This paper presents Instruct-Imagen, a model that tackles heterogeneous image generation tasks and generalizes across unseen tasks. We introduce multi-modal in-struction for image generation, a task representation artic-ulating a range of generation intents with precision. It uses natural language to amalgamate disparate modalities (e.g., text, edge, style, subject, etc.), such that abundant generation intents can be standardized in a uniform format. We then build Instruct - Imagen by fine-tuning a pre-trained text-to-image diffusion model with two stages. First, we adapt the model using the retrieval-augmented training, to enhance model's capabilities to ground its generation on external multi-modal context. Subsequently, we fine-tune the adapted model on diverse image generation tasks that requires vision-language understanding (e.g., subject-driven generation, etc.), each paired with a multi-modal instruction encapsulating the task's essence. Human evaluation on various image generation datasets re-veals that Instruct-Imagen matches or surpasses prior task-specific models in-domain and demonstrates promising generalization to unseen and more complex tasks. Our evaluation suite will be made publicly available.

Tal Schuster, Adam Lelkes, Haitian Sun, Jai Gupta, Jonathan Berant, William Cohen, Donald Metzler. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Prediction-powered inference (PPI) is a method that improves statistical estimates based on limited human-labeled data. PPI achieves this by combining small amounts of human-labeled data with larger amounts of data labeled by a reasonably accurate---but potentially biased---automatic system, in a way that results in tighter confidence intervals for certain parameters of interest (e.g., the mean performance of a language model). In this paper, we propose a method called Stratified Prediction-Powered Inference (StratPPI), in which we show that the basic PPI estimates can be considerably improved by employing simple data stratification strategies. Without making any assumptions on the underlying automatic labeling system or data distribution, we derive an algorithm for computing provably valid confidence intervals for parameters of any dimensionality that is based on stratified sampling. In particular, we show both theoretically and empirically that, with appropriate choices of stratification and sample allocation, our approach can provide substantially tighter confidence intervals than unstratified approaches. Specifically, StratPPI is expected to improve in cases where the performance of the autorater varies across different conditional distributions of the target data.

Large-scale generative language and vision-language models (LLMs and VLMs) excel in few-shot in-context learning for decision making and instruction following. However, they require high-quality exemplar demonstrations to be included in their context window. In this work, we ask: Can LLMs and VLMs generate their own examples from generic, sub-optimal demonstrations? We propose In-Context Abstraction Learning (ICAL), a method that builds a memory of multimodal experience from sub-optimal demonstrations and human feedback. Given a task demonstration that may contain inefficiencies or mistakes, a VLM abstracts the trajectory into a generalized program by correcting inefficient actions and annotating cognitive abstractions: causal relationships, object state changes, temporal subgoals, and task-relevant visual elements. These abstractions are iteratively improved and adapted through human feedback while the agent attempts to execute the trajectory in a similar environment. The resulting examples, when used as exemplars in the prompt, significantly improve decision-making in retrieval-augmented LLM and VLM agents. Moreover, as the agent's library of examples grows, it becomes more efficient, relying less on human feedback and requiring fewer environment interactions per demonstration. Our ICAL agent surpasses the state-of-the-art in dialogue-based instruction following in TEACh, multimodal web agents in VisualWebArena, and action anticipation in Ego4D. In TEACh, we achieve a 12.6% improvement in goal-condition success. In VisualWebArena, our task success rate improves over the SOTA from 14.3% to 22.7% using GPT4V. In Ego4D action forecasting, we improve over few-shot GPT-4V and remain competitive with supervised models. We show finetuning our retrieval-augmented in-context agent yields additional improvements. Our approach significantly reduces reliance on manual prompt engineering and consistently outperforms in-context learning from action plans that lack such abstractions.

We propose a new knowledge representation (KR) based on knowledge bases (KBs) derived from text, based on question generation and entity linking. We argue that the proposed type of KB has many of the key advantages of a traditional symbolic KB: in particular, it consists of small modular components, which can be combined compositionally to answer complex queries, including relational queries and queries involving ``multi-hop'' inferences. However, unlike a traditional KB, this information store is well-aligned with common user information needs. We present one such KB, called a QEDB, and give qualitative evidence that the atomic components are high-quality and meaningful, and that atomic components can be combined in ways similar to the triples in a symbolic KB. We also show experimentally that questions reflective of typical user questions are more easily answered with a QEDB than a symbolic KB.

John Wieting, Jonathan Clark, William Cohen, Graham Neubig, Taylor Berg-Kirkpatrick. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Existing state-of-the-art methods for opendomain question-answering (ODQA) use anopen book approach in which information is first retrieved from a large text corpus or knowledge base (KB) and then reasoned over to produce an answer.A recent alternative is to retrieve from a collection of previouslygenerated question-answer pairs; this has several practical advantages including being more memory and compute-efficient.Questionanswer pairs are also appealing in that they can be viewed as an intermediate between text and KB triples: like KB triples, they often concisely express a single relationship, but like text, have much higher coverage than traditional KBs.In this work, we describe a new QA system that augments a text-to-text model with a large memory of question-answer pairs, and a new pre-training task for the latent step of question retrieval.The pre-training task substantially simplifies training and greatly improves performance on smaller QA benchmarks.Unlike prior systems of this sort, our QA system can also answer multi-hop questions that do not explicitly appear in the collection of stored question-answer pairs.

We introduce a new in-context learning paradigm to measure Large Language Models' (LLMs) ability to learn novel words during inference.In particular, we rewrite Winogradstyle co-reference resolution problems by replacing the key concept word with a synthetic but plausible word that the model must understand to complete the task.Solving this task requires the model to make use of the dictionary definition of the new word given in the prompt.This benchmark addresses word acquisition, one important aspect of the diachronic degradation known to afflict LLMs.As LLMs are frozen in time at the moment they are trained, they are normally unable to reflect the way language changes over time.We show that the accuracy of LLMs compared to the original Winograd tasks decreases radically in our benchmark, thus identifying a limitation of current models and providing a benchmark to measure future improvements in LLMs ability to do in-context learning.

Retrieval-augmented language models such as Fusion-in-Decoder are powerful, setting the state of the art on a variety of knowledge-intensive tasks. However, they are also expensive, due to the need to encode a large number of retrieved passages. Some work avoids this cost by pre-encoding a text corpus into a memory and retrieving dense representations directly. However, pre-encoding memory incurs a severe quality penalty as the memory representations are not conditioned on the current input. We propose LUMEN, a hybrid between these two extremes, pre-computing the majority of the retrieval representation and completing the encoding on the fly using a live encoder that is conditioned on the question and fine-tuned for the task. We show that LUMEN significantly outperforms pure memory on multiple question-answering tasks while being much cheaper than FiD, and outperforms both for any given compute budget. Moreover, the advantage of LUMEN over FiD increases with model size.

Recent text-to-image generation models like DreamBooth have made remarkable progress in generating highly customized images of a target subject, by fine-tuning an ``expert model'' for a given subject from a few examples. However, this process is expensive, since a new expert model must be learned for each subject. In this paper, we present SuTI, a Subject-driven Text-to-Image generator that replaces subject-specific fine tuning with {in-context} learning. Given a few demonstrations of a new subject, SuTI can instantly generate novel renditions of the subject in different scenes, without any subject-specific optimization. SuTI is powered by {apprenticeship learning}, where a single apprentice model is learned from data generated by a massive number of subject-specific expert models. Specifically, we mine millions of image clusters from the Internet, each centered around a specific visual subject. We adopt these clusters to train a massive number of expert models, each specializing in a different subject. The apprentice model SuTI then learns to imitate the behavior of these fine-tuned experts. SuTI can generate high-quality and customized subject-specific images 20x faster than optimization-based SoTA methods. On the challenging DreamBench and DreamBench-v2, our human evaluation shows that SuTI significantly outperforms existing models like InstructPix2Pix, Textual Inversion, Imagic, Prompt2Prompt, Re-Imagen and DreamBooth.

