Albert Sun, Varun Nair, Elliot Schumacher, Anitha Kannan. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Factual correctness is often the limiting factor in practical applications of natural language generation in high-stakes domains such as healthcare.An essential requirement for maintaining factuality is the ability to deal with rare tokens.This paper focuses on rare tokens that appear in both the source and the reference sequences, and which, when missed during generation, decrease the factual correctness of the output text.For high-stake domains that are also knowledge-rich, we show how to use knowledge to (a) identify which rare tokens that appear in both source and reference are important and (b) uplift their conditional probability.We introduce the "utilization rate" that encodes knowledge and serves as a regularizer by maximizing the marginal probability of selected tokens.We present a study in a knowledgerich domain of healthcare, where we tackle the problem of generating after-visit care instructions based on patient-doctor dialogues.We verify that, in our dataset, specific medical concepts with high utilization rates are underestimated by conventionally trained sequence-tosequence models.We observe that correcting this with our approach to knowledge injection reduces the uncertainty of the model as well as improves factuality and coherence without negatively impacting fluency. 1

People increasingly search online for answers to their medical questions but the rate at which medical questions are asked online significantly exceeds the capacity of qualified people to answer them. This leaves many questions unanswered or inadequately answered. Many of these questions are not unique, and reliable identification of similar questions would enable more efficient and effective question answering schema. COVID-19 has only exacerbated this problem. Almost every government agency and healthcare organization has tried to meet the informational need of users by building online FAQs, but there is no way for people to ask their question and know if it is answered on one of these pages. While many research efforts have focused on the problem of general question similarity, these approaches do not generalize well to domains that require expert knowledge to determine semantic similarity, such as the medical domain. In this paper, we show how a double fine-tuning approach of pretraining a neural network on medical question-answer pairs followed by fine-tuning on medical question-question pairs is a particularly useful intermediate task for the ultimate goal of determining medical question similarity. While other pretraining tasks yield an accuracy below 78.7% on this task, our model achieves an accuracy of 82.6% with the same number of training examples, an accuracy of 80.0% with a much smaller training set, and an accuracy of 84.5% when the full corpus of medical question-answer data is used. We also describe a currently live system that uses the trained model to match user questions to COVID-related FAQs.

Many structured prediction problems (particularly in vision and language domains) are ambiguous, with multiple outputs being ‘correct’ for an input {–} e.g. there are many ways of describing an image, multiple ways of translating a sentence; however, exhaustively annotating the applicability of all possible outputs is intractable due to exponentially large output spaces (e.g. all English sentences). In practice, these problems are cast as multi-class prediction, with the likelihood of only a sparse set of annotations being maximized {–} unfortunately penalizing for placing beliefs on plausible but unannotated outputs. We make and test the following hypothesis {–} for a given input, the annotations of its neighbors may serve as an additional supervisory signal. Specifically, we propose an objective that transfers supervision from neighboring examples. We first study the properties of our developed method in a controlled toy setup before reporting results on multi-label classification and two image-grounded sequence modeling tasks {–} captioning and question generation. We evaluate using standard task-specific metrics and measures of output diversity, finding consistent improvements over standard maximum likelihood training and other baselines.

Interactive agents are becoming increasingly common in many application domains, such as education, healthcare and personal assistance. The success of such embodied agents relies on their ability to have sustained engagement with their human users. Such engagement requires agents to be socially intelligent, equipped with the ability to understand and reciprocate both verbal and non-verbal cues. While there has been tremendous progress in verbal communication, mostly driven by the success of speech recognition and question-answering, teaching agents to appropriately react to facial expressions has received less attention. In this paper, we focus on non-verbal facial cues for face-to-face communication between a user and an embodied agent. We propose a method that automatically learns to update the agent's facial expressions based on the user's expressions. We adopt a learning scheme and train a deep neural network on hundreds of videos, containing pairs of people engaging in a conversation, and without external human supervision. Our experimental results show the efficacy of our model in sustained long-term prediction of the agent's facial landmarks. We present comparative results showing that our model significantly outperforms baseline approaches and provide insightful human studies to better understand our model's qualitative performance. We release our dataset to further encourage research in this field.

