12 Research Papers Accepted to EMNLP 2022

Papers from CS researchers were accepted to the Empirical Methods in Natural Language Processing (EMNLP) 2022. EMNLP is a leading conference in artificial intelligence and natural language processing. Aside from presenting their research papers, several researchers also organized workshops to gather conference attendees for discussions about current issues confronting NLP and computer science.

Workshops

Massively Multilingual Natural Language Understanding
Jack FitzGerald Amazon Alexa, Kay Rottmann Amazon Alexa, Julia Hirschberg Columbia University, Mohit Bansal University of North Carolina, Anna Rumshisky University of Massachusetts Lowell, and Charith Peris Amazon Alexa

3rd Workshop on Figurative Language Processing
Debanjan Ghosh Educational Testing Service, Beata Beigman Klebanov Educational Testing Service, Smaranda Muresan Columbia University, Anna Feldman Montclair State University, Soujanya Poria Singapore University of Technology and Design, and Tuhin Chakrabarty Columbia University

Sharing Stories and Lessons Learned
Diyi Yang Stanford University, Pradeep Dasigi Allen Institute for AI, Sherry Tongshuang Wu Carnegie Mellon University, Tuhin Chakrabarty Columbia University, Yuval Pinter Ben-Gurion University of the Negev, and Mike Zheng Shou National University of Singapore

Accepted Papers

Help me write a Poem – Instruction Tuning as a Vehicle for Collaborative Poetry Writing
Tuhin Chakrabarty Columbia University, Vishakh Padmakumar New York University, He He New York University

Abstract
Recent work in training large language models (LLMs) to follow natural language instructions has opened up exciting opportunities for natural language interface design. Building on the prior success of large language models in the realm of computer assisted creativity, in this work, we present CoPoet, a collaborative poetry writing system, with the goal of to study if LLM’s actually improve the quality of the generated content. In contrast to auto-completing a user’s text, CoPoet is controlled by user instructions that specify the attributes of the desired text, such as Write a sentence about ‘love’ or Write a sentence ending in ‘fly’. The core component of our system is a language model fine-tuned on a diverse collection of instructions for poetry writing. Our model is not only competitive to publicly available LLMs trained on instructions (InstructGPT), but also capable of satisfying unseen compositional instructions. A study with 15 qualified crowdworkers shows that users successfully write poems with CoPoet on diverse topics ranging from Monarchy to Climate change, which are preferred by third-party evaluators over poems written without the system.

FLUTE: Figurative Language Understanding through Textual Explanations
Tuhin Chakrabarty Columbia University, Arkadiy Saakyan Columbia University, Debanjan Ghosh Educational Testing Service, and Smaranda Muresan Columbia University

Abstract
Figurative language understanding has been recently framed as a recognizing textual entailment (RTE) task (a.k.a. natural language inference (NLI)). However, similar to classical RTE/NLI datasets they suffer from spurious correlations and annotation artifacts. To tackle this problem, work on NLI has built explanation-based datasets such as eSNLI, allowing us to probe whether language models are right for the right reasons. Yet no such data exists for figurative language, making it harder to assess genuine understanding of such expressions. To address this issue, we release FLUTE, a dataset of 9,000 figurative NLI instances with explanations, spanning four categories: Sarcasm, Simile, Metaphor, and Idioms. We collect the data through a Human-AI collaboration framework based on GPT-3, crowd workers, and expert annotators. We show how utilizing GPT-3 in conjunction with human annotators (novices and experts) can aid in scaling up the creation of datasets even for such complex linguistic phenomena as figurative language. The baseline performance of the T5 model fine-tuned on FLUTE shows that our dataset can bring us a step closer to developing models that understand figurative language through textual explanations.