In attempts to "explain" predictions of machine learning models, researchers have proposed hundreds of techniques for attributing predictions to features that are deemed important. While these attributions are often claimed to hold the potential to improve human "understanding" of the models, surprisingly little work explicitly evaluates progress towards this aspiration. In this paper, we conduct a crowdsourcing study, where participants interact with deception detection models that have been trained to distinguish between genuine and fake hotel reviews. They are challenged both to simulate the model on fresh reviews, and to edit reviews with the goal of lowering the probability of the originally predicted class. Successful manipulations would lead to an adversarial example. During the training (but not the test) phase, input spans are highlighted to communicate salience. Through our evaluation, we observe that for a linear bag-of-words model, participants with access to the feature coefficients during training are able to cause a larger reduction in model confidence in the testing phase when compared to the no-explanation control. For the BERT-based classifier, popular local explanations do not improve their ability to reduce the model confidence over the no-explanation case. Remarkably, when the explanation for the BERT model is given by the (global) attributions of a linear model trained to imitate the BERT model, people can effectively manipulate the model.

We describe a Question Answering (QA) dataset that contains complex questions with conditional answers, i.e. the answers are only applicable when certain conditions apply.Answering the questions requires compositional logical reasoning across complex context.We call this dataset ConditionalQA.In addition to conditional answers, the dataset also features:(1) long context documents with information that is related in logically complex ways; (2) multi-hop questions that require compositional logical reasoning; (3) a combination of extractive questions, yes/no questions, questions with multiple answers, and not-answerable questions; (4) questions asked without knowing the answers.We show that ConditionalQA is challenging for many of the existing QA models, especially in selecting answer conditions.We believe that this dataset will motivate further research in understanding complex documents to answer hard questions. 1

Abstractive summarization models often generate inconsistent summaries containing factual errors or hallucinated content.Recent works focus on correcting factual errors in generated summaries via post-editing.Such correction models are trained using adversarial nonfactual summaries constructed using heuristic rules for injecting errors.However, generating non-factual summaries using heuristics often does not generalize well to actual model errors.In this work, we propose to generate hard, representative synthetic examples of nonfactual summaries through infilling language models.With this data, we train a more robust fact-correction model to post-edit the summaries to improve factual consistency.Through quantitative and qualitative experiments on two popular summarization datasets-CNN/DM and XSum-we show that our approach vastly outperforms prior methods in correcting erroneous summaries.Our model-FACTEDITimproves factuality scores by over ∼11 points on CNN/DM and over ∼31 points on XSum on average across multiple summarization models, producing more factual summaries while maintaining competitive summarization quality. 1 The first vaccine for Ebola was approved by the FDA in 2019 in the US, five years after the initial outbreak in 2014.To produce the vaccine, scientists had to sequence the DNA of Ebola, then identify possible vaccines, and finally show successful clinical trials.Scientists say a vaccine for COVID-19 is unlikely to be ready this year, although clinical trials have already started.Scientists believe a vaccine for Covid-19 might not be ready this year.The first vaccine for Ebola took 5 years to be approved by the FDA.Scientists believe a vaccine for Ebola might not be ready this year.The first vaccine for Ebola took 5 years to be produced by the CBP.

While language Models store a massive amount of world knowledge implicitly in their parameters, even very large models often fail to encode information about rare entities and events, while incurring huge computational costs.Recently, retrieval-augmented models, such as REALM, RAG, and RETRO, have incorporated world knowledge into language generation by leveraging an external nonparametric index and have demonstrated impressive performance with constrained model sizes.However, these methods are restricted to retrieving only textual knowledge, neglecting the ubiquitous amount of knowledge in other modalities like images -much of which contains information not covered by any text.To address this limitation, we propose the first Multimodal Retrieval-Augmented Transformer (MuRAG), which accesses an external non-parametric multimodal memory to augment language generation.MuRAG is pretrained with a mixture of large-scale imagetext and text-only corpora using a joint contrastive and generative loss.We perform experiments on two different datasets that require retrieving and reasoning over both images and text to answer a given query: We-bQA, and MultimodalQA.Our results show that MuRAG achieves state-of-the-art accuracy, outperforming existing models by 10-20% absolute on both datasets and under both distractor and full-wiki settings.

In this paper, we demonstrate that information retrieval can be accomplished with a single Transformer, in which all information about the corpus is encoded in the parameters of the model. To this end, we introduce the Differentiable Search Index (DSI), a new paradigm that learns a text-to-text model that maps string queries directly to relevant docids; in other words, a DSI model answers queries directly using only its parameters, dramatically simplifying the whole retrieval process. We study variations in how documents and their identifiers are represented, variations in training procedures, and the interplay between models and corpus sizes. Experiments demonstrate that given appropriate design choices, DSI significantly outperforms strong baselines such as dual encoder models. Moreover, DSI demonstrates strong generalization capabilities, outperforming a BM25 baseline in a zero-shot setup.

Abstract Many facts come with an expiration date, from the name of the President to the basketball team Lebron James plays for. However, most language models (LMs) are trained on snapshots of data collected at a specific moment in time. This can limit their utility, especially in the closed-book setting where the pretraining corpus must contain the facts the model should memorize. We introduce a diagnostic dataset aimed at probing LMs for factual knowledge that changes over time and highlight problems with LMs at either end of the spectrum—those trained on specific slices of temporal data, as well as those trained on a wide range of temporal data. To mitigate these problems, we propose a simple technique for jointly modeling text with its timestamp. This improves memorization of seen facts from the training time period, as well as calibration on predictions about unseen facts from future time periods. We also show that models trained with temporal context can be efficiently “refreshed” as new data arrives, without the need for retraining from scratch.

Abstract While many methods purport to explain predictions by highlighting salient features, what aims these explanations serve and how they ought to be evaluated often go unstated. In this work, we introduce a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. Crucially, the explanations are available to the student during training, but are not available at test time. Compared with prior proposals, our approach is less easily gamed, enabling principled, automatic, model-agnostic evaluation of attributions. Using our framework, we compare numerous attribution methods for text classification and question answering, and observe quantitative differences that are consistent (to a moderate to high degree) across different student model architectures and learning strategies.1

Although large neural language models (LMs) like BERT can be finetuned to yield state-of-the-art results on many NLP tasks, it is often unclear what these models actually learn. Here we study using such LMs to fill in entities in human-authored comparative questions, like ``Which country is older, India or _____?''---i.e., we study the ability of neural LMs to ask (not answer) reasonable questions. We show that accuracy in this fill-in-the-blank task is well-correlated with human judgements of whether a question is reasonable, and that these models can be trained to achieve nearly human-level performance in completing comparative questions in three different subdomains. However, analysis shows that what they learn fails to model any sort of broad notion of which entities are semantically comparable or similar---instead the trained models are very domain-specific, and performance is highly correlated with co-occurrences between specific entities observed in the training set. This is true both for models that are pretrained on general text corpora, as well as models trained on a large corpus of comparison questions. Our study thus reinforces recent results on the difficulty of making claims about a deep model's world knowledge or linguistic competence based on performance on specific benchmark problems. We make our evaluation datasets publicly available to foster future research on complex understanding and reasoning in such models at standards of human interaction.

This work presents a sparse-attention Transformer architecture for modeling documents that contain large tables.Tables are ubiquitous on the web, and are rich in information.However, more than 20% of relational tables on the web have 20 or more rows (Cafarella et al., 2008), and these large tables present a challenge for current Transformer models, which are typically limited to 512 tokens.Here we propose MATE, a novel Transformer architecture designed to model the structure of web tables.MATE uses sparse attention in a way that allows heads to efficiently attend to either rows or columns in a table.This architecture scales linearly with respect to speed and memory, and can handle documents containing more than 8000 tokens with current accelerators.MATE also has a more appropriate inductive bias for tabular data, and sets a new state-of-the-art for three table reasoning datasets.For HY-BRIDQA (Chen et al., 2020b), a dataset that involves large documents containing tables, we improve the best prior result by 19 points.

