A few large, homogenous, pre-trained models undergird many machine learning systems -and often, these models contain harmful stereotypes learned from the internet.We investigate the bias transfer hypothesis: the theory that social biases (such as stereotypes) internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning.For two classification tasks, we find that reducing intrinsic bias with controlled interventions before finetuning does little to mitigate the classifier's discriminatory behavior after fine-tuning.Regression analysis suggests that downstream disparities are better explained by biases in the fine-tuning dataset.Still, pre-training plays a role: simple alterations to co-occurrence rates in the fine-tuning dataset are ineffective when the model has been pre-trained.Our results encourage practitioners to focus more on dataset quality and context-specific harms.

In this paper, we propose conjugate energy-based models (CEBMs), a new class of energy-based models that define a joint density over data and latent variables. The joint density of a CEBM decomposes into an intractable distribution over data and a tractable posterior over latent variables. CEBMs have similar use cases as variational autoencoders, in the sense that they learn an unsupervised mapping from data to latent variables. However, these models omit a generator network, which allows them to learn more flexible notions of similarity between data points. Our experiments demonstrate that conjugate EBMs achieve competitive results in terms of image modelling, predictive power of latent space, and out-of-domain detection on a variety of datasets.

We consider the problem of utility maximization in online ranking applications while also satisfying a pre-defined fairness constraint. We consider batches of items which arrive over time, already ranked using an existing ranking model. We propose online post-processing for re-ranking these batches to enforce adherence to the pre-defined fairness constraint, while maximizing a specific notion of utility. To achieve this goal, we propose two deterministic re-ranking policies. In addition, we learn a re-ranking policy based on a novel variation of learning to search. Extensive experiments on real world and synthetic datasets demonstrate the effectiveness of our proposed policies both in terms of adherence to the fairness constraint and utility maximization. Furthermore, our analysis shows that the performance of the proposed policies depends on the original data distribution w.r.t the fairness constraint and the notion of utility.

Practitioners apply neural networks to increasingly complex problems in natural language processing, such as syntactic parsing and semantic role labeling that have rich output structures. Many such structured-prediction problems require deterministic constraints on the output values; for example, in sequence-to-sequence syntactic parsing, we require that the sequential outputs encode valid trees. While hidden units might capture such properties, the network is not always able to learn such constraints from the training data alone, and practitioners must then resort to post-processing. In this paper, we present an inference method for neural networks that enforces deterministic constraints on outputs without performing rule-based post-processing or expensive discrete search. Instead, in the spirit of gradient-based training, we enforce constraints with gradient-based inference (GBI): for each input at test-time, we nudge continuous model weights until the network’s unconstrained inference procedure generates an output that satisfies the constraints. We study the efficacy of GBI on three tasks with hard constraints: semantic role labeling, syntactic parsing, and sequence transduction. In each case, the algorithm not only satisfies constraints, but improves accuracy, even when the underlying network is stateof-the-art.

Recent work has explored transforming data sets into smaller, approximate summaries in order to scale Bayesian inference. We examine a related problem in which the parameters of a Bayesian model are very large and expensive to store in memory, and propose more compact representations of parameter values that can be used during inference. We focus on a class of graphical models that we refer to as latent Dirichlet-Categorical models, and show how a combination of two sketching algorithms known as count-min sketch and approximate counters provide an efficient representation for them. We show that this sketch combination – which, despite having been used before in NLP applications, has not been previously analyzed – enjoys desirable properties. We prove that for this class of models, when the sketches are used during Markov Chain Monte Carlo inference, the equilibrium of sketched MCMC converges to that of the exact chain as sketch parameters are tuned to reduce the error rate.

