by far the largest available dataset for this task, with images for approximately 10,000 words in each of 100 languages. We run experiments on a dozen high resource languages and 20 low resources languages, demonstrating the effect of word concreteness and part-of-speech on translation quality. %We find that while image features work best for concrete nouns, they are sometimes effective on other parts of speech. To improve image-based translation, we introduce a novel method of predicting word concreteness from images, which improves on a previous state-of-the-art unsupervised technique. This allows us to predict when image-based translation may be effective, enabling consistent improvements to a state-of-the-art text-based word translation system. Our code and the Massively Multilingual Image Dataset (MMID) are available at http://multilingual-images.org/.
Adjectives like “warm”, “hot”, and “scalding” all describe temperature but differ in intensity. Understanding these differences between adjectives is a necessary part of reasoning about natural language. We propose a new paraphrase-based method to automatically learn the relative intensity relation that holds between a pair of scalar adjectives. Our approach analyzes over 36k adjectival pairs from the Paraphrase Database under the assumption that, for example, paraphrase pair “really hot” <–> “scalding” suggests that “hot” < “scalding”. We show that combining this paraphrase evidence with existing, complementary pattern- and lexicon-based approaches improves the quality of systems for automatically ordering sets of scalar adjectives and inferring the polarity of indirect answers to “yes/no” questions.
Vector space embedding models like word2vec, GloVe, and fastText are extremely popular representations in natural language processing (NLP) applications. We present Magnitude, a fast, lightweight tool for utilizing and processing embeddings. Magnitude is an open source Python package with a compact vector storage file format that allows for efficient manipulation of huge numbers of embeddings. Magnitude performs common operations up to 60 to 6,000 times faster than Gensim. Magnitude introduces several novel features for improved robustness like out-of-vocabulary lookups.
Bilingual lexicon induction is the task of inducing word translations from monolingual corpora in two languages. In this article we present the most comprehensive analysis of bilingual lexicon induction to date. We present experiments on a wide range of languages and data sizes. We examine translation into English from 25 foreign languages: Albanian, Azeri, Bengali, Bosnian, Bulgarian, Cebuano, Gujarati, Hindi, Hungarian, Indonesian, Latvian, Nepali, Romanian, Serbian, Slovak, Somali, Spanish, Swedish, Tamil, Telugu, Turkish, Ukrainian, Uzbek, Vietnamese, and Welsh. We analyze the behavior of bilingual lexicon induction on low-frequency words, rather than testing solely on high-frequency words, as previous research has done. Low-frequency words are more relevant to statistical machine translation, where systems typically lack translations of rare words that fall outside of their training data. We systematically explore a wide range of features and phenomena that affect the quality of the translations discovered by bilingual lexicon induction. We provide illustrative examples of the highest ranking translations for orthogonal signals of translation equivalence like contextual similarity and temporal similarity. We analyze the effects of frequency and burstiness, and the sizes of the seed bilingual dictionaries and the monolingual training corpora. Additionally, we introduce a novel discriminative approach to bilingual lexicon induction. Our discriminative model is capable of combining a wide variety of features that individually provide only weak indications of translation equivalence. When feature weights are discriminatively set, these signals produce dramatically higher translation quality than previous approaches that combined signals in an unsupervised fashion (e.g., using minimum reciprocal rank). We also directly compare our model’s performance against a sophisticated generative approach, the matching canonical correlation analysis (MCCA) algorithm used by Haghighi et al. (2008). Our algorithm achieves an accuracy of 42% versus MCCA’s 15%.
Recognizing and distinguishing antonyms from other types of semantic relations is an essential part of language understanding systems. In this paper, we present a novel method for deriving antonym pairs using paraphrase pairs containing negation markers. We further propose a neural network model, AntNET, that integrates morphological features indicative of antonymy into a path-based relation detection algorithm. We demonstrate that our model outperforms state-of-the-art models in distinguishing antonyms from other semantic relations and is capable of efficiently handling multi-word expressions.