We present the Open Predicate Query Language (OPQL); a method for constructing a virtual KB (VKB) trained entirely from text. Large Knowledge Bases (KBs) are indispensable for a wide-range of industry applications such as question answering and recommendation. Typically, KBs encode world knowledge in a structured, readily accessible form derived from laborious human annotation efforts. Unfortunately, while they are extremely high precision, KBs are inevitably highly incomplete and automated methods for enriching them are far too inaccurate. Instead, OPQL constructs a VKB by encoding and indexing a set of relation mentions in a way that naturally enables reasoning and can be trained without any structured supervision. We demonstrate that OPQL outperforms prior VKB methods on two different KB reasoning tasks and, additionally, can be used as an external memory integrated into a language model (OPQL-LM) leading to improvements on two open-domain question answering tasks.

Bill Yuchen Lin, Haitian Sun, Bhuwan Dhingra, Manzil Zaheer, Xiang Ren, William Cohen. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Pat Verga, Haitian Sun, Livio Baldini Soares, William Cohen. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

The deductive closure of an ideal knowledge base (KB) contains exactly the logical queries that the KB can answer. However, in practice KBs are both incomplete and over-specified, failing to answer some queries that have real-world answers. \emph{Query embedding} (QE) techniques have been recently proposed where KB entities and KB queries are represented jointly in an embedding space, supporting relaxation and generalization in KB inference. However, experiments in this paper show that QE systems may disagree with deductive reasoning on answers that do not require generalization or relaxation. We address this problem with a novel QE method that is more faithful to deductive reasoning, and show that this leads to better performance on complex queries to incomplete KBs. Finally we show that inserting this new QE module into a neural question-answering system leads to substantial improvements over the state-of-the-art.

We present an implementation of a probabilistic first-order logic called TensorLog, in which classes of logical queries are compiled into differentiable functions in a neural-network infrastructure such as Tensorflow or Theano. This leads to a close integration of probabilistic logical reasoning with deep-learning infrastructure: in particular, it enables high-performance deep learning frameworks to be used for tuning the parameters of a probabilistic logic. The integration with these frameworks enables use of GPU-based parallel processors for inference and learning, making TensorLog the first highly parallellizable probabilistic logic. Experimental results show that TensorLog scales to problems involving hundreds of thousands of knowledge-base triples and tens of thousands of examples.

Automatically constructed datasets for generating text from semi-structured data (tables), such as WikiBio (Lebret et al., 2016), often contain reference texts that diverge from the information in the corresponding semistructured data.We show that metrics which rely solely on the reference texts, such as BLEU and ROUGE, show poor correlation with human judgments when those references diverge.We propose a new metric, PAR-ENT, which aligns n-grams from the reference and generated texts to the semi-structured data before computing their precision and recall.Through a large scale human evaluation study of table-to-text models for WikiBio, we show that PARENT correlates with human judgments better than existing text generation metrics.We also adapt and evaluate the information extraction based evaluation proposed in Wiseman et al. (2017), and show that PAR-ENT has comparable correlation to it, while being easier to use.We show that PARENT is also applicable when the reference texts are elicited from humans using the data from the WebNLG challenge.1 * Work done during an internship at Google.

Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, Xinghua Lu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Haitian Sun, Tania Bedrax-Weiss, William Cohen. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

The ability to inferring latent psychological traits from human behavior is key to developing personalized human-interacting machine learning systems. Approaches to infer such traits range from surveys to manually-constructed experiments and games. However, these traditional games are limited because they are typically designed based on heuristics. In this paper, we formulate the task of designing behavior diagnostic games that elicit distinguishable behavior as a mutual information maximization problem, which can be solved by optimizing a variational lower bound. Our framework is instantiated by using prospect theory to model varying player traits, and Markov Decision Processes to parameterize the games. We validate our approach empirically, showing that our designed games can successfully distinguish among players with different traits, outperforming manually-designed ones by a large margin.

Open Domain Question Answering (QA) is evolving from complex pipelined systems to end-to-end deep neural networks.Specialized neural models have been developed for extracting answers from either text alone or Knowledge Bases (KBs) alone.In this paper we look at a more practical setting, namely QA over the combination of a KB and entitylinked text, which is appropriate when an incomplete KB is available with a large text corpus.Building on recent advances in graph representation learning we propose a novel model, GRAFT-Net, for extracting answers from a question-specific subgraph containing text and KB entities and relations.We construct a suite of benchmark tasks for this problem, varying the difficulty of questions, the amount of training data, and KB completeness.We show that GRAFT-Net is competitive with the state-of-the-art when tested using either KBs or text alone, and vastly outperforms existing methods in the combined setting.

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, Christopher D. Manning. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

We propose a technique for declaratively specifying strategies for semi-supervised learning (SSL). SSL methods based on different assumptions perform differently on different tasks, which leads to difficulties applying them in practice. In this paper, we propose to use entropy to unify many types of constraints. Our method can be used to easily specify ensembles of semi-supervised learners, as well as agreement constraints and entropic regularization constraints between these learners, and can be used to model both well-known heuristics such as co-training, and novel domain-specific heuristics. Besides, our model is flexible as to the underlying learning mechanism. Compared to prior frameworks for specifying SSL techniques, our technique achieves consistent improvements on a suite of well-studied SSL benchmarks, and obtains a new state-of-the-art result on a difficult relation extraction task.

Modern deep transfer learning approaches have mainly focused on learning generic feature vectors from one task that are transferable to other tasks, such as word embeddings in language and pretrained convolutional features in vision. However, these approaches usually transfer unary features and largely ignore more structured graphical representations. This work explores the possibility of learning generic latent relational graphs that capture dependencies between pairs of data units (e.g., words or pixels) from large-scale unlabeled data and transferring the graphs to downstream tasks. Our proposed transfer learning framework improves performance on various tasks including question answering, natural language inference, sentiment analysis, and image classification. We also show that the learned graphs are generic enough to be transferred to different embeddings on which the graphs have not been trained (including GloVe embeddings, ELMo embeddings, and task-specific RNN hidden units), or embedding-free units such as image pixels.

We propose a framework to improve the performance of distantly-supervised relation extraction, by jointly learning to solve two related tasks: concept-instance extraction and relation extraction. We further extend this framework to make a novel use of document structure: in some small, well-structured corpora, sections can be identified that correspond to relation arguments, and distantly-labeled examples from such sections tend to have good precision. Using these as seeds we extract additional relation examples by applying label propagation on a graph composed of noisy examples extracted from a large unstructured testing corpus. Combined with the soft constraint that concept examples should have the same type as the second argument of the relation, we get significant improvements over several state-of-the-art approaches to distantly-supervised relation extraction, and reasonable extraction performance even with very small set of distant labels.

In this paper we study the problem of answering cloze-style questions over documents.Our model, the Gated-Attention (GA) Reader 1 , integrates a multi-hop architecture with a novel attention mechanism, which is based on multiplicative interactions between the query embedding and the intermediate states of a recurrent neural network document reader.This enables the reader to build query-specific representations of tokens in the document for accurate answer selection.The GA Reader obtains state-of-the-art results on three benchmarks for this task-the CNN & Daily Mail news stories and the Who Did What dataset.The effectiveness of multiplicative interaction is demonstrated by an ablation study, and by comparing to alternative compositional operators for implementing the gated-attention.

We study the problem of semi-supervised question answering--utilizing unlabeled text to boost the performance of question answering models.We propose a novel training framework, the Generative Domain-Adaptive Nets.In this framework, we train a generative model to generate questions based on the unlabeled text, and combine model-generated questions with human-generated questions for training question answering models.We develop novel domain adaptation algorithms, based on reinforcement learning, to alleviate the discrepancy between the modelgenerated data distribution and the humangenerated data distribution.Experiments show that our proposed framework obtains substantial improvement from unlabeled text.

We propose a general approach to modeling semi-supervised learning (SSL) algorithms. Specifically, we present a declarative language for modeling both traditional supervised classification tasks and many SSL heuristics, including both well-known heuristics such as co-training and novel domain-specific heuristics. In addition to representing individual SSL heuristics, we show that multiple heuristics can be automatically combined using Bayesian optimization methods. We experiment with two classes of tasks, link-based text classification and relation extraction. We show modest improvements on well-studied link-based classification benchmarks, and state-of-the-art results on relation-extraction tasks for two realistic domains.