We present a novel training framework for neural sequence models, particularly for grounded dialog generation. The standard training paradigm for these models is maximum likelihood estimation (MLE), or minimizing the cross-entropy of the human responses. Across a variety of domains, a recurring problem with MLE trained generative neural dialog models (G) is that they tend to produce 'safe' and generic responses like "I don't know", "I can't tell"). In contrast, discriminative dialog models (D) that are trained to rank a list of candidate human responses outperform their generative counterparts; in terms of automatic metrics, diversity, and informativeness of the responses. However, D is not useful in practice since it can not be deployed to have real conversations with users. Our work aims to achieve the best of both worlds -- the practical usefulness of G and the strong performance of D -- via knowledge transfer from D to G. Our primary contribution is an end-to-end trainable generative visual dialog model, where G receives gradients from D as a perceptual (not adversarial) loss of the sequence sampled from G. We leverage the recently proposed Gumbel-Softmax (GS) approximation to the discrete distribution -- specifically, a RNN is augmented with a sequence of GS samplers, which coupled with the straight-through gradient estimator enables end-to-end differentiability. We also introduce a stronger encoder for visual dialog, and employ a self-attention mechanism for answer encoding along with a metric learning loss to aid D in better capturing semantic similarities in answer responses. Overall, our proposed model outperforms state-of-the-art on the VisDial dataset by a significant margin (2.67% on recall@10). The source code can be downloaded from https://github.com/jiasenlu/visDial.pytorch

The rapid proliferation of hand-held devices has led to the development of rich, interactive and immersive applications, such as e-readers for electronic books. These applications motivate retrieval systems that can implicitly satisfy any information need of the reader by exploiting the context of the user's interactions. Such retrieval systems differ from traditional search engines in that the queries constructed using the context are typically complex objects (including the document and its structure).

Recent studies have shown the importance of using online videos along with textual material in educational instruction, especially for better content retention and improved concept understanding. A key question is how to select videos to maximize student engagement, particularly when there are multiple possible videos on the same topic. While there are many aspects that drive student engagement, in this paper we focus on presenter speaking styles in the video. We use crowd-sourcing to explore speaking style dimensions in online educational videos, and identify six broad dimensions: liveliness, speaking rate, pleasantness, clarity, formality and confidence. We then propose techniques based solely on acoustic features for automatically identifying a subset of the dimensions. Finally, we perform video re-ranking experiments to learn how users apply their speaking style preferences to augment textbook material. Our findings also indicate how certain dimensions are correlated with perceptions of general pleasantness of the voice.

Images are often used to convey many different concepts or illustrate many different stories. We propose an algorithm to mine multiple diverse, relevant, and interesting text snippets for images on the web. Our algorithm scales to all images on the web. For each image, all webpages that contain it are considered. The top-K text snippet selection problem is posed as combinatorial subset selection with the goal of choosing an optimal set of snippets that maximizes a combination of relevancy, interestingness, and diversity. The relevancy and interestingness are scored by machine learned models. Our algorithm is run at scale on the entire image index of a major search engine resulting in the construction of a database of images with their corresponding text snippets. We validate the quality of the database through a large-scale comparative study. We showcase the utility of the database through two web-scale applications: (a) augmentation of images on the web as webpages are browsed and (b)~an image browsing experience (similar in spirit to web browsing) that is enabled by interconnecting semantically related images (which may not be visually related) through shared concepts in their corresponding text snippets.

The emergence of tablet devices, cloud computing, and abundant online multimedia content presents new opportunities to transform traditional paper-based textbooks into tablet-based electronic textbooks, and to further augment the educational experience by enriching them with relevant supplementary materials. Given a candidate set of relevant educational videos for augmenting an electronic textbook, how do we assign the videos at the appropriate granularity (a collection of logical units in the book)? We propose a rigorous formulation of the video assignment problem and present an algorithm for assigning each video to the optimum subset of logical units. Our experimental evaluation using a diverse collection of educational videos relevant to multiple chapters in a textbook demonstrates the efficacy of the proposed techniques for inferring the granularity at which a relevant video should be assigned.

AutoCaption is a system that helps a smartphone user generate a caption for their photos. It operates by uploading the photo to a cloud service where a number of parallel modules are applied to recognize a variety of entities and relations. The outputs of the modules are combined to generate a large set of candidate captions, which are returned to the phone. The phone client includes a convenient user interface that allows users to select their favorite caption, reorder, add, or delete words to obtain the grammatical style they prefer. The user can also select from multiple candidates returned by the recognition modules.

We present study navigator, an algorithmically-generated aid for enhancing the experience of studying from electronic textbooks. The study navigator for a section of the book consists of helpful concept references for understanding this section. Each concept reference is a pair consisting of a concept phrase explained elsewhere and the link to the section in which it has been explained. We propose a novel reader model for textbooks and an algorithm for generating the study navigator based on this model. We also present the results of an extensive user study that demonstrates the efficacy of the proposed system across textbooks on different subjects from different grades.