Fine-tuned Language Models are Continual Learners
Thomas Scialom Columbia University, Tuhin Chakrabarty Columbia University, and Smaranda Muresan Columbia University

Abstract
Recent work on large language models relies on the intuition that most natural language processing tasks can be described via natural language instructions and that models trained on these instructions show strong zero-shot performance on several standard datasets. However, these models even though impressive still perform poorly on a wide range of tasks outside of their respective training and evaluation sets.To address this limitation, we argue that a model should be able to keep extending its knowledge and abilities, without forgetting previous skills. In spite of the limited success of Continual Learning, we show that Fine-tuned Language Models can be continual learners.We empirically investigate the reason for this success and conclude that Continual Learning emerges from self-supervision pre-training. Our resulting model Continual-T0 (CT0) is able to learn 8 new diverse language generation tasks, while still maintaining good performance on previous tasks, spanning in total of 70 datasets. Finally, we show that CT0 is able to combine instructions in ways it was never trained for, demonstrating some level of instruction compositionality.

Multitask Instruction-based Prompting for Fallacy Recognition
Tariq Alhindi Columbia University, Tuhin Chakrabarty Columbia University, Elena Musi University of Liverpool, and Smaranda Muresan Columbia University

Abstract
Fallacies are used as seemingly valid arguments to support a position and persuade the audience about its validity. Recognizing fallacies is an intrinsically difficult task both for humans and machines. Moreover, a big challenge for computational models lies in the fact that fallacies are formulated differently across the datasets with differences in the input format (e.g., question-answer pair, sentence with fallacy fragment), genre (e.g., social media, dialogue, news), as well as types and number of fallacies (from 5 to 18 types per dataset). To move towards solving the fallacy recognition task, we approach these differences across datasets as multiple tasks and show how instruction-based prompting in a multitask setup based on the T5 model improves the results against approaches built for a specific dataset such as T5, BERT or GPT-3. We show the ability of this multitask prompting approach to recognize 28 unique fallacies across domains and genres and study the effect of model size and prompt choice by analyzing the per-class (i.e., fallacy type) results. Finally, we analyze the effect of annotation quality on model performance, and the feasibility of complementing this approach with external knowledge.

CONSISTENT: Open-Ended Question Generation From News Articles
Tuhin Chakrabarty Columbia University, Justin Lewis The New York Times R&D, and Smaranda Muresan Columbia University

Abstract
Recent work on question generation has largely focused on factoid questions such as who, what, where, when about basic facts. Generating open-ended why, how, what, etc. questions that require long-form answers have proven more difficult. To facilitate the generation of open-ended questions, we propose CONSISTENT, a new end-to-end system for generating open-ended questions that are answerable from and faithful to the input text. Using news articles as a trustworthy foundation for experimentation, we demonstrate our model’s strength over several baselines using both automatic and human=based evaluations. We contribute an evaluation dataset of expert-generated open-ended questions.We discuss potential downstream applications for news media organizations.

SafeText: A Benchmark for Exploring Physical Safety in Language Models
Sharon Levy University of California, Santa Barbara, Emily Allaway Columbia University, Melanie Subbiah Columbia University, Lydia Chilton Columbia University, Desmond Patton Columbia University, Kathleen McKeown Columbia University, and William Yang Wang University of California, Santa Barbara

Abstract
Understanding what constitutes safe text is an important issue in natural language processing and can often prevent the deployment of models deemed harmful and unsafe. One such type of safety that has been scarcely studied is commonsense physical safety, i.e. text that is not explicitly violent and requires additional commonsense knowledge to comprehend that it leads to physical harm. We create the first benchmark dataset, SafeText, comprising real-life scenarios with paired safe and physically unsafe pieces of advice. We utilize SafeText to empirically study commonsense physical safety across various models designed for text generation and commonsense reasoning tasks. We find that state-of-the-art large language models are susceptible to the generation of unsafe text and have difficulty rejecting unsafe advice. As a result, we argue for further studies of safety and the assessment of commonsense physical safety in models before release.