Semi-supervised learning methods based on generative adversarial networks (GANs) obtained strong empirical results, but it is not clear 1) how the discriminator benefits from joint training with a generator, and 2) why good semi-supervised classification performance and a good generator cannot be obtained at the same time. Theoretically we show that given the discriminator objective, good semi-supervised learning indeed requires a bad generator, and propose the definition of a preferred generator. Empirically, we derive a novel formulation based on our analysis that substantially improves over feature matching GANs, obtaining state-of-the-art results on multiple benchmark datasets.

We study the problem of learning probabilistic first-order logical rules for knowledge base reasoning. This learning problem is difficult because it requires learning the parameters in a continuous space as well as the structure in a discrete space. We propose a framework, Neural Logic Programming, that combines the parameter and structure learning of first-order logical rules in an end-to-end differentiable model. This approach is inspired by a recently-developed differentiable logic called TensorLog [5], where inference tasks can be compiled into sequences of differentiable operations. We design a neural controller system that learns to compose these operations. Empirically, our method outperforms prior work on multiple knowledge base benchmark datasets, including Freebase and WikiMovies.

Recently, deep learning methods have been shown to improve the performance of recommender systems over traditional methods, especially when review text is available. For example, a recent model, DeepCoNN, uses neural nets to learn one latent representation for the text of all reviews written by a target user, and a second latent representation for the text of all reviews for a target item, and then combines these latent representations to obtain state-of-the-art performance on recommendation tasks. We show that (unsurprisingly) much of the predictive value of review text comes from reviews of the target user for the target item. We then introduce a way in which this information can be used in recommendation, even when the target user's review for the target item is not available. Our model, called TransNets, extends the DeepCoNN model by introducing an additional latent layer representing the target user-target item pair. We then regularize this layer, at training time, to be similar to another latent representation of the target user's review of the target item. We show that TransNets and extensions of it improve substantially over the previous state-of-the-art.

Distant labeling for information extraction (IE) suffers from noisy training data. We describe a way of reducing the noise associated with distant IE by identifying coupling constraints between potential instance labels. As one example of coupling,items in a list are likely to have the same label.A second example of coupling comes from analysis of document structure: in some corpora,sections can be identified such that items in the same section are likely to have the same label. Such sections do not exist in all corpora, but we show that augmenting a large corpus with coupling constraints from even a small, well-structured corpus can improve performance substantially, doubling F1 on one task.

We present a semi-supervised learning framework based on graph embeddings. Given a graph between instances, we train an embedding for each instance to jointly predict the class label and the neighborhood context in the graph. We develop both transductive and inductive variants of our method. In the transductive variant of our method, the class labels are determined by both the learned embeddings and input feature vectors, while in the inductive variant, the embeddings are defined as a parametric function of the feature vectors, so predictions can be made on instances not seen during training. On a large and diverse set of benchmark tasks, including text classification, distantly supervised entity extraction, and entity classification, we show improved performance over many of the existing models.

We propose a novel extension of the encoder-decoder framework, called a review network. The review network is generic and can enhance any existing encoder- decoder model: in this paper, we consider RNN decoders with both CNN and RNN encoders. The review network performs a number of review steps with attention mechanism on the encoder hidden states, and outputs a thought vector after each review step; the thought vectors are used as the input of the attention mechanism in the decoder. We show that conventional encoder-decoders are a special case of our framework. Empirically, we show that our framework improves over state-of- the-art encoder-decoder systems on the tasks of image captioning and source code captioning.

Improving the performance of recommender systems using knowledge graphs is an important task. There have been many hybrid systems proposed in the past that use a mix of content-based and collaborative filtering techniques to boost the performance. More recently, some work has focused on recommendations that use external knowledge graphs (KGs) to supplement content-based recommendation. In this paper, we investigate three methods for making KG based recommendations using a general-purpose probabilistic logic system called ProPPR. The simplest of the models, EntitySim, uses only the links of the graph. We then extend the model to TypeSim that also uses the types of the entities to boost its generalization capabilities. Next, we develop a graph based latent factor model, GraphLF, which combines the strengths of latent factorization with graphs. We compare our approaches to a recently proposed state-of-the-art graph recommendation method on two large datasets, Yelp and MovieLens-100K. The experiments illustrate that our approaches can give large performance improvements. Additionally, we demonstrate that knowledge graphs give maximum advantage when the dataset is sparse, and gradually become redundant as more training data becomes available, and hence are most useful in cold-start settings.

In an entity classification task, topic or concept hierarchies are often incomplete. Previous work by Dalvi et al. [12] has showed that in non-hierarchical semi-supervised classification tasks, the presence of such unanticipated classes can cause semantic drift for seeded classes. The Exploratory learning [12] method was proposed to solve this problem; however it is limited to the flat classification task. This paper builds such exploratory learning methods for hierarchical classification tasks.

Whereas people learn many different types of knowledge from diverse experiences over many years, most current machine learning systems acquire just a single function or data model from just a single data set. We propose a never-ending learning paradigm for machine learning, to better reflect the more ambitious and encompassing type of learning performed by humans. As a case study, we describe the Never-Ending Language Learner (NELL), which achieves some of the desired properties of a never-ending learner, and we discuss lessons learned. NELL has been learning to read the web 24 hours/day since January 2010, and so far has acquired a knowledge base with over 80 million confidence-weighted beliefs (e.g., servedWith(tea, biscuits)). NELL has also learned millions of features and parameters that enable it to read these beliefs from the web. Additionally, it has learned to reason over these beliefs to infer new beliefs, and is able to extend its ontology by synthesizing new relational predicates. NELL can be tracked online at http://rtw.ml.cmu.edu, and followed on Twitter at @CMUNELL.

Ni Lao, Einat Minkov, William Cohen. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Dana Movshovitz-Attias, William W. Cohen. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

William Yang Wang, William W. Cohen. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Because of polysemy, distant labeling for information extraction leads to noisy training data.We describe a procedure for reducing this noise by using label propagation on a graph in which the nodes are entity mentions, and mentions are coupled when they occur in coordinate list structures.We show that this labeling approach leads to good performance even when off-the-shelf classifiers are used on the distantly-labeled data.

We outline a learning framework that aims at identifying useful contextual cues for knowledge-based word sense disambiguation. The usefulness of individual context words is evaluated based on diverse lexico-statistical and syntactic information, as well as simple word distance. Experiments using two dif-ferent knowledge-based methods and bench-mark datasets show significant improvements due to context modeling, beating the conven-tional window-based approach. 1

While there has been much research on automatically constructing structured Knowledge Bases (KBs), most of it has focused on generating facts to populate a KB. However, a useful KB must go beyond facts. For example, glosses (short natural language definitions) have been found to be very useful in tasks such as Word Sense Disambiguation. However, the important problem of Automatic Gloss Finding, i.e., assigning glosses to entities in an initially gloss-free KB, is relatively unexplored. We address that gap in this paper. In particular, we propose GLOFIN, a hierarchical semi-supervised learning algorithm for this problem which makes effective use of limited amounts of supervision and available ontological constraints. To the best of our knowledge, GLOFIN is the first system for this task. Through extensive experiments on real-world datasets, we demonstrate GLOFIN's effectiveness. It is encouraging to see that GLOFIN outperforms other state-of-the-art SSL algorithms, especially in low supervision settings. We also demonstrate GLOFIN's robustness to noise through experiments on a wide variety of KBs, ranging from user contributed (e.g., Freebase) to automatically constructed (e.g., NELL). To facilitate further research in this area, we have made the datasets and code used in this paper publicly available.