Advertisers typically have a fairly accurate idea of the interests of their target audience. However, today's online advertising systems are unable to leverage this information. The reasons are two-fold. First, there is no agreed upon vocabulary of interests for advertisers and advertising systems to communicate. More importantly, advertising systems lack a mechanism for mapping users to the interest vocabulary.

Recent work in commerce search has shown that understanding the semantics in user queries enables more effective query analysis and retrieval of relevant products. However, due to lack of sufficient domain knowledge, user queries often include terms that cannot be mapped directly to any product attribute. For example, a user looking for designer handbags might start with such a query because she is not familiar with the manufacturers, the price ranges, and/or the material that gives a handbag designer appeal. Current commerce search engines treat terms such as designer as keywords and attempt to match them to contents such as product reviews and product descriptions, often resulting in poor user experience.

Proxemics is the study of how people interact. We present a computational formulation of visual proxemics by attempting to label each pair of people in an image with a subset of physically based “touch codes.” A baseline approach would be to first perform pose estimation and then detect the touch codes based on the estimated joint locations. We found that this sequential approach does not perform well because pose estimation step is too unreliable for images of interacting people, due to difficulties with occlusion and limb ambiguities. Instead, we propose a direct approach where we build an articulated model tuned for each touch code. Each such model contains two people, connected in an appropriate manner for the touch code in question. We fit this model to the image and then base classification on the fitting error. Experiments show that this approach significantly outperforms the sequential baseline as well as other related approches.

Good textbooks are organized in a systematically progressive fashion so that students acquire new knowledge and learn new concepts based on known items of information. We provide a diagnostic tool for quantitatively assessing the comprehension burden that a textbook imposes on the reader due to non-sequential presentation of concepts. We present a formal definition of comprehension burden and propose an algorithmic approach for computing it. We apply the tool to a corpus of high school textbooks from India and empirically examine its effectiveness in helping authors identify sections of textbooks that can benefit from reorganizing the material presented.

We introduce an entity-centric search experience, called Active Objects, in which entity-bearing queries are paired with actions that can be performed on the entities. For example, given a query for a specific flashlight, we aim to present actions such as reading reviews, watching demo videos, and finding the best price online. In an annotation study conducted over a random sample of user query sessions, we found that a large proportion of queries in query logs involve actions on entities, calling for an automatic approach to identifying relevant actions for entity-bearing queries. In this paper, we pose the problem of finding actions that can be performed on entities as the problem of probabilistic inference in a graphical model that captures how an entity bearing query is generated. We design models of increasing complexity that capture latent factors such as entity type and intended actions that determine how a user writes a query in a search box, and the URL that they click on. Given a large collection of real-world queries and clicks from a commercial search engine, the models are learned efficiently through maximum likelihood estimation using an EM algorithm. Given a new query, probabilistic inference enables recommendation of a set of pertinent actions and hosts. We propose an evaluation methodology for measuring the relevance of our recommended actions, and show empirical evidence of the quality and the diversity of the discovered actions.

Textbooks have a direct bearing on the quality of education imparted to the students. Therefore, it is of paramount importance that the educational content of textbooks should provide rich learning experience to the students. Recent studies on understanding learning behavior suggest that the incorporation of digital visual material can greatly enhance learning. However, textbooks used in many developing regions are largely text-oriented and lack good visual material. We propose techniques for finding images from the web that are most relevant for augmenting a section of the textbook, while respecting the constraint that the same image is not repeated in different sections of the same chapter. We devise a rigorous formulation of the image assignment problem and present a polynomial time algorithm for solving the problem optimally. We also present two image mining algorithms that utilize orthogonal signals and hence obtain different sets of relevant images. Finally, we provide an ensembling algorithm for combining the assignments. To empirically evaluate our techniques, we use a corpus of high school textbooks in use in India. Our user study utilizing the Amazon Mechanical Turk platform indicates that the proposed techniques are able to obtain images that can help increase the understanding of the textbook material.

Product classification in Commerce search (e.g., Google Product Search, Bing Shopping) involves associating categories to offers of products from a large number of merchants. The categorized offers are used in many tasks including product taxonomy browsing and matching merchant offers to products in the catalog. Hence, learning a product classifier with high precision and recall is of fundamental importance in order to provide high quality shopping experience. A product offer typically consists of a short textual description and an image depicting the product. Traditional approaches to this classification task is to learn a classifier using only the textual descriptions of the products. In this paper, we show that the use of images, a weaker signal in our setting, in conjunction with the textual descriptions, a more discriminative signal, can considerably improve the precision of the classification task, irrespective of the type of classifier being used. We present a novel classification approach, Confusion Driven Probabilistic Fusion++ (CDPF++), that is cognizant of the disparity in the discriminative power of different types of signals and hence makes use of the confusion matrix of dominant signal (text in our setting) to prudently leverage the weaker signal (image), for an improved performance. Our evaluation performed on data from a major Commerce search engine's catalog shows a 12% (absolute) improvement in precision at 100% coverage, and a 16% (absolute) improvement in recall at 90% precision compared to classifiers that only use textual description of products. In addition, CDPF++ also provides a more accurate classifier based only on the dominant signal (text) that can be used in situations in which only the dominant signal is available during application time.