Learning to Revise References for Faithful Summarization
Griffin Adams Columbia University, Han-Chin Shing Amazon AWS AI, Qing Sun Amazon AWS AI, Christopher Winestock Amazon AWS AI, Kathleen McKeown Columbia University, and Noémie Elhadad Columbia University

Abstract
In real-world scenarios with naturally occurring datasets, reference summaries are noisy and may contain information that cannot be inferred from the source text. On large news corpora, removing low quality samples has been shown to reduce model hallucinations. Yet, for smaller, and/or noisier corpora, filtering is detrimental to performance. To improve reference quality while retaining all data, we propose a new approach: to selectively re-write unsupported reference sentences to better reflect source data. We automatically generate a synthetic dataset of positive and negative revisions by corrupting supported sentences and learn to revise reference sentences with contrastive learning. The intensity of revisions is treated as a controllable attribute so that, at inference, diverse candidates can be over-generated-then-rescored to balance faithfulness and abstraction. To test our methods, we extract noisy references from publicly available MIMIC-III discharge summaries for the task of hospital-course summarization, and vary the data on which models are trained. According to metrics and human evaluation, models trained on revised clinical references are much more faithful, informative, and fluent than models trained on original or filtered data.

Mitigating Covertly Unsafe Text within Natural Language Systems
Alex Mei University of California, Santa Barbara, Anisha Kabir University of California, Santa Barbara, Sharon Levy University of California, Santa Barbara, Melanie Subbiah Columbia University, Emily Allaway Columbia University, John N. Judge University of California, Santa Barbara, Desmond Patton University of Pennsylvania, Bruce Bimber University of California, Santa Barbara, Kathleen McKeown Columbia University, and William Yang Wang University of California, Santa Barbara

Abstract
An increasingly prevalent problem for intelligent technologies is text safety, as uncontrolled systems may generate recommendations to their users that lead to injury or life-threatening consequences. However, the degree of explicitness of a generated statement that can cause physical harm varies. In this paper, we distinguish types of text that can lead to physical harm and establish one particularly underexplored category: covertly unsafe text. Then, we further break down this category with respect to the system’s information and discuss solutions to mitigate the generation of text in each of these subcategories. Ultimately, our work defines the problem of covertly unsafe language that causes physical harm and argues that this subtle yet dangerous issue needs to be prioritized by stakeholders and regulators. We highlight mitigation strategies to inspire future researchers to tackle this challenging problem and help improve safety within smart systems.

Affective Idiosyncratic Responses to Music
Sky CH-Wang Columbia University, Evan Li Columbia University, Oliver Li Columbia University, Smaranda Muresan Columbia University, and Zhou Yu Columbia University

Abstract
Affective responses to music are highly personal. Despite consensus that idiosyncratic factors play a key role in regulating how listeners emotionally respond to music, precisely measuring the marginal effects of these variables has proved challenging. To address this gap, we develop computational methods to measure affective responses to music from over 403M listener comments on a Chinese social music platform. Building on studies from music psychology in systematic and quasi-causal analyses, we test for musical, lyrical, contextual, demographic, and mental health effects that drive listener affective responses. Finally, motivated by the social phenomenon known as 网抑云 (wǎng-yì-yún), we identify influencing factors of platform user self-disclosures, the social support they receive, and notable differences in discloser user activity.

Robots-Dont-Cry: Understanding Falsely Anthropomorphic Utterances in Dialog Systems
David Gros University of California, Davis, Yu Li Columbia University, and Zhou Yu Columbia University

Abstract
Dialog systems are often designed or trained to output human-like responses. However, some responses may be impossible for a machine to truthfully say (e.g. “that movie made me cry”). Highly anthropomorphic responses might make users uncomfortable or implicitly deceive them into thinking they are interacting with a human. We collect human ratings on the feasibility of approximately 900 two-turn dialogs sampled from 9 diverse data sources. Ratings are for two hypothetical machine embodiments: a futuristic humanoid robot and a digital assistant. We find that for some data-sources commonly used to train dialog systems, 20-30% of utterances are not viewed as possible for a machine. Rating is marginally affected by machine embodiment. We explore qualitative and quantitative reasons for these ratings. Finally, we build classifiers and explore how modeling configuration might affect output permissibly, and discuss implications for building less falsely anthropomorphic dialog systems.