Graph-based Semi-supervised learning (SSL) algorithms have been successfully used in a large number of applications. These methods classify initially unlabeled nodes by propagating label information over the structure of graph starting from seed nodes. Graph-based SSL algorithms usually scale linearly with the number of distinct labels (m), and require O(m) space on each node. Unfortunately, there exist many applications of practical significance with very large m over large graphs, demanding better space and time complexity. In this paper, we propose MAD-Sketch, a novel graph-based SSL algorithm which compactly stores label distribution on each node using Count-min Sketch, a randomized data structure. We present theoretical analysis showing that under mild conditions, MAD-Sketch can reduce space complexity at each node from O(m) to O(\log(m)), and achieve similar savings in time complexity as well. We support our analysis through experiments on multiple real world datasets. We observe that MAD-Sketch achieves similar performance as existing state-of-the-art graph-based SSL algorithms, while requiring smaller memory footprint and at the same time achieving up to 10x speedup. We find that MAD-Sketch is able to scale to datasets with one million labels, which is beyond the scope of existing graph-based SSL algorithms.

A key challenge in information and knowledge management is to automatically discover the underlying structures and patterns from large collections of extracted information. This paper presents a novel structure-learning method for a new, scalable probabilistic logic called ProPPR. Our approach builds on the recent success of meta-interpretive learning methods in Inductive Logic Programming (ILP), and we further extends it to a framework that enables robust and efficient structure learning of logic programs on graphs: using an abductive second-order probabilistic logic, we show how first-order theories can be automatically generated via parameter learning. To learn better theories, we then propose an iterated structural gradient approach that incrementally refines the hypothesized space of learned first-order structures. In experiments, we show that the proposed method further improves the results, outperforming competitive baselines such as Markov Logic Networks (MLNs) and FOIL on multiple datasets with various settings; and that the proposed approach can learn structures in a large knowledge base in a tractable fashion.

Dependency parsing is a core task in NLP, and it is widely used by many applications such as information extraction, question answering, and machine translation. In the era of social media, a big challenge is that parsers trained on traditional newswire corpora typically suffer from the domain mismatch issue, and thus perform poorly on social media data. We present a new GFL/FUDG-annotated Chinese treebank with more than 18K tokens from Sina Weibo (the Chinese equivalent of Twitter). We formulate the dependency parsing problem as many small and parallelizable arc prediction tasks: for each task, we use a programmable probabilistic firstorder logic to infer the dependency arc of a token in the sentence. In experiments, we show that the proposed model outperforms an off-the-shelf Stanford Chinese parser, as well as a strong MaltParser baseline that is trained on the same in-domain data.

In this paper, we propose the CBS topic model, a probabilistic graphical model, to derive the user communities in microblogging networks based on the sentiments they express on their generated content and behaviors they adopt. As a topic model, CBS can uncover hidden topics and derive user topic distribution. In addition, our model associates topic-specific sentiments and behaviors with each user community. Notably, CBS has a general framework that accommodates multiple types of behaviors simultaneously. Our experiments on two Twitter datasets show that the CBS model can effectively mine the representative behaviors and emotional topics for each community. We also demonstrate that CBS model perform as well as other state-of-the-art models in modeling topics, but outperforms the rest in mining user communities.

Abstract We consider a dependency-parsed text corpus as an instance of a labeled directed graph, where nodes represent words and weighted directed edges represent the syntactic relations between them. We show that graph walks, combined with existing techniques of supervised learning that model local and global information about the graph walk process, can be used to derive a task-specific word similarity measure in this graph. We also propose and evaluate a new learning method in this framework, a path-constrained graph walk variant, in which the walk process is guided by high-level knowledge about meaningful edge sequences (paths) in the graph. Empirical evaluation on the tasks of named entity coordinate term extraction and general word synonym extraction show that this framework is preferable to, or competitive with, vector-based models when learning is applied, and using small to moderate size text corpora.

In political contexts, it is known that people act as "motivated reasoners", i.e., information is evaluated first for emotional affect, and this emotional reaction influences later deliberative reasoning steps. As social media becomes a more and more prevalent way of receiving political information, it becomes important to understand more completely the interaction between information, emotion, social community, and information-sharing behavior. In this paper, we describe a high-precision classifier for politically-oriented tweets, and an accurate classifier of a Twitter user's political affiliation. Coupled with existing sentiment-analysis tools for microblogs, these methods enable us to systematically study the interaction of emotion and sharing in a large corpus of politically-oriented microblog messages, collected from just before the 2012 US presidential election. In particular, we seek to understand how information sharing is influenced by the political affiliation of the sender and receiver of a message, and the sentiment associated with the message.

Many information-management tasks (including classification, retrieval, information extraction, and information integration) can be formalized as inference in an appropriate probabilistic first-order logic. However, most probabilistic first-order logics are not efficient enough for realistically-sized instances of these tasks. One key problem is that queries are typically answered by "grounding" the query---i.e., mapping it to a propositional representation, and then performing propositional inference---and with a large database of facts, groundings can be very large, making inference and learning computationally expensive. Here we present a first-order probabilistic language which is well-suited to approximate "local" grounding: in particular, every query $Q$ can be approximately grounded with a small graph. The language is an extension of stochastic logic programs where inference is performed by a variant of personalized PageRank. Experimentally, we show that the approach performs well on an entity resolution task, a classification task, and a joint inference task; that the cost of inference is independent of database size; and that speedup in learning is possible by multi-threading.

We present a pseudo-observed variable based regularization technique for latent variable mixed-membership models that provides a mechanism to impose preferences on the characteristics of aggregate functions of latent and observed variables. The regularization framework is used to regularize topic models, which are latent variable mixed membership models for language modeling. In many domains, documents and words often exhibit only a slight degree of mixed-membership behavior that is inadequately modeled by topic models which are overly liberal in permitting mixed-membership behavior. The regularization introduced in the paper is used to control the degree of polysemy of words permitted by topic models and to prefer sparsity in topic distributions of documents in a manner that is much more flexible than permitted by modification of priors. The utility of the regularization in exploiting sentiment-indicative features is evaluated internally using document perplexity and externally by using the models to predict star counts in movie and product reviews based on the content of the reviews. Results of our experiments show that using the regularization to finely control the behavior of topic models leads to better perplexity and lower mean squared error rates in the star-prediction task.

In this paper, we propose a single low-dimensional representation for entities found in different datasets on the web. Our proposed PIC-D embeddings can represent large D-partite graphs using small number of dimensions enabling fast similarity queries. Our experiments show that this representation can be constructed in small amount of time (linear in number of dimensions). We demonstrate how it can be used for variety of similarity queries like set expansion, automatic set instance acquisition, and column classification. Our approach results in comparable precision with respect to task specific baselines and up to two orders of magnitude improvement in terms of query response time.

One essential goal in education is to improve understanding of how humans acquire knowledge and how students vary in their abilities to learn. Building an intelligent agent that models student learning would be a significant achievement in the learning sciences. SimStudent is a state-of-the-art intelligent agent that simulates a human's learning process. However, SimStudent has only been living in the world of graphical user interfaces. To construct a more human-like learning agent, we integrate SimStudent with a cognitive robot, Calliope5KP, to create a physical agent that is able to learn skill knowledge by interacting with users in the physical world. We demonstrate the integration in a tic-tac-toe game, and show that the SimStudent robot is able to learn reasonably well with 12 games.

Many event monitoring systems rely on counting known keywords in streaming text data to detect sudden spikes in frequency. But the dynamic and conversational nature of Twitter makes it hard to select known keywords for monitoring. Here we consider a method of automatically finding noun phrases (NPs) as keywords for event monitoring in Twitter. Finding NPs has two aspects, identifying the boundaries for the subsequence of words which represent the NP, and classifying the NP to a specific broad category such as politics, sports, etc. To classify an NP, we define the feature vector for the NP using not just the words but also the author's behavior and social activities. Our results show that we can classify many NPs by using a sample of training data from a knowledge-base.

We describe a novel learnable proximity measure based on personalized PageRank (also known as "random walk with reset"). Instead of introducing one weight per edge label, as in most prior work, we introduce one weight for each edge label sequence. We show that this approach is advantageous for a number of real-world tasks, including querying graph databases, recommendation tasks, and inference in large, noisy knowledge bases.