An e-commerce catalog typically comprises of specifications for millions of products. The search engine receives millions of sales offers from thousands of independent merchants that must be matched to the right products. We describe the challenges that a system for matching unstructured offers to structured product descriptions must address, drawing upon our experience from building such a system for Bing Shopping. The heart of our system is a data-driven component that learns the matching function off-line, which is then applied at run-time for matching offers to products. We provide the design of this and other critical components of the system as well as the details of the extensive experiments we performed to assess the readiness of the system. This system is currently deployed in an experimental Commerce Search Engine and is used to match all the offers received by Bing Shopping to the Bing product catalog.

Classification is a core task in knowledge discovery and data mining, and there has been substantial research effort in developing sophisticated classification models. In a parallel thread, recent work from the NLP community suggests that for tasks such as natural language disambiguation even a simple algorithm can outperform a sophisticated one, if it is provided with large quantities of high quality training data. In those applications, training data occurs naturally in text corpora, and high quality training data sets running into billions of words have been reportedly used.

Given a terabyte click log, can we build an efficient and effective click model? It is commonly believed that web search click logs are a gold mine for search business, because they reflect users' preference over web documents presented by the search engine. Click models provide a principled approach to inferring user-perceived relevance of web documents, which can be leveraged in numerous applications in search businesses. Due to the huge volume of click data, scalability is a must.We present the click chain model (CCM), which is based on a solid, Bayesian framework. It is both scalable and incremental, perfectly meeting the computational challenges imposed by the voluminous click logs that constantly grow. We conduct an extensive experimental study on a data set containing 8.8 million query sessions obtained in July 2008 from a commercial search engine. CCM consistently outperforms two state-of-the-art competitors in a number of metrics, with over 9.7% better log-likelihood, over 6.2% better click perplexity and much more robust (up to 30%) prediction of the first and the last clicked position.

This paper presents a novel method for location recognition, which exploits an epitomic representation to achieve both high efficiency and good generalization. A generative model based on epitomic image analysis captures the appearance and geometric structure of an environment while allowing for variations due to motion, occlusions, and non-Lambertian effects. The ability to model translation and scale invariance together with the fusion of diverse visual features yields enhanced generalization with economical training. Experiments on both existing and new labeled image databases result in recognition accuracy superior to state of the art with real-time computational performance.

This paper presents a novel method for location recognition, which exploits an epitomic representation to achieve both high efficiency and good generalization. A generative model based on epitomic image analysis captures the appearance and geometric structure of an environment while allowing for variations due to motion, occlusions and non-Lambertian effects. The ability to model translation and scale invariance together with the fusion of diverse visual features yield enhanced generalization with economical training. Experiments on both existing and new labelled image databases result in recognition accuracy superior to state of the art with real-time computational performance.

A jigsaw is a recently proposed generative model that describes an image as a composition of non-overlapping patches of varying shape, extracted from a latent image. By learning the latent jigsaw image which best explains a set of images, it is possible to discover the shape, size and appearance of repeated structures in the images. A challenge when learning this model is the very large space of possible jigsaw pixels which can potentially be used to explain each image pixel. The previous method of inference for this model scales linearly with the number of jigsaw pixels, making it unusable for learning the large jigsaws needed for many practical applications. In this paper, we make three contributions that enable the learning of large jigsaws -a novel sparse belief propagation algorithm, a hybrid method which significantly improves the sparseness of this algorithm, and a method that uses these techniques to make learning of large jigsaws feasible. We provide detailed analysis of how our hybrid inference method leads to significant savings in memory and computation time. To demonstrate the success of our method, we present experimental results applying large jigsaws to an object recognition task.