Just Fine-tune Twice: Selective Differential Privacy for Large Language Models
Weiyan Shi Columbia University, Ryan Patrick Shea Columbia University, Si Chen Columbia University, Chiyuan Zhang Google Research, Ruoxi Jia Virginia Tech, and Zhou Yu Columbia University

Abstract
Protecting large language models from privacy leakage is becoming increasingly crucial with their wide adoption in real-world products. Yet applying *differential privacy* (DP), a canonical notion with provable privacy guarantees for machine learning models, to those models remains challenging due to the trade-off between model utility and privacy loss. Utilizing the fact that sensitive information in language data tends to be sparse, Shi et al. (2021) formalized a DP notion extension called *Selective Differential Privacy* (SDP) to protect only the sensitive tokens defined by a policy function. However, their algorithm only works for RNN-based models. In this paper, we develop a novel framework, *Just Fine-tune Twice* (JFT), that achieves SDP for state-of-the-art large transformer-based models. Our method is easy to implement: it first fine-tunes the model with *redacted* in-domain data, and then fine-tunes it again with the *original* in-domain data using a private training mechanism. Furthermore, we study the scenario of imperfect implementation of policy functions that misses sensitive tokens and develop systematic methods to handle it. Experiments show that our method achieves strong utility compared to previous baselines. We also analyze the SDP privacy guarantee empirically with the canary insertion attack.

Focus! Relevant and Sufficient Context Selection for News Image Captioning
Mingyang Zhou University of California, Davis, Grace Luo University of California, Berkeley, Anna Rohrbach University of California, Berkeley, and Zhou Yu Columbia University

Abstract
News Image Captioning requires describing an image by leveraging additional context from a news article. Previous works only coarsely leverage the article to extract the necessary context, which makes it challenging for models to identify relevant events and named entities. In our paper, we first demonstrate that by combining more fine-grained context that captures the key named entities (obtained via an oracle) and the global context that summarizes the news, we can dramatically improve the model’s ability to generate accurate news captions. This begs the question, how to automatically extract such key entities from an image? We propose to use the pre-trained vision and language retrieval model CLIP to localize the visually grounded entities in the news article and then capture the non-visual entities via an open relation extraction model. Our experiments demonstrate that by simply selecting a better context from the article, we can significantly improve the performance of existing models and achieve new state-of-the-art performance on multiple benchmarks.

Natural Language Processing and Spoken Language Processing groups present papers at EMNLP 2018

Columbia researchers presented their work at the Empirical Methods in Natural Language Processing (EMNLP) in Brussels, Belgium.

Professor Julia Hirschberg gave a keynote talk on the work done by the Spoken Language Processing Group on how to automatically detect deception in spoken language – how to identify cues in trusted speech vs. mistrusted speech and how these features differ by speaker and by listener. Slides from the talk can be viewed here.

Five teams with computer science undergrad and PhD students from the Natural Language Processing Group (NLP) also attended the conference to showcase their work on text summarization, analysis of social media, and fact checking.

Robust Document Retrieval and Individual Evidence Modeling for Fact Extraction and Verification
Tuhin Chakrabarty Computer Science Department, Tariq Alhindi Computer Science Department, and Smaranda Muresan Computer Science Department and Data Science Institute

”Given the difficult times, we are living in, it’s extremely necessary to be perfect with our facts,” said Tuhin Chakrabarty, lead researcher of the paper. “Misinformation spreads like wildfire and has long-lasting impacts. This motivated us to delve into the area of fact extraction and verification.”

This paper presents the ColumbiaNLP submission for the FEVER Workshop Shared Task. Their system is an end-to-end pipeline that extracts factual evidence from Wikipedia and infers a decision about the truthfulness of the claim based on the extracted evidence.

Fact checking is a type of investigative journalism where experts examine the claims published by others for their veracity. The claims can range from statements made by public figures to stories reported by other publishers. The end goal of a fact checking system is to provide a verdict on whether the claim is true, false, or mixed. Several organizations such as FactCheck.org and PolitiFact are devoted to such activities.

The FEVER Shared task aims to evaluate the ability of a system to verify information using evidence from Wikipedia. Given a claim involving one or more entities (mapping to Wikipedia pages), the system must extract textual evidence (sets of sentences from Wikipedia pages) that supports or refutes the claim and then using this evidence, it must label the claim as Supported, Refuted or NotEnoughInfo.