Spectral clustering methods are elegant and effective graph-based node clustering methods, but they do not allow mixed membership clustering. We describe an approach that first transforms the data from a node-centric representation to an edge-centric one, and then use this representation to define a scalable and competitive mixed membership alternative to spectral clustering methods. Experimental results show the proposed approach improves substantially in mixed membership clustering tasks over node clustering methods.

We describe a open-domain information extraction method for extracting concept-instance pairs from an HTML corpus. Most earlier approaches to this problem rely on combining clusters of distributionally similar terms and concept-instance pairs obtained with Hearst patterns. In contrast, our method relies on a novel approach for clustering terms found in HTML tables, and then assigning concept names to these clusters using Hearst patterns. The method can be efficiently applied to a large corpus, and experimental results on several datasets show that our method can accurately extract large numbers of concept-instance pairs.

Identifying latent groups of entities from observed interactions between pairs of entities is a frequently encountered problem in areas like analysis of protein interactions and social networks. We present a model that combines aspects of mixed membership stochastic block models and topic models to improve entity-entity link modeling by jointly modeling links and text about the entities that are linked. We apply the model to two datasets: a protein-protein interaction (PPI) dataset supplemented with a corpus of abstracts of scientific publications annotated with the proteins in the PPI dataset and an Enron email corpus. The model is evaluated by inspecting induced topics to understand the nature of the data and by quantitative methods such as functional category prediction of proteins and perplexity which exhibit improvements when joint modeling is used over baselines that use only link or text information.

Building an intelligent agent, which simulates human-level learning appropriate for learning math, science, or a second language, could potentially benefit both education in understanding human learning, and artificial intelligence in creating human-level intelligence. Recently, we have proposed an efficient approach to acquiring procedural knowledge using transfer learning. However, it operated as a separate module. In this paper, we describe how to integrate this module into a machine-learning agent, SimStudent, that learns procedural knowledge from examples and through problem solving. We illustrate this method in the domain of algebra, after which we consider directions for future research in this area.

The goal of semi-supervised learning (SSL) methods is to reduce the amount of labeled training data required by learning from both labeled and unlabeled instances. Macskassy and Provost (2007) proposed the weighted-vote relational neighbor classifier (wvRN) as a simple yet effective baseline for semi-supervised learning on network data. It is similar to many recent graph-based SSL methods and is shown to be essentially the same as the Gaussian-field harmonic functions classifier proposed by Zhu et al. (2003) and proves to be very effective on some benchmark network datasets. We describe another simple and intuitive semi-supervised learning method based on random graph walk that outperforms wvRN by a large margin on several benchmark datasets when very few labels are available. Additionally, we show that using authoritative instances as training seeds --- instances that arguably cost much less to label --- dramatically reduces the amount of labeled data required to achieve the same classification accuracy. For some existing state-of-the-art semi-supervised learning methods the labeled data needed is reduced by a factor of 50.

Many recommendation and retrieval tasks can be represented as proximity queries on a labeled directed graph, with typed nodes representing documents, terms, and metadata, and labeled edges representing the relationships between them. Recent work has shown that the accuracy of the widely-used random-walk-based proximity measures can be improved by supervised learning - in particular, one especially effective learning technique is based on Path-Constrained Random Walks (PCRW), in which similarity is defined by a learned combination of constrained random walkers, each constrained to follow only a particular sequence of edge labels away from the query nodes. The PCRW based method significantly outperformed unsupervised random walk based queries, and models with learned edge weights. Unfortunately, PCRW query systems are expensive to evaluate. In this study we evaluate the use of approximations to the computation of the PCRW distributions, including fingerprinting, particle filtering, and truncation strategies. In experiments on several recommendation and retrieval problems using two large scientific publications corpora we show speedups of factors of 2 to 100 with little loss in accuracy.

Markov networks (MNs) can incorporate arbitrarily complex features in modeling relational data. However, this flexibility comes at a sharp price of training an exponentially complex model. To address this challenge, we propose a novel relational learning approach, which consists of a restricted class of relational MNs (RMNs) called relation tree-based RMN (treeRMN), and an efficient Hidden Variable Detection algorithm called Contrastive Variable Induction (CVI). On one hand, the restricted treeRMN only considers simple (e.g., unary and pairwise) features in relational data and thus achieves computational efficiency; and on the other hand, the CVI algorithm efficiently detects hidden variables which can capture long range dependencies. Therefore, the resultant approach is highly efficient yet does not sacrifice its expressive power. Empirical results on four real datasets show that the proposed relational learning method can achieve similar prediction quality as the state-of-the-art approaches, but is significantly more efficient in training; and the induced hidden variables are semantically meaningful and crucial to improve the training speed and prediction qualities of treeRMNs.

Relational or semistructured data is naturally represented by a graph, where nodes denote entities and directed typed edges represent the relations between them. Such graphs are heterogeneous, describing different types of objects and links. We represent personal information as a graph that includes messages , terms , persons , dates , and other object types, and relations like sent-to and has-term . Given the graph, we apply finite random graph walks to induce a measure of entity similarity, which can be viewed as a tool for performing search in the graph. Experiments conducted using personal email collections derived from the Enron corpus and other corpora show how the different tasks of alias finding , threading , and person name disambiguation can be all addressed as search queries in this framework, where the graph-walk-based similarity metric is preferable to alternative approaches, and further improvements are achieved with learning. While researchers have suggested to tune edge weight parameters to optimize the graph walk performance per task, we apply reranking to improve the graph walk results, using features that describe high-level information such as the paths traversed in the walk. High performance, together with practical runtimes, suggest that the described framework is a useful search system in the PIM domain, as well as in other semistructured domains.

A major source of information (often the most crucial and informative part) in scholarly articles from scientific journals, proceedings and books are the figures that directly provide images and other graphical illustrations of key experimental results and other scientific contents. In biological articles, a typical figure often comprises multiple panels, accompanied by either scoped or global captioned text. Moreover, the text in the caption contains important semantic entities such as protein names, gene ontology, tissues labels, etc., relevant to the images in the figure. Due to the avalanche of biological literature in recent years, and increasing popularity of various bio-imaging techniques, automatic retrieval and summarization of biological information from literature figures has emerged as a major unsolved challenge in computational knowledge extraction and management in the life science. We present a new structured probabilistic topic model built on a realistic figure generation scheme to model the structurally annotated biological figures, and we derive an efficient inference algorithm based on collapsed Gibbs sampling for information retrieval and visualization. The resulting program constitutes one of the key IR engines in our SLIF system that has recently entered the final round (4 out 70 competing systems) of the Elsevier Grand Challenge on Knowledge Enhancement in the Life Science. Here we present various evaluations on a number of data mining tasks to illustrate our method.

In this work we try to bridge the gap often encountered by researchers who find themselves with few or no labeled examples from their desired target domain, yet still have access to large amounts of labeled data from other related, but distinct source domains, and seemingly no way to transfer knowledge from one to the other. Experimentally, we focus on the problem of extracting protein mentions from academic publications in the field of biology, where the source domain data are abstracts labeled with protein mentions, and the target domain data are wholly unlabeled captions. We mine the large number of such full text articles freely available on the Internet in order to supplement the limited amount of annotated data available. By exploiting the explicit and implicit common structure of the different subsections of these documents, including the unlabeled full text, we are able to generate robust features that are insensitive to changes in marginal and conditional distributions of classes and data across domains. We supplement these domain-insensitive features with automatically obtained high-confidence positive and negative predictions on the target domain to learn extractors that generalize well from one section of a document to another. Finally, lacking labeled target testing data, we employ comparative user preference studies to evaluate the relative performance of the proposed methods with respect to existing baselines.