WordNet has facilitated important research in natural language processing but its usefulness is somewhat limited by its relatively small lexical coverage. The Paraphrase Database (PPDB) covers 650 times more words, but lacks the semantic structure of WordNet that would make it more directly useful for downstream tasks. We present a method for mapping words from PPDB to WordNet synsets with 89% accuracy. The mapping also lays important groundwork for incorporating WordNet’s relations into PPDB so as to increase its utility for semantic reasoning in applications.
Bilingual Lexicon Induction is the task of learning word translations without bilingual parallel corpora. We model this task as a matrix completion problem, and present an effective and extendable framework for completing the matrix. This method harnesses diverse bilingual and monolingual signals, each of which may be incomplete or noisy. Our model achieves state-of-the-art performance for both high and low resource languages.
Semantic relation knowledge is crucial for natural language understanding. We introduce “KnowYourNyms?”, a web-based game for learning semantic relations. While providing users with an engaging experience, the application collects large amounts of data that can be used to improve semantic relation classifiers. The data also broadly informs us of how people perceive the relationships between words, providing useful insights for research in psychology and linguistics.
There is a relationship between what we say and where we say it. Word embeddings are usually trained assuming that semantically-similar words occur within the same textual contexts. We investigate the extent to which semantically-similar words occur within the same geospatial contexts. We enrich a corpus of geolocated Twitter posts with physical data derived from Google Places and OpenStreetMap, and train word embeddings using the resulting geospatial contexts. Intrinsic evaluation of the resulting vectors shows that geographic context alone does provide useful information about semantic relatedness.
The role of word sense disambiguation in lexical substitution has been questioned due to the high performance of vector space models which propose good substitutes without explicitly accounting for sense. We show that a filtering mechanism based on a sense inventory optimized for substitutability can improve the results of these models. Our sense inventory is constructed using a clustering method which generates paraphrase clusters that are congruent with lexical substitution annotations in a development set. The results show that lexical substitution can still benefit from senses which can improve the output of vector space paraphrase ranking models.
In certain fields, real-time knowledge from events can help in making informed decisions. In order to extract pertinent real-time knowledge related to an event, it is important to identify the named entities and their corresponding aliases related to the event. The problem of identifying aliases of named entities that spike has remained unexplored. In this paper, we introduce an algorithm, EntitySpike, that identifies entities that spike in popularity in tweets from a given time period, and constructs an alias list for these spiked entities. EntitySpike uses a temporal heuristic to identify named entities with similar context that occur in the same time period (within minutes) during an event. Each entity is encoded as a vector using this temporal heuristic. We show how these entity-vectors can be used to create a named entity alias list. We evaluated our algorithm on a dataset of temporally ordered tweets from a single event, the 2013 Grammy Awards show. We carried out various experiments on tweets that were published in the same time period and show that our algorithm identifies most entity name aliases and outperforms a competitive baseline.
n this work we adapt machine translation (MT) to grammatical error correction, identifying how components of the statistical MT pipeline can be modified for this task and analyzing how each modification impacts system performance. We evaluate the contribution of each of these components with standard evaluation metrics and automatically characterize the morphological and lexical transformations made in system output. Our model rivals the current state of the art using a fraction of the training data.
Most recent sentence simplification systems use basic machine translation models to learn lexical and syntactic paraphrases from a manually simplified parallel corpus. These methods are limited by the quality and quantity of manually simplified corpora, which are expensive to build. In this paper, we conduct an in-depth adaptation of statistical machine translation to perform text simplification, taking advantage of large-scale paraphrases learned from bilingual texts and a small amount of manual simplifications with multiple references. Our work is the first to design automatic metrics that are effective for tuning and evaluating simplification systems, which will facilitate iterative development for this task.
Simple Wikipedia has dominated simplification research in the past 5 years. In this opinion paper, we argue that focusing on Wikipedia limits simplification research. We back up our arguments with corpus analysis and by highlighting statements that other researchers have made in the simplification literature. We introduce a new simplification dataset that is a significant improvement over Simple Wikipedia, and present a novel quantitative-comparative approach to study the quality of simplification data resources.