Patch-based appearance models are used in a wide range of computer vision ap- plications. To learn such models it has previously been necessary to specify a suitable set of patch sizes and shapes by hand. In the jigsaw model presented here, the shape, size and appearance of patches are learned automatically from the repeated structures in a set of training images. By learning such irregularly shaped ‘jigsaw pieces’, we are able to discover both the shape and the appearance of object parts without supervision. When applied to face images, for example, the learned jigsaw pieces are surprisingly strongly associated with face parts of different shapes and scales such as eyes, noses, eyebrows and cheeks, to name a few. We conclude that learning the shape of the patch not only improves the accuracy of appearance-based part detection but also allows for shape-based part detection. This enables parts of similar appearance but different shapes to be dis- tinguished; for example, while foreheads and cheeks are both skin colored, they have markedly different shapes.

By mapping a set of input images to points in a low-dimensional manifold or subspace, it is possible to efficiently account for a small number of degrees of freedom. For example, images of a person walking can be mapped to a one-dimensional manifold that measures the phase of the person's gait. However, when the object is moving around the frame and being occluded by other objects, standard manifold modeling techniques (e.g., principal components analysis, factor analysis, locally linear embedding) try to account for global motion and occlusion. We show how factor analysis can be incorporated into a generative model of layered, 2.5-dimensional vision, to jointly locate objects, resolve occlusion ambiguities, and learn models of the appearance manifolds of objects. We demonstrate the algorithm on a video consisting of four occluding objects, two of which are people who are walking, and occlude each other for most of the duration of the video. Whereas standard manifold modeling techniques fail to extract information about the gaits, the layered model successfully extracts a periodic representation of the gait of each person.

We present novel simple appearance and shape models that we call epitomes. The epitome of an image is its miniature, condensed version containing the essence of the textural and shape properties of the image. As opposed to previously used simple image models, such as templates or basis functions, the size of the epitome is considerably smaller than the size of the image or object it represents, but the epitome still contains most constitutive elements needed to reconstruct the image. A collection of images often shares an epitome, e.g., when images are a few consecutive frames from a video sequence, or when they are photographs of similar objects. A particular image in a collection is defined by its epitome and a smooth mapping from the epitome to the image pixels. When the epitomic representation is used within a hierarchical generative model, appropriate inference algorithms can be derived to extract the epitome from a single image or a collection of images and at the same time perform various inference tasks, such as image segmentation, motion estimation, object removal and super-resolution.

Dimensionality reduction techniques such as principal component analy- sis and factor analysis are used to discover a linear mapping between high dimensional data samples and points in a lower dimensional subspace. In [6], Jojic and Frey introduced mixture of transformation-invariant component analyzers (MTCA) that can account for global transforma- tions such as translations and rotations, perform clustering and learn lo- cal appearance deformations by dimensionality reduction. However, due to enormous computational requirements of the EM algorithm for learn- ing the model, O( is the dimensionality of a data sample, MTCA was not practical for most applications. In this paper, we demon- strate how fast Fourier transforms can reduce the computation to the or- . With this speedup, we show the effectiveness of MTCA der of in various applications - tracking, video textures, clustering video se- quences, object recognition, and object detection in images.

Factor analysis and principal components analysis can be used to model linear relationships between observed variables and linearly map high-dimensional data to a lower-dimensional hidden space. In factor analysis, the observations are modeled as a linear com(cid:173) bination of normally distributed hidden variables. We describe a nonlinear generalization of factor analysis, called "product analy(cid:173) sis", that models the observed variables as a linear combination of products of normally distributed hidden variables. Just as fac(cid:173) tor analysis can be viewed as unsupervised linear regression on unobserved, normally distributed hidden variables, product anal(cid:173) ysis can be viewed as unsupervised linear regression on products of unobserved, normally distributed hidden variables. The map(cid:173) ping between the data and the hidden space is nonlinear, so we use an approximate variational technique for inference and learn(cid:173) ing. Since product analysis is a generalization of factor analysis, product analysis always finds a higher data likelihood than factor analysis. We give results on pattern recognition and illumination(cid:173) invariant image clustering.

One way to approximate inference in richly-connected graphical models is to apply the sum-product algorithm (a.k.a. probabil(cid:173) ity propagation algorithm), while ignoring the fact that the graph has cycles. The sum-product algorithm can be directly applied in Gaussian networks and in graphs for coding, but for many condi(cid:173) tional probability functions - including the sigmoid function - di(cid:173) rect application of the sum-product algorithm is not possible. We introduce "accumulator networks" that have low local complexity (but exponential global complexity) so the sum-product algorithm can be directly applied. In an accumulator network, the probability of a child given its parents is computed by accumulating the inputs from the parents in a Markov chain or more generally a tree. After giving expressions for inference and learning in accumulator net(cid:173) works, we give results on the "bars problem" and on the problem of extracting translated, overlapping faces from an image.