Detecting Gang-Involved Escalation on Social Media Using Context
Serina Chang Computer Science Department, Ruiqi Zhong Computer Science Department, Ethan Adams Computer Science Department, Fei-Tzin Lee Computer Science Department, Siddharth Varia Computer Science Department, Desmond Patton School of Social Work, William Frey School of Social Work, Chris Kedzie Computer Science Department, and Kathleen McKeown Computer Science Department

This research is a collaboration between Professor Kathy McKeown’s NLP lab and the Columbia School of Social Work. Professor Desmond Patton, from the School of Social Work and a member of the Data Science Institute, discovered that gang-involved youth in cities such as Chicago increasingly turn to social media to grieve the loss of loved ones, which may escalate into aggression toward rival gangs and plans for violence.

The team created a machine learning system that can automatically detect aggression and loss in the social media posts of gang-involved youth. They developed an approach with the hope to eventually use a system that can save critical time, scale reach, and intervene before more young lives are lost.

The system features the use of word embeddings and lexicons, automatically derived from a large domain-specific corpus which the team constructed. They also created context features that capture user’s recent posts, both in semantic and emotional content, and their interactions with other users in the dataset. Incorporating domain-specific resources and context feature in a Convolutional Neural Network (CNN) that leads to a significant improvement over the prior state-of-the-art.

The dataset used spans the public Twitter posts of nearly 300 users from a gang-involved community in Chicago. Youth volunteers and violence prevention organizations helped identify users and annotate the dataset for aggression and loss. Here are two examples of labeled tweets, both of which the system was able to classify correctly. Names are blocked out to preserve the privacy of users.

For semantics, which were represented by word embeddings, the researchers found that it was optimal to include 90 days of recent tweet history. While for emotion, where an emotion lexicon was employed, only two days of recent tweets were needed. This matched insight from prior social work research, which found that loss is significantly likely to precede aggression in a two-day window. They also found that emotions fluctuate more quickly than semantics so the tighter context window would be able to capture more fine-grained fluctuation.

“We took this context-driven approach because we believed that interpreting emotion in a given tweet requires context, including what the users had been saying recently, how they had been feeling, and their social dynamics with others,” said Serina Chang, an undergraduate computer science student. One thing that surprised them was the extent to which different types of context offered different types of information, as demonstrated by the contrasting contributions of the semantic-based user history feature and the emotion-based one. Continued Chang, “As we hypothesized, adding context did result in a significant performance improvement in our neural net model.”

Team SWEEPer: Joint Sentence Extraction and Fact Checking with Pointer Networks
Christopher Hidey Columbia University, Mona Diab Amazon AI Lab

Automated fact checking of textual claims is of increasing interest in today’s world. Previous research has investigated fact checking in political statements, news articles, and community forums.

“Through our model we can fact check claims and find specific statements that support the evidence,” said Christopher Hidey, a fourth year PhD student. “This is a step towards addressing the propagation of misinformation online.”

As part of the FEVER community shared task, the researchers developed models that given a statement would jointly find a Wikipedia article and a sentence related to the statement, and then predict whether the statement is supported by that sentence.

For example, given the claim “Lorelai Gilmore’s father is named Robert,” one could find the Wikipedia article on Lorelai Gilmore and extract the third sentence “Lorelai has a strained relationship with her wealthy parents, Richard and Emily, after running away as a teen to raise her daughter on her own” to show that the claim is false.

Credit : Wikipedia – https://en.wikipedia.org/wiki/Lorelai_Gilmore

One aspect of this problem that the team observed was how poorly TF-IDF, a standard technique in information retrieval and natural language processing, performed at retrieving Wikipedia articles and sentences. Their custom model improved performance by 35 points in terms of recall over a TF-IDF baseline, achieving 90% recall for 5 articles. Overall, the model retrieved the correct sentence and predicted the veracity of the claim 50% of the time.

Where is your Evidence: Improving Fact-checking by Justification Modeling
Tariq Alhindi Computer Science Department, Savvas Petridis Computer Science Department, Smaranda Muresan Computer Science Department and Data Science Institute

The rate of which misinformation is spreading on the web is faster than the rate of manual fact-checking conducted by organizations like Politifact.com and Factchecking.org. For this paper the researchers wanted to explore how to automate parts or all of the fact-checking process. A poster with their findings was presented as part of the FEVER workshop.