Translation of the output of automatic speech recognition (ASR) systems, also known as speech translation, has received a lot of research interest recently. This is especially true for programs such as DARPA BOLT which focus on improving spontaneous human-human conversation across languages. However, this research is hindered by the dearth of datasets developed for this explicit purpose. For Egyptian Arabic-English, in particular, no parallel speechtranscription-translation dataset exists in the same domain. In order to support research in speech translation, we introduce the Callhome Egyptian Arabic-English Speech Translation Corpus. This supplements the existing LDC corpus with four reference translations for each utterance in the transcripts. The result is a three-way parallel dataset of Egyptian Arabic Speech, transcriptions and English translations.
This paper presents a multi-dialect, multi-genre, human annotated corpus of dialectal Arabic. We collected utterances in five Arabic dialects: Levantine, Gulf, Egyptian, Iraqi and Maghrebi. We scraped newspaper websites for user commentary and Twitter for two distinct types of dialectal content. To the best of the authors knowledge, this work is the most diverse corpus of dialectal Arabic in both the source of the content and the number of dialects. Every utterance in the corpus was human annotated on Amazons Mechanical Turk; this stands in contrast to Al-Sabbagh and Girju (2012) where only a small subset was human annotated in order to train a classifier to automatically annotate the remainder of the corpus. We provide a discussion of the methodology used for the annotation in addition to the performance of the individual workers. We extend the Arabic dialect identification task to the Iraqi and Maghrebi dialects and improve the results of Zaidan and Callison-Burch (2011a) on Levantine, Gulf and Egyptian.
We release a massive expansion of the paraphrase database (PPDB) that now includes a collection of paraphrases in 23 different languages. The resource is derived from large volumes of bilingual parallel data. Our collection is extracted and ranked using state of the art methods. The multilingual PPDB has over a billion paraphrase pairs in total, covering the following languages: Arabic, Bulgarian, Chinese, Czech, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Polish, Portugese, Romanian, Russian, Slovak, Slovenian, and Swedish.
We present the American Local News Corpus (ALNC), containing over 4 billion words of text from 2,652 online newspapers in the United States. Each article in the corpus is associated with a timestamp, state, and city. All 50 U.S. states and 1,924 cities are represented. We detail our method for taking daily snapshots of thousands of local and national newspapers and present two example corpus analyses. The first explores how different sports are talked about over time and geography. The second compares per capita murder rates with news coverage of murders across the 50 states. The ALNC is about the same size as the Gigaword corpus and is growing continuously. Version 1.0 is available for research use.
We present a large scale study of the languages spoken by bilingual workers on Mechanical Turk (MTurk). We establish a methodology for determining the language skills of anonymous crowd workers that is more robust than simple surveying. We validate workers’ self-reported language skill claims by measuring their ability to correctly translate words, and by geolocating workers to see if they reside in countries where the languages are likely to be spoken. Rather than posting a one-off survey, we posted paid tasks consisting of 1,000 assignments to translate a total of 10,000 words in each of 100 languages. Our study ran for several months, and was highly visible on the MTurk crowdsourcing platform, increasing the chances that bilingual workers would complete it. Our study was useful both to create bilingual dictionaries and to act as census of the bilingual speakers on MTurk. We use this data to recommend languages with the largest speaker populations as good candidates for other researchers who want to develop crowdsourced, multilingual technologies. To further demonstrate the value of creating data via crowdsourcing, we hire workers to create bilingual parallel corpora in six Indian languages, and use them to train statistical machine translation systems.
We present MultiP (Multi-instance Learning Paraphrase Model), a new model suited to identify paraphrases within the short messages on Twitter. We jointly model paraphrase relations between word and sentence pairs and assume only sentence-level annotations during learning. Using this principled latent variable model alone, we achieve the performance competitive with a state-of-the-art method which combines a latent space model with a feature-based supervised classifier. Our model also captures lexically divergent paraphrases that differ from yet complement previous methods; combining our model with previous work significantly outperforms the state-of-the-art. In addition, we present a novel annotation methodology that has allowed us to crowdsource a paraphrase corpus from Twitter. We make this new dataset available to the research community.