The prevailing wisdom is that a model's fairness and its accuracy are in tension with one another. However, there is a pernicious {\em modeling-evaluating dualism} bedeviling fair machine learning in which phenomena such as label bias are appropriately acknowledged as a source of unfairness when designing fair models, only to be tacitly abandoned when evaluating them. We investigate fairness and accuracy, but this time under a variety of controlled conditions in which we vary the amount and type of bias. We find, under reasonable assumptions, that the tension between fairness and accuracy is illusive, and vanishes as soon as we account for these phenomena during evaluation. Moreover, our results are consistent with an opposing conclusion: fairness and accuracy are sometimes in accord. This raises the question, {\em might there be a way to harness fairness to improve accuracy after all?} Since most notions of fairness are with respect to the model's predictions and not the ground truth labels, this provides an opportunity to see if we can improve accuracy by harnessing appropriate notions of fairness over large quantities of {\em unlabeled} data with techniques like posterior regularization and generalized expectation. Indeed, we find that semi-supervision not only improves fairness, but also accuracy and has advantages over existing in-processing methods that succumb to selection bias on the training set.

We present a method that consumes a large corpus of multilingual text and produces a single, unified word embedding in which the word vectors generalize across languages. In contrast to current approaches that require language identification, our method is agnostic about the languages with which the documents in the corpus are expressed, and does not rely on parallel corpora to constrain the spaces. Instead we utilize a small set of human provided word translations---which are often freely and readily available. We can encode such word translations as hard constraints in the model's objective functions; however, we find that we can more naturally constrain the space by allowing words in one language to borrow distributional statistics from context words in another language. We achieve this via a process we term artificial code-switching. As the name suggests, we induce code-switching so that words across multiple languages appear in contexts together. Not only do embedding models trained on code-switched data learn common cross-lingual structure, the common structure allows an NLP model trained in a source language to generalize to multiple target languages (achieving up to 80% of the accuracy of models trained with target-language data).

We propose an embarrassingly parallel, memory efficient inference algorithm for latent variable models in which the complete data likelihood is in the exponential family. The algorithm is a stochastic cellular automaton and converges to a valid maximum a posteriori fixed point. Applied to latent Dirichlet allocation we find that our algorithm is over an order or magnitude faster than the fastest current approaches. A simple C++/MPI implementation on a 20-node Amazon EC2 cluster samples at more than 1 billion tokens per second. We process 3 billion documents and achieve predictive power competitive with collapsed Gibbs sampling and variational inference.

Traditional approaches to probabilistic inference such as loopy belief propagation and Gibbs sampling typically compute marginals for it all the unobserved variables in a graphical model. However, in many real-world applications the user's interests are focused on a subset of the variables, specified by a query. In this case it would be wasteful to uniformly sample, say, one million variables when the query concerns only ten. In this paper we propose a query-specific approach to MCMC that accounts for the query variables and their generalized mutual information with neighboring variables in order to achieve higher computational efficiency. Surprisingly there has been almost no previous work on query-aware MCMC. We demonstrate the success of our approach with positive experimental results on a wide range of graphical models.

In the database community, work on information extraction (IE) has centered on two themes: how to effectively manage IE tasks, and how to manage the uncertainties that arise in the IE process in a scalable manner. Recent work has proposed a probabilistic database (PDB) based declarative IE system that supports a leading statistical IE model, and an associated inference algorithm to answer top-k-style queries over the probabilistic IE outcome. Still, the broader problem of effectively supporting general probabilistic inference inside a PDB-based declarative IE system remains open. In this paper, we explore the in-database implementations of a wide variety of inference algorithms suited to IE, including two Markov chain Monte Carlo algorithms, the Viterbi and the sum-product algorithms. We describe the rules for choosing appropriate inference algorithms based on the model, the query and the text, considering the trade-off between accuracy and runtime. Based on these rules, we describe a hybrid approach to optimize the execution of a single probabilistic IE query to employ different inference algorithms appropriate for different records. We show that our techniques can achieve up to 10-fold speedups compared to the non-hybrid solutions proposed in the literature.