Many of the recently proposed algorithms for learning feature-based ranking functions are based on the pairwise preference framework, in which instead of taking documents in isolation, document pairs are used as instances in the learning process. One disadvantage of this process is that a noisy relevance judgment on a single document can lead to a large number of mis-labeled document pairs. This can jeopardize robustness and deteriorate overall ranking performance. In this paper we study the effects of outlying pairs in rank learning with pairwise preferences and introduce a new meta-learning algorithm capable of suppressing these undesirable effects. This algorithm works as a second optimization step in which any linear baseline ranker can be used as input. Experiments on eight different ranking datasets show that this optimization step produces statistically significant performance gains over state-of-the-art methods.

Set expansion refers to expanding a partial set of "seed" objects into a more complete set. One system that does set expansion is SEAL (set expander for any language), which expands entities automatically by utilizing resources from the Web in a language independent fashion. In a previous study, SEAL showed good set expansion performance using three seed entities; however, when given a larger set of seeds (e.g., ten), SEAL's expansion method performs poorly. In this paper, we present iterative SEAL (iSEAL), which allows a user to provide many seeds. Briefly, iSEAL makes several calls to SEAL, each call using a small number of seeds. We also show that iSEAL can be used in a "bootstrapping" manner, where each call to SEAL uses a mixture of user-provided and self-generated seeds. We show that the bootstrapping version of iSEAL obtains better results than SEAL even when using fewer user-provided seeds. In addition, we compare the performance of various ranking algorithms used in iSEAL, and show that the choice of ranking method has a small effect on performance when all seeds are user-provided, but a large effect when iSEAL is bootstrapped. In particular, we show that random walk with restart is nearly as good as Bayesian sets with user-provided seeds, and performs best with bootstrapped seeds.

In this work, we address the problem of joint modeling of text and citations in the topic modeling framework. We present two different models called the Pairwise-Link-LDA and the Link-PLSA-LDA models.

Set expansion refers to expanding a given partial set of objects into a more complete set. A well-known example system that does set expansion using the web is Google Sets. In this paper, we propose a novel method for expanding sets of named entities. The approach can be applied to semi-structured documents written in any markup language and in any human language. We present experimental results on 36 benchmark sets in three languages, showing that our system is superior to Google Sets in terms of mean average precision.

Management of personal information such as email messages, calendar entries, to-do items, and workstation documents is one of the most highly visible current uses of computer technology. I will present experimental evidence that machine learning techniques can be effectively used to improve personal information management tools in two ways. First, machine learning can be used to improve performance on certain types of difficult searches, notably searches that require some awareness of context. Second, machine learning can be used to reduce the chance of certain high-cost errors. One type of high-cost error we consider is the “dropped ball”—i.e., losing track of a task that has been delegated, in part or whole, to others. The second type of high-cost error is an “email leak”—i.e., mistakenly sending a sensitive email message to the wrong recipient.

The widespread use of email has raised serious privacy concerns. A critical issue is how to prevent email information leaks, i.e., when a message is accidentally addressed to non-desired recipients. This is an increasingly common problem that can severely harm individuals and corporations — for instance, a single email leak can potentially cause expensive law suits, brand reputation damage, negotiation setbacks and severe financial losses. In this paper we present the first attempt to solve this problem. We begin by redefining it as an outlier detection task, where the unintended recipients are the outliers. Then we combine real email examples (from the Enron Corpus) with carefully simulated leak-recipients to learn textual and network patterns associated with email leaks. This method was able to detect email leaks in almost 82% of the test cases, significantly outperforming all other baselines. More importantly, in a separate set of experiments we applied the proposed method to the task of finding real cases of email leaks. The result was encouraging: a variation of the proposed technique was consistently successful in finding two real cases of email leaks. Not only does this paper introduce the important problem of email leak detection, but also presents an effective solution that can be easily implemented in any email client — with no changes in the email server side.

In collective classification, classes are predicted simultaneously for a group of related instances, rather than predicting a class for each instance separately. Collective classification has been widely used for classification on relational datasets. However, the inference procedure used in collective classification usually requires many iterations and thus is expensive. We propose stacked graphical learning, a meta-learning scheme in which a base learner is augmented by expanding one instance's features with predictions on other related instances. Stacked graphical learning is efficient, especially during inference, capable of capturing dependencies easily, and can be implemented with any kind of base learner. In experiments on eight datasets, stacked graphical learning is 40 to 80 times faster than Gibbs sampling during inference.

To learn concepts over massive data streams, it is essential to design inference and learning methods that operate in real time with limited memory. Online learning methods such as perceptron or Winnow are naturally suited to stream processing; however, in practice multiple passes over the same training data are required to achieve accuracy comparable to state-of-the-art batch learners. In the current work we address the problem of training an on-line learner with a single passover the data. We evaluate several existing methods, and also propose a new modification of Margin Balanced Winnow, which has performance comparable to linear SVM. We also explore the effect of averaging, a.k.a. voting, on online learning. Finally, we describe how the new Modified Margin Balanced Winnow algorithm can be naturally adapted to perform feature selection. This scheme performs comparably to widely-used batch feature selection methods like information gain or Chi-square, with the advantage of being able to select features on-the-fly. Taken together, these techniques allow single-pass online learning to be competitive with batch techniques, and still maintain the advantages of on-line learning.

Similarity measures for text have historically been an important tool for solving information retrieval problems. In many interesting settings, however, documents are often closely connected to other documents, as well as other non-textual objects: for instance, email messages are connected to other messages via header information. In this paper we consider extended similarity metrics for documents and other objects embedded in graphs, facilitated via a lazy graph walk. We provide a detailed instantiation of this framework for email data, where content, social networks and a timeline are integrated in a structural graph. The suggested framework is evaluated for two email-related problems: disambiguating names in email documents, and threading. We show that reranking schemes based on the graph-walk similarity measures often outperform baseline methods, and that further improvements can be obtained by use of appropriate learning methods.

BACKGROUND: One step in the model organism database curation process is to find, for each article, the identifier of every gene discussed in the article. We consider a relaxation of this problem suitable for semi-automated systems, in which each article is associated with a ranked list of possible gene identifiers, and experimentally compare methods for solving this geneId ranking problem. In addition to baseline approaches based on combining named entity recognition (NER) systems with a "soft dictionary" of gene synonyms, we evaluate a graph-based method which combines the outputs of multiple NER systems, as well as other sources of information, and a learning method for reranking the output of the graph-based method. RESULTS: We show that named entity recognition (NER) systems with similar F-measure performance can have significantly different performance when used with a soft dictionary for geneId-ranking. The graph-based approach can outperform any of its component NER systems, even without learning, and learning can further improve the performance of the graph-based ranking approach. CONCLUSION: The utility of a named entity recognition (NER) system for geneId-finding may not be accurately predicted by its entity-level F1 performance, the most common performance measure. GeneId-ranking systems are best implemented by combining several NER systems. With appropriate combination methods, usefully accurate geneId-ranking systems can be constructed based on easily-available resources, without resorting to problem-specific, engineered components.

We consider classification of email messages as to whether or not they contain certain "email acts", such as a request or a commitment. We show that exploiting the sequential correlation among email messages in the same thread can improve email-act classification. More specifically, we describe a new text-classification algorithm based on a dependency-network based collective classification method, in which the local classifiers are maximum entropy models based on words and certain relational features. We show that statistically significant improvements over a bag-of-words baseline classifier can be obtained for some, but not all, email-act classes. Performance improvements obtained by collective classification appears to be consistent across many email acts suggested by prior speech-act theory.

We consider the problem of improving named entity recognition (NER) systems by using external dictionaries---more specifically, the problem of extending state-of-the-art NER systems by incorporating information about the similarity of extracted entities to entities in an external dictionary. This is difficult because most high-performance named entity recognition systems operate by sequentially classifying words as to whether or not they participate in an entity name; however, the most useful similarity measures score entire candidate names. To correct this mismatch we formalize a semi-Markov extraction process, which is based on sequentially classifying segments of several adjacent words, rather than single words. In addition to allowing a natural way of coupling high-performance NER methods and high-performance similarity functions, this formalism also allows the direct use of other useful entity-level features, and provides a more natural formulation of the NER problem than sequential word classification. Experiments in multiple domains show that the new model can substantially improve extraction performance over previous methods for using external dictionaries in NER.

We describe semi-Markov conditional random fields (semi-CRFs), a con- ditionally trained version of semi-Markov chains. Intuitively, a semi- CRF on an input sequence x outputs a “segmentation” of x, in which labels are assigned to segments (i.e., subsequences) of x rather than to individual elements xi of x. Importantly, features for semi-CRFs can measure properties of segments, and transitions within a segment can be non-Markovian. In spite of this additional power, exact learning and inference algorithms for semi-CRFs are polynomial-time—often only a small constant factor slower than conventional CRFs. In experiments on five named entity recognition problems, semi-CRFs generally outper- form conventional CRFs.

From the standpoint of the automated extraction of scientific knowledge, an important but little-studied part of scientific publications are the figures and accompanying captions. Captions are dense in information, but also contain many extra-grammatical constructs, making them awkward to process with standard information extraction methods. We propose a scheme for "understanding" captions in biomedical publications by extracting and classifying "image pointers" (references to the accompanying image). We evaluate a number of automated methods for this task, including hand-coded methods, methods based on existing learning techniques, and methods based on novel learning techniques. The best of these methods leads to a usefully accurate tool for caption-understanding, with both recall and precision in excess of 94% on the most important single class in a combined extraction/classification task.

Part of the process of data integration is determining which sets of identifiers refer to the same real-world entities. In integrating databases found on the Web or obtained by using information extraction methods, it is often possible to solve this problem by exploiting similarities in the textual names used for objects in different databases. In this paper we describe techniques for clustering and matching identifier names that are both scalable and adaptive, in the sense that they can be trained to obtain better performance in a particular domain. An experimental evaluation on a number of sample datasets shows that the adaptive method sometimes performs much better than either of two non-adaptive baseline systems, and is nearly always competitive with the best baseline system.

Most text categorization systems use simple models of documents and document collections. In this paper we describe a technique that im- proves a simple web page classifier’s performance on pages from a new, unseen web site, by exploiting link structure within a site as well as page structure within hub pages. On real-world test cases, this technique significantly and substantially improves the accuracy of a bag-of-words classifier, reducing error rate by about half, on average. The system uses a variant of co-training to exploit unlabeled data from a new site. Pages are labeled using the base classifier; the results are used by a restricted wrapper-learner to propose potential “main-category anchor wrappers”; and finally, these wrappers are used as features by a third learner to find a categorization of the site that implies a simple hub structure, but which also largely agrees with the original bag-of-words classifier.

A program that makes an existing website look like a database is called a wrapper. Wrapper learning is the problem of learning website wrappers from examples. We present a wrapper-learning system called WL2 that can exploit several different representations of a document. Examples of such different representations include DOM-level and token-level representations, as well as two-dimensional geometric views of the rendered page (for tabular data) and representations of the visual appearance of text asm it will be rendered. Additionally, the learning system is modular, and can be easily adapted to new domains and tasks. The learning system described is part of an "industrial-strength" wrapper management system that is in active use at WhizBang Labs. Controlled experiments show that the learner has broader coverage and a faster learning rate than earlier wrapper-learning systems.

The growing need to manage and exploit the proliferation of online data sources is opening up new opportunities for bringing people closer to the resources they need. For instance, consider a recommendation service through which researchers can receive daily pointers to journal papers in their fields of interest. We survey some of the known approaches to the problem of technical paper recommendation and ask how they can be extended to deal with multiple information sources. More specifically, we focus on a variant of this problem -- recommending conference paper submissions to reviewing committee members -- which offers us a testbed to try different approaches. Using WHIRL -- an information integration system -- we are able to implement different recommendation algorithms derived from information retrieval principles. We also use a novel autonomous procedure for gathering reviewer interest information from the Web. We evaluate our approach and compare it to other methods using preference data provided by members of the AAAI-98 conference reviewing committee along with data about the actual submissions.

The web contains a large quantity of unstructured information. In many cases, it is possible to heuristically extract structured information, but the resulting databases are "soft": they contain inconsistencies and duplication, and lack unique, consistently-used object identifiers. Examples include large bibliographic databases harvested from raw scientific papers or databases constructed by merging heterogeneous "hard" databases. Here we formally model a soft database as a noisy version of some unknown hard database. We then consider the hardening problem, i.e., the problem of inferring the most likely underlying hard database given a particular soft database. A key feature of our approach is that hardening is global --- many sources of evidence for a given hard fact are taken into account. We formulate hardening as an optimization problem and give a nontrivial nearly linear time algorithm for finding a local optimum. Categories and Subject Descriptors H.4.m [Information Systems]: M...

The integration of distributed, heterogeneous databases, such as those available on the World Wide Web, poses many problems. Herer we consider the problem of integrating data from sources that lack common object identifiers. A solution to this problem is proposed for databases that contain informal, natural-language “names” for objects; most Web-based databases satisfy this requirement, since they usually present their information to the end-user through a veneer of text. We describe WHIRL, a “soft” database management system which supports “similarity joins,” based on certain robust, general-purpose similarity metrics for text. This enables fragments of text (e.g., informal names of objects) to be used as keys. WHIRL includes textual objects as a built-in type, similarity reasoning as a built-in predicate, and answers every query with a list of answer substitutions that are ranked according to an overall score. Experiments show that WHIRL is much faster than naive inference methods, even for short queries, and efficient on typical queries to real-world databases with tens of thousands of tuples. Inferences made by WHIRL are also surprisingly accurate, equaling the accuracy of hand-coded normalization routines on one benchmark problem, and outerperforming exact matching with a plausible global domain on a second.

No abstract available.

We evaluate a class of learning algorithms known as inductive logic programming (ILP) methods on the task of predicting fault density in C++ classes. Using these methods, a large space of possible hypotheses is searched in an automated fashion; further, the hypotheses are based directly on an abstract logical representation of the software, eliminating the need to manually propose numerical metrics that predict fault density. We compare two ILP systems, FOIL and FLIPPER, and conclude that FLIPPER generally outperforms FOIL on this problem. We analyze the reasons for the differing performance of these two systems, and based on the analysis, propose two extensions to FLIPPER: a user-directed bias towards easy-to-evaluate clauses, and an extension that allows FLIPPER to learn "counting clauses". Counting clauses augment logic programs with a variation of the "number restrictions" used in description logics, and significantly improve performance on this problem when prior knowledge is used. We also evaluate the use of ILP techniques for automatic generation of Boolean indicators and numeric metrics from the calling tree representation.

There are many applications in which it is desirable to order rather than classify instances. Here we consider the problem of learning how to order instances given feedback in the form of preference judgments, i.e., statements to the effect that one instance should be ranked ahead of another. We outline a two-stage approach in which one first learns by conventional means a binary preference function indicating whether it is advisable to rank one instance before another. Here we consider an on-line algorithm for learning preference functions that is based on Freund and Schapire's 'Hedge' algorithm. In the second stage, new instances are ordered so as to maximize agreement with the learned preference function. We show that the problem of finding the ordering that agrees best with a learned preference function is NP-complete. Nevertheless, we describe simple greedy algorithms that are guaranteed to find a good approximation. Finally, we show how metasearch can be formulated as an ordering problem, and present experimental results on learning a combination of 'search experts', each of which is a domain-specific query expansion strategy for a web search engine.

Two recently implemented machine-learning algorithms, RIPPER and sleeping-experts for phrases , are evaluated on a number of large text categorization problems. These algorithms both construct classifiers that allow the “context” of a word w to affect how (or even whether) the presence or absence of w will contribute to a classification. However, RIPPER and sleeping-experts differ radically in many other respects: differences include different notions as to what constitutes a context, different ways of combining contexts to construct a classifier, different methods to search for a combination of contexts, and different criteria as to what contexts should be included in such a combination. In spite of these differences, both RIPPER and sleeping-experts perform extremely well across a wide variety of categorization problems, generally outperforming previously applied learning methods. We view this result as a confirmation of the usefulness of classifiers that represent contextual information.

Most databases contain “name constants” like course numbers, personal names