Machine translation (MT) draws from several different disciplines, making it a complex subject to teach. There are excellent pedagogical texts, but problems in MT and current algorithms for solving them are best learned by doing. As a centerpiece of our MT course, we devised a series of open-ended challenges for students in which the goal was to improve performance on carefully constrained instances of four key MT tasks: alignment, decoding, evaluation, and reranking. Students brought a diverse set of techniques to the problems, including some novel solutions which performed remarkably well. A surprising and exciting outcome was that student solutions or their combinations fared competitively on some tasks, demonstrating that even newcomers to the field can help improve the state-of-the-art on hard NLP problems while simultaneously learning a great deal. The problems, baseline code, and results are freely available.
Research into the translation of the output of automatic speech recognition (ASR) systems is hindered by the dearth of datasets developed for that explicit purpose. For SpanishEnglish translation, in particular, most parallel data available exists only in vastly different domains and registers. In order to support research on cross-lingual speech applications, we introduce the Fisher and Callhome Spanish-English Speech Translation Corpus, supplementing existing LDC audio and transcripts with (a) ASR 1-best, lattice, and oracle output produced by the Kaldi recognition system and (b) English translations obtained on Amazon’s Mechanical Turk. The result is a four-way parallel dataset of Spanish audio, transcriptions, ASR lattices, and English translations of approximately 38 hours of speech, with defined training, development, and held-out test sets. We conduct baseline machine translation experiments using models trained on the provided training data, and validate the dataset by corroborating a number of known results in the field, including the utility of in-domain (information, conversational) training data, increased performance translating lattices (instead of recognizer 1-best output), and the relationship between word error rate and BLEU score.
We describe a unified and coherent syntactic framework for supporting a semantically-informed syntactic approach to statistical machine translation. Semantically enriched syntactic tags assigned to the target-language training texts improved translation quality. The resulting system significantly outperformed a linguistically naive baseline model (Hiero), and reached the highest scores yet reported on the NIST 2009 Urdu-English translation task. This finding supports the hypothesis (posed by many researchers in the MT community, e.g., in DARPA GALE) that both syntactic and semantic information are critical for improving translation quality—and further demonstrates that large gains can be achieved for low-resource languages with different word order than English.
Much of the previous work on transliteration has depended on resources and attributes specific to particular language pairs. In this work, rather than focus on a single language pair, we create robust models for transliterating from all languages in a large, diverse set to English. We create training data for 150 languages by mining name pairs from Wikipedia. We train 13 systems and analyze the effects of the amount of training data on transliteration performance. We also present an analysis of the types of errors that the systems make. Our analyses are particularly valuable for building machine translation systems for low resource languages, where creating and integrating a transliteration module for a language with few NLP resources may provide substantial gains in translation performance.
The Joint Probability Model proposed by Marcu and Wong (2002) provides a probabilistic framework for modeling phrase-based statistical machine transla- tion (SMT). The model’s usefulness is, however, limited by the computational complexity of estimating parameters at the phrase level. We present a method of constraining the search space of the Joint Probability Model based on statistically and linguistically motivated word align- ments. This method reduces the complexity and size of the Joint Model and allows it to display performance superior to the standard phrase-based models for small amounts of training material.
This paper describes a program that automatically selects the best translation from a set of translations produced by multiple commercial machine translation engines. The program is simplified by assuming that the most fluent item in the set is the best translation. Fluency is determined using a trigram language model. Results are provided illustrating how well the program performs for human ranked data as compared to each of its constituent engines.
User feedback has often been proposed as a method for improving the accuracy of machine translation systems, but useful feedback can also serve a number of secondary benefits, including increasing user confidence in the MT technology and expanding the potential audience of users. Amikai, Inc. has produced a number of communication tools which embed translation technology and which attempt to improve the user experience by maximizing useful user interaction and feedback. As MT continues to develop, further attention needs to be paid to developing the overall user experience, which can improve the utility of translation tools even when translation quality itself plateaus.