Six papers from CS researchers were accepted to the 16th conference of the European Chapter of the Association for Computational Linguistics (EACL). As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.
Below are brief descriptions and links to the papers.
This paper presents a new clustering paradigm for news streams, where clusters have a one-to-one correspondence with real-world events (for example, the Suez canal blockage). An important aspect of this problem is that the number of clusters is unknown and varies with time (new events occur and old events cease to be of relevance). The proposed paradigm follows a pipeline approach – where representations are built for each new article, comparisons are made with existing clusters to pick the most compatible one, and finally, a clustering decision is produced.
A surprising observation from this work is that contextual embeddings (from models like BERT), in contrast to their overwhelming success in many NLP problems, achieve sub-par performance by themselves on this clustering problem. However, when combined with other representations (like TF-IDF and timestamps) and fine-tuned with task-specific augmentations, they achieve new state-of-the-art performance. Another interesting observation is that the widely reported B-Cubed metrics are biased towards large clusters and hence don’t capture cluster fragmentation on smaller clusters as well. Since clusters corresponding to emerging events are small and errors made on such clusters are highly undesirable, the authors suggest using an additional metric CEAF-e to evaluate models for this task.
Segmenting Subtitles for Correcting ASR Segmentation Errors David Wan Columbia University, Chris Kedzie Columbia University, Faisal Ladakh Columbia University, Elsbeth Turcan Columbia University, Petra Galuszkova University of Maryland, Elena Zotkina University of Maryland, Zhengping Jiang Columbia University, Peter Bell University of Edinburgh, and Kathleen McKeown Columbia University
For the task of spoken language translation, the usual approach is to have a pipeline consisting of Automatic Speech Recognition (ASR) that transforms audio files into words and utterances in the original language and a Machine Translation (MT) that translate the utterances into the target language. However this setup may suffer from input-output mismatches: ASR segments utterances by acoustic information such as pauses, and thus may produce run-on sentences or sentence fragments, but MT is usually trained on proper sentences without such issues and may not perform well under such setting. This paper proposes the use of an intermediate model to segment utterances into sentences to improve performance in MT as well as other downstream tasks.
One crucial problem for developing such models is the lack of suitable training data for segmentation, especially when the languages involved are low-resourced. To this end, this paper also proposes a way to use subtitles dataset as proxy speech data as well as creating synthetic acoustic utterances that mimic common ASR errors for the model to learn to fix. Using a simple neural tagging model, the authors of this paper show improvement over the baseline ASR segmentation on MT for Lithuanian, Bulgarian, and Farisi. A surprising finding is that the segmentation model most improves the translation quality of more syntactically complex segments.
It has been well-documented for several languages that human interlocutors tend to adapt their linguistic productions to become more similar to each other. This behavior, known as entrainment, affects lexical choice as well, both with regard to specific words, such as referring expressions, and overall style.
Lexical entrainment is the behavior that causes the words that speakers use in a conversation to become more similar over time. Entrainment more broadly is a human behavior causing interlocutors to adapt to each other to become more similar. Its effects are measurable but entrainment itself is not a measure.
This paper offers the first investigation of such lexical entrainment in Hebrew.
The analysis of Hebrew speakers interacting in a Map Task, a popular experimental setup, provides rich evidence of lexical entrainment. No clear pattern of differences is found between speaker pairs by the combination of their genders, nor between speakers by their individual gender. However, speakers in a position of less power are found to entrain more than those with greater power, which matches theoretical accounts.
Overall, the results mostly accord with those for American English. There is, however, a surprising lack of entrainment on a list of hedge words that were previously found to be highly entrained in English. This might be due to cultural differences between American and Israeli speakers that render adoption of a more tentative style less appropriate in the Hebrew context.