“In order to come up with reliable fact-checking systems we need to understand the current manual process and identify opportunities for automation,” said Tariq Alhindi, lead author on the paper. They looked at the LIAR dataset – around 10,000 claims classified by Politifact.com to one of six degrees of truth – pants-on-fire, false, mostly-false, half-true, mostly-true, true. Continued Alhindi, we also looked at the fact-checking article for each claim and automatically extracted justification sentences of a given verdict and used them in our models, after removing all sentences that contain the verdict (e.g. true or false).

Feature-based machine learning models and neural networks were used to develop models that can predict whether a given statement is true or false. Results showed that using some sort of justification or evidence always improves the results of fake-news detection models.

“What was most surprising about the results is that adding features from the extracted justification sentences consistently improved the results no matter what classifier we used or what other features we included,” shared Alhindi, a PhD student. “However, we were surprised that the improvement was consistent even when we compare traditional feature-based linear machine learning models against state of the art deep learning models.”

Their research extends the previous work done on this data set which only looked at the linguistic cues of the claim and/or the metadata of the speaker (history, venue, party-affiliation, etc.). The researchers also released the extended dataset to the community to allow further work on this dataset with the extracted justifications.

Content Selection in Deep Learning Models of Summarization
Chris Kedzie Columbia University, Kathleen McKeown Columbia University, Hal Daume III University of Maryland, College Park

Recently, a specific type of machine learning, called deep learning, has made strides in reaching human level performance on hard to articulate problems, that is, things people do subconsciously like recognizing faces or understanding speech. And so, natural language processing researchers have turned to these models for the task of identifying the most important phrases and sentences in text documents, and have trained them to imitate the decisions a human editor might make when selecting content for a summary.

“Deep learning models have been successful in summarizing natural language texts, news articles and online comments,” said Chris Kedzie, a fifth year PhD student. “What we wanted to know is how they are doing it.”

While these deep learning models are empirically successful, it is not clear how they are performing this task. By design, they are learning to create their own representation of words and sentences, and then using them to predict whether a sentence is important – if it should go into a summary of the document. But just what kinds of information are they using to create these representations?

One hypotheses the researchers had was that certain types of words were more informative than others. For example, in a news article, nouns and verbs might be more important than adjectives and adverbs for identifying the most important information since such articles are typically written in a relatively objective manner.

To see if this was so, they trained models to predict sentence importance on redacted datasets, where either nouns, verbs, adjectives, adverbs, or function words were removed and compared them to models trained on the original data.

On a dataset of personal stories published on Reddit, adjectives and adverbs were the key to achieving the best performance. This made intuitive sense in that people tend to use intensifiers to highlight the most important or climactic moments in their stories with sentences like, “And those were the WORST customers I ever served.”

What surprised the researchers were the news articles – removing any one class of words did not dramatically decrease model performance. Either important content was broadly distributed across all kinds of words or there was some other signal that the model was using.

They suspected that sentence order was important because journalists are typically instructed to write according to the inverted pyramid style with the most important information at the top of the article. It was possible that the models were implicitly learning this and simply selecting sentences from the article lead.

Two pieces of evidence confirmed this. First, looking at a histogram of sentence positions selected as important, the models overwhelmingly preferred the lead of the article. Second, in a follow up experiment, the sentence ordered was shuffled to remove sentence position as a viable signal from which to learn. On news articles, model performance dropped significantly, leading to the conclusion that sentence position was most responsible for model performance on news documents.

The result concerned the researchers as they want models to be trained to truly understand human language and not use simple and brittle heuristics (like sentence position). “To connect this to broader trends in machine learning, we should be very concerned and careful about what signals are being exploited by our models, especially when making sensitive decisions,” Kedzie continued. ”The signals identified by the model as helpful may not truly capture the problem we are trying to solve, and worse yet, may be exploiting biases in the dataset that we do not wish it to learn.”

However, Kedzie sees this as an opportunity to improve the utility of word representations so that models are better able to use the article content itself. Along these lines, in the future, he hopes to show that by quantifying the surprisal or novelty of a particular word or phrase, models are able to make better sentence importance predictions. Just as people might remember the most surprising and unexpected parts of a good story.

Tag: EMNLP