Incorporating probabilities into the semantics of incomplete databases has posed many challenges, forcing systems to sacrifice modeling power, scalability, or treatment of relational algebra operators. We propose an alternative approach where the underlying relational database always represents a single world, and an external factor graph encodes a distribution over possible worlds; Markov chain Monte Carlo (MCMC) inference is then used to recover this uncertainty to a desired level of fidelity. Our approach allows the efficient evaluation of arbitrary queries over probabilistic databases with arbitrary dependencies expressed by graphical models with structure that changes during inference. MCMC sampling provides efficiency by hypothesizing modifications to possible worlds rather than generating entire worlds from scratch. Queries are then run over the portions of the world that change, avoiding the onerous cost of running full queries over each sampled world. A significant innovation of this work is the connection between MCMC sampling and materialized view maintenance techniques: we find empirically that using view maintenance techniques is several orders of magnitude faster than naively querying each sampled world. We also demonstrate our system's ability to answer relational queries with aggregation, and demonstrate additional scalability through the use of parallelization on a real-world complex model of information extraction. This framework is sufficiently expressive to support probabilistic inference not only for answering queries, but also for inferring missing database content from raw evidence.

Large, relational factor graphs with structure defined by first-order logic or other languages give rise to notoriously difficult inference problems. Because unrolling the structure necessary to represent distributions over all hypotheses has exponential blow-up, solutions are often derived from MCMC. However, because of limitations in the design and parameterization of the jump function, these sampling-based methods suffer from local minima|the system must transition through lower-scoring configurations before arriving at a better MAP solution. This paper presents a new method of explicitly selecting fruitful downward jumps by leveraging reinforcement learning (RL). Rather than setting parameters to maximize the likelihood of the training data, parameters of the factor graph are treated as a log-linear function approximator and learned with temporal difference (TD); MAP inference is performed by executing the resulting policy on held out test data. Our method allows efficient gradient updates since only factors in the neighborhood of variables affected by an action need to be computed|we bypass the need to compute marginals entirely. Our method provides dramatic empirical success, producing new state-of-the-art results on a complex joint model of ontology alignment, with a 48\% reduction in error over state-of-the-art in that domain.

Recently, many advanced machine learning approaches have been proposed for coreference resolution; however, all of the discriminatively-trained models reason over mentions rather than entities. That is, they do not explicitly contain variables indicating the “canonical” values for each attribute of an entity (e.g., name, venue, title, etc.). This canonicalization step is typically implemented as a post-processing routine to coreference resolution prior to adding the extracted entity to a database. In this paper, we propose a discriminatively-trained model that jointly performs coreference resolution and canonicalization, enabling features over hypothesized entities. We validate our approach on two different coreference problems: newswire anaphora resolution and research paper citation matching, demonstrating improvements in both tasks and achieving an error reduction of up to 62% when compared to a method that reasons about mentions only.

The automatic consolidation of database records from many heterogeneous sources into a single repository requires solving several information integration tasks. Although tasks such as coreference, schema matching, and canonicalization are closely related, they are most commonly studied in isolation. Systems that do tackle multiple integration problems traditionally solve each independently, allowing errors to propagate from one task to another. In this paper, we describe a discriminatively-trained model that reasons about schema matching, coreference, and canonicalization jointly. We evaluate our model on a real-world data set of people and demonstrate that simultaneously solving these tasks reduces errors over a cascaded or isolated approach. Our experiments show that a joint model is able to improve substantially over systems that either solve each task in isolation or with the conventional cascade. We demonstrate nearly a 50% error reduction for coreference and a 40% error reduction for schema matching.

Modern optical, character recognition software relies on human interaction to correct mis recognized characters. Even though the software often reliably identifies low-confidence output, the simple language and vocabulary models employed are insufficient to automatically correct mistakes. This paper demonstrates that topic models, which automatically detect and represent an article's semantic context, reduces error by 7% over a global word distribution in a simulated OCR correction task. Detecting and leveraging context in this manner is an important step towards improving OCR.

It is becoming increasingly common to construct databases from information automatically culled from many heterogeneous sources. For example, a research publication database can be constructed by automatically extracting titles, authors, and conference information from online papers. A common difficulty in consolidating data from multiple sources is that records are referenced in a variety of ways (e.g. abbreviations, aliases, and misspellings). Therefore, it can be difficult to construct a single, standard representation to present to the user. We refer to the task of constructing this representation as canonicalization. Despite its importance, there is little existing work on canonicalization.