Entity-level Factual Consistency of Abstractive Text Summarization Feng Nan Amazon Web Services, Ramesh Nallapati Amazon Web Services, Zhiguo Wang Amazon Web Services, Cicero Nogueira dos Santos Amazon Web Services, Henghui Zhu Amazon Web Services, Dejiao Zhang Amazon Web Services, Kathleen McKeown Amazon Web Services & Columbia University, Bing Xiang Amazon Web Services
A key challenge for abstractive summarization is ensuring factual consistency of the generated summary with respect to the original document. For example, state-of-the-art models trained on existing datasets exhibit entity hallucination, generating names of entities that are not present in the source document.
The paper proposes a set of new metrics to quantify the entity-level factual consistency of generated summaries and shows that the entity hallucination problem can be alleviated by simply filtering the training data. In addition, the paper introduces a summary-worthy entity classification task to the training process as well as a joint entity and summary generation approach, which yields further improvements in entity-level metrics.
Detecting arguments in online interactions is useful to understand how conflicts arise and get resolved. Users often use figurative language, such as sarcasm, either as persuasive devices or to attack the opponent by an ad hominem argument. To further our understanding of the role of sarcasm in shaping the disagreement space, the paper presents a thorough experimental setup using a corpus annotated with both argumentative moves (agree/disagree) and sarcasm. The research exploits joint modeling in terms of (a) applying discrete features that are useful in detecting sarcasm to the task of argumentative relation classification (agree/disagree/none), and (b) multitask learning for argumentative relation classification and sarcasm detection using deep learning architectures (e.g., dual Long ShortTerm Memory (LSTM) with hierarchical attention and Transformer-based architectures). The paper shows that modeling sarcasm improves the argumentative relation classification task (agree/disagree/none) in all setups.
Ideological attitudes and stances are often expressed through subtle meanings of words and phrases. Understanding these connotations is critical to recognize the cultural and emotional perspectives of the speaker. In this paper, the researchers use distant labeling to create a new lexical resource representing connotation aspects for nouns and adjectives. Their analysis shows that it aligns well with human judgments. Additionally, they present a method for creating lexical representations that capture connotations within the embedding space and show that using the embeddings provides a statistically significant improvement on the task of stance detection when data is limited.
Assistant Professor Carl Vondrick has won the National Science Foundation’s (NSF) Faculty Early Career Development award for his proposal program to develop machine perception systems that robustly detect and track objects even when they disappear from sight, thereby enabling machines to build spatial awareness of their surroundings.
Shuran Song and Carl Vondrick are among the awardees chosen for their artificial intelligence (AI) research. The program aims to use AI for societal good.
The award recognizes Moti Yung’s significant contributions to the field of cryptography, specifically, coinventing malicious cryptography and pioneering contributions to distributed cryptosystems.
Research from the department was accepted to the 35th AAAI Conference on Artificial Intelligence. The conference promotes research in artificial intelligence (AI) and scientific exchange among AI researchers, practitioners, scientists, and engineers in affiliated disciplines.
One of the most exciting applications of modern artificial intelligence is to automatically discover scientific laws from experimental data. This is not a trivial problem as it involves searching for a complex mathematical relationship over a large set of explanatory variables and operators that can be combined in an infinite number of ways. Inspired by the incredible success of deep learning in computer vision, the authors tackle this problem by adapting various successful network architectures into the symbolic law discovery pipeline. The novelty of this new approach is in (1) encoding the input data as an image with super-resolution, (2) developing an appropriate deep network pipeline, and (3) predicting the importance of each mathematical operator from the relationship image. This allowed to prior the exponentially large search with the predicted importance of the symbolic operators, which can significantly accelerate the discovery process.
The model was then applied to a variety of plausible relationships—both simulated and from physics and mathematics domains—involving different dimensions and constituents. The authors show that their model is able to identify the underlying operators from data, achieving a high accuracy and AUC (91% and 0.96 on average resp.) for systems with as many as ten independent variables. Their method significantly outperforms the current state of the art in terms of data fitting (R^2), discovery rate (recovering the true relationship), and succinctness (output formula complexity). The discovered equations can be seen as first drafts of scientific laws that can be helpful to the scientists for (1) hypothesis building, and (2) understanding the complex underlying structure of the studied phenomena. This novel approach holds a real promise to help speed up the rate of scientific discovery.
One of the most common methods for policy learning used throughout the empirical sciences is the use of randomization of the treatment assignment. This method is considered the gold standard within many disciplines and can be traced back, at least, to Fisher (Fisher 1935) and Neyman (Neyman 1923). Whenever human subjects are at the center of the experiment, unfortunately, issues of non-compliance arise. Namely, subjects do not necessarily follow the experimental protocol and end up doing what they want. It is well-understood that under such conditions, unobserved confounding bias will emerge. For instance, subjects who did not comply with the treatment assignment may be precisely those who would have responded adversely to the treatment. Therefore, the actual causal effects of the treatment, when it is applied uniformly to the population, might be substantially less effective than the data reveals. Moreover, since one does not observe how subjects decide/respond to the realized treatment, the actual treatment effects are not uniquely computably from the collected data, called non-identifiable.
Robins (1989) and Manski (1990) derived the first informative bounds over the causal effects from studies with imperfect compliance under a set of non-parametric assumptions called instrumental variables (IV). In their seminal work, Balke and Pearl (1994a, 1997) improved earlier results by employing an algebraic method to derive analytic expressions of the causal bounds, which are provably optimal. However, this approach assumes the primary outcome to be discrete and finite. Solving such a program could be intractable when high-dimensional context variables are present.
This paper presents novel non-parametric methods to bound causal effects on the continuous outcome from studies with imperfect compliance. These methods could be generalized to settings with a high-dimensional context. Perhaps surprisingly, this paper introduced a latent data representation that could characterize all constraints on the observational and interventional distributions implied by IV assumptions, even when the primary outcome is continuous. Such representation allows one to reduce the original bounding problem to a series of linear programs. Solve these programs, therefore, leads to tight causal bounds.
Learning causal effects from observational data is a pervasive challenge found throughout the data-intensive sciences. General methods of determining the identifiability of causal effect from a combination of observational data and causal knowledge about the underlying system have been well-understood in theory. In practice, however, there are still challenges to estimating identifiable causal functionals from finite samples. Recently, a novel approach, named double/debiased machine learning (DML) (Chernozhukov et al. 2018), has been proposed to learn parameters leveraging modern machine learning techniques, which are both robust to model misspecification (‘doubly robust’) and slow convergence (‘debiased’). Still, DML has only been used for causal estimation in settings when the back-door condition (also known as conditional ignorability) holds.
This paper aims to bridge this gap by developing a general class of estimators for any identifiable causal functionals that exhibit robustness properties of DML estimators, which the authors called ‘DML-ID.’ In particular, they provide a complete procedure for deriving an essential ingredient of the DML estimator called an influence function (IF) and construct a general class of estimators based on the IF. This means that one can estimate any causal functional and enjoy two robustness properties, doubly robustness and debiasedness.
The prevailing framework for solving referring expression grounding is based on a two-stage process: 1) detecting proposals with an object detector and 2) grounding the referent to one of the proposals. Existing two-stage solutions mostly focus on the grounding step, which aims to align the expressions with the proposals.
In this paper, the researchers argue that these methods overlook an obvious mismatch between the roles of proposals in the two stages: they generate proposals solely based on the detection confidence (i.e., expression-agnostic), hoping that the proposals contain all right instances in the expression (i.e., expression-aware). Due to this mismatch, current two-stage methods suffer from a severe performance drop between detected and ground-truth proposals.
The paper proposes Ref-NMS, which is the first method to yield expression-aware proposals at the first stage. Ref-NMS regards all nouns in the expression as critical objects, and introduces a lightweight module to predict a score for aligning each box with a critical object. These scores can guide the NMS operation to filter out the boxes irrelevant to the expression, increasing the recall of critical objects, resulting in a significantly improved grounding performance.
Since RefNMS is agnostic to the grounding step, it can be easily integrated into any state-of-the-art two-stage method. Extensive ablation studies on several backbones, benchmarks, and tasks consistently demonstrate the superiority of Ref-NMS. Codes are available at: https://github.com/ChopinSharp/ref-nms.
Jihye Kwon, a computer engineering PhD student, talks about her research projects and what it took to win a Best Paper award.
Jihye Kwon
What drew you to computer engineering, specifically the application of machine learning to computer-aided design? What questions or issues do you hope to answer?
I was attracted to the concept of a computer: a machine that performs calculations. I found it very interesting how modern computers evolved from executing one instruction at a time to executing many instructions simultaneously by exploiting multiple levels of parallelism. Still, various challenges remained, or newly arose, so I dreamed about designing a brand-new computer system. That is what I had in mind when coming to Columbia.
At the beginning of my PhD, I experimented and learned how to design the core parts of special-purpose computers, using computer-aided design tools. I also explored machine learning from both theoretical and practical perspectives. These activities led me to work on my current research problems.
In advanced computer-aided design of computer systems, computers solve many complex optimization problems in steps to generate a final design. They do so as guided by the designers via means of the configurable ‘knobs’. My focus is on the designers’ work.
For a target system, designers run the computer-aided design tools repeatedly with the many different knob configurations until the tools output final designs with optimal or desired properties, e.g., in timing, area, and power. I wondered if machines can learn, from designers’ previous work, how to configure the knobs to optimize a new target system. Can designers virtually collaborate across time and tasks through the machine learning models? These are the main questions that I hope to answer.
Could you talk about your research and how you collaborated with other groups? Was this something you considered when applying to Columbia – that there are opportunities to do multi-disciplinary work?
When I was applying to Columbia, I wished I could have collaboration opportunities to study and work in the interdisciplinary research communities at the center of New York City. I wanted to explore applications of computer science in different areas to eventually gain insight and inspiration for my own research, which is centered at computer engineering.
Fortunately, these were realized as I worked with my advisor, Professor Luca Carloni. I was invited to join the project “Energy Efficient Computing with Chip-Based Photonics”, which is a part of a large initiative supported by the government and industry. In this project, I worked closely with Lightwave Research Laboratory in Electrical Engineering on a new optical computing system. We proposed the concept of a next-generation computing system that is co-designed with silicon photonics and electronic circuitry, in order to overcome the fundamental and physical limitations of today’s computers.
Another project on optical communication was initiated from a student project that I mentored in my advisor’s class, Embedded Scalable Platforms. This project investigated the use of photonic switches in optically-connected memory systems for deep learning applications.
Outside Columbia, I have also collaborated with researchers at IBM TJ Watson Research Center via my summer internships on the project of auto-tuning computer-aided design flows for commercial supercomputers. All these collaborations opened new horizons for me.
You won the MLCAD 2020 Best Paper award for your research, can you talk about your process – how did the research come about? How long did it take you to complete the work? What were the things you had to overcome?
In this work, I proposed a novel machine learning approach for computer-aided design optimization of hardware accelerators. I wanted to address this problem because it is computationally very expensive to explore the entire optimization space. It took me about one year to complete the work. One of the biggest difficulties I faced was the limited availability of the data for applying machine learning to the problem.
Then, I found out that transfer learning has been recently successfully applied in other areas with limited data. In transfer learning, a model trained for a related problem (e.g., natural image recognition) is transferred to aid the machine learning for the target problem (e.g., face recognition). Hence, I tried to apply transfer learning to my research problem. I trained a neural network model for a different accelerator design, and transferred the model to predict the design properties of a target accelerator.
However, the transferred model did not perform well in this case. I realized that due to the diverse characteristics of the accelerators, I needed to distinguish which piece of the source information should be transferred. Based on this intuition, I constructed a series of new models, and eventually, proposed one with promising performance. While it was a long process of building new models without knowing the answers, my advisor greatly encouraged me in our discussions to keep moving forward, and it was very rewarding in the end.
Looking back, how have you grown as a researcher and a person?
Besides expanding my problem-solving capabilities and technical skills, I have grown to be a better presenter and communicator. One of the tasks of a researcher is to explain one’s work to various groups and different types of audiences. I had a number of opportunities to present my work at academic conferences, seminars at companies, lightning talks, and annual project reviews. Initially, I struggled to meet the audience’s interests whose expertise spans a diverse range of areas and levels. Through those opportunities, I have received very helpful feedback, I have tried to ask myself questions from other people’s perspectives and progressively learned to keep a good balance between abstraction and elaboration.
Also, by interacting with a lot of students with heterogeneous backgrounds in the classes I TA’ed, I have learned to understand what their questions mean and where they come from. Based on that, I tried to adjust my answers to have more relatable conversations. From those conversations, sometimes the students found the topics very interesting, and sometimes I learned something new from them. It was such a great pleasure to inspire others and to be inspired. I think those experiences have made me a better researcher and person.
In Fall 2017, I received an invitation from WiCS’ president, Julia Di, and was impressed by the passionate and caring board members working on the common goal of supporting the advancement of womxn in computer science. In my second year I launched the WiCS Lightning Talks for students with research experience to share their work and stories. The goal was for young students to get to know more about research and demystify the process.
I am one of the few women at Columbia in my research area of computer engineering and would like to contribute to inspiring the next generation to join us.
What was the highlight of your time at Columbia?
Every moment was special for me. Some of the highlights were during happy hour with members of the fishbowl. The fishbowl is a large office occupied by the majority of PhD students in computer engineering. We call it the fishbowl, because it is surrounded by large windows and students inside look like small fishes. Once, my colleagues talked about their memories of old computers that I had never seen. I enjoyed imagining the machines from their descriptions, and thinking about different types and generations of computers.
What is your advice to students on how to navigate their time at Columbia?
Explore, experience, and exploit. There are recommended lists of classes, activities, and companies, depending on your track and interests, but no one is exactly like you. There is such a great variety of opportunities and resources at Columbia and in New York City. I hope you can spend enough time exploring them and get involved in many ways before determining your academic and career goals.
Is there anything else that you think people should know?
Columbia is beautiful in the snow! It gets pretty windy in the winter, so please be aware if you are coming from warmer places. There are many places where you can study but Avery Library is my favorite library on campus. If you have any questions or opinions on this Q&A story, please feel free to drop me a line!
The researchers developed AwareDX – Analysing Women At Risk for Experiencing Drug toXicity – a machine learning algorithm that identifies and predicts differences in adverse drug effects between men and women by analyzing 50 years’ worth of reports in an FDA database. The algorithm automatically corrects for biases in these data that stem from an overrepresentation of male subjects in clinical research trials.
Though men and women can have different responses to medications – the sleep aid Ambien, for example, metabolizes more slowly in women, causing next-day grogginess – doctors may not know about these differences because most clinical trial data itself is biased toward men. This trickles down to impact prescribing guidelines, drug marketing, and ultimately, patients’ health. Unfortunately, pharmaceutical companies have a history of ignoring complex problems and clinical trials have singularly studied men, not even including women. As a result, there is a lot less information about how women respond to drugs compared to men. The research tries to bridge this information gap.
Buy Weed Seeds from premium breeders of the best marijuana strains. Discreetly ships across the USA with free weed seeds in every order.
The Robot Operating System (ROS) is the most popular framework for robotics development. In this paper, the researchers conducted the first major empirical study of ROS, with the goal of understanding how developers collaborate across the many technical disciplines that coalesce in robotics.
Building a complete robot is a difficult task that involves bridging many technical disciplines. ROS aims to simplify development by providing reusable libraries, tools, and conventions for building a robot. Still, as building a robot requires domain expertise in software, mechanical, and electrical engineering, as well as artificial intelligence and robotics, ROS faces knowledge-based barriers to collaboration. The researchers wanted to understand how the necessity of domain-specific knowledge impacts the open-source collaboration model in ROS.
Virtually no one is an expert in every subdomain of robotics: experts who create computer vision packages likely need to rely on software designed by mechanical engineers to implement motor control. As a result, the researchers found that development in ROS is centered around a few unique subgroups each devoted to a different specialty in robotics (i.e. perception, motion). This is unlike other ecosystems, where competing implementations are the norm.
Performance has a major impact on the overall quality of a software project. Performance bugs—bugs that substantially decrease run-time—have long been studied in software engineering, and yet they remain incredibly difficult for developers to handle. In this project, the researchers leveraged contemporary methods in machine learning to create graph embeddings of Python code that can be used to automatically predict performance.
Using un-optimized programming language concepts can lead to performance bugs and the researchers hypothesized that statistical language embeddings could help reveal these patterns. By transforming code samples into graphs that captured the control and data flow of a program, the researchers studied how various unsupervised embeddings of these graphs could be used to predict performance.
Implementing “sort” by hand as opposed to using the built-in Python sort function is an example of a choice that typically slows down a program’s run-time. When the researchers embedded the AST and data flow of a code snippet in Euclidean space (using DeepWalk), patterns like this were captured in the embedding and allowed classifiers to learn which structures are correlated with various levels of performance.
“I was surprised by how often research changes directions,” said Sophia Kolak. In both projects, they started out with one set of questions but answered completely different ones by the end. “It showed me that, in addition to persistence, research requires open-mindedness.”
Yanda Chen Honorable Mention
Cross-language Sentence Selection Via Data Augmentation and Rationale Training Yanda Chen Columbia University, Chris Kedzie Columbia University, Suraj Nair University of Maryland, Petra Galuscakova University of Maryland, Rui Zhang Yale University, Douglas Oard University of Maryland, and Kathleen McKeown Columbia University
In this project, the researchers proposed a new approach to cross-language sentence selection, where they used models to predict sentence-level query relevance with English queries over sentences within document collections in low-resource languages such as Somali, Swahili, and Tagalog.
The system is used as part of cross-lingual information retrieval and query-focused summarization system. For example, if a user puts in a query word “business activity” and specifies Swahili as the language of source documents, then the system will automatically retrieve the Swahili documents that are related to “business activity” and produce short summaries that are then translated from Swahili to English.
A major challenge of the project was the lack of training data for low-resource languages. To tackle this problem, the researchers proposed to generate a relevance dataset of query-sentence pairs through data augmentation based on parallel corpora collected from the web. To mitigate the spurious correlations learned by the model, they proposed the idea of rationale training where they first trained a phrase-based statistical machine translation system and used the alignment information to provide additional supervision for the models.
The approach achieved state-of-the-art results on both text and speech across three languages – Somali, Swahili, and Tagalog.
Dean Boyce's statement on amicus brief filed by President Bollinger
President Bollinger announced that Columbia University along with many other academic institutions (sixteen, including all Ivy League universities) filed an amicus brief in the U.S. District Court for the Eastern District of New York challenging the Executive Order regarding immigrants from seven designated countries and refugees. Among other things, the brief asserts that “safety and security concerns can be addressed in a manner that is consistent with the values America has always stood for, including the free flow of ideas and people across borders and the welcoming of immigrants to our universities.”
This recent action provides a moment for us to collectively reflect on our community within Columbia Engineering and the importance of our commitment to maintaining an open and welcoming community for all students, faculty, researchers and administrative staff. As a School of Engineering and Applied Science, we are fortunate to attract students and faculty from diverse backgrounds, from across the country, and from around the world. It is a great benefit to be able to gather engineers and scientists of so many different perspectives and talents – all with a commitment to learning, a focus on pushing the frontiers of knowledge and discovery, and with a passion for translating our work to impact humanity.
I am proud of our community, and wish to take this opportunity to reinforce our collective commitment to maintaining an open and collegial environment. We are fortunate to have the privilege to learn from one another, and to study, work, and live together in such a dynamic and vibrant place as Columbia.
Sincerely,
Mary C. Boyce
Dean of Engineering
Morris A. and Alma Schapiro Professor
Payal Chandak
Sophia Kolak
Yanda Chen
Add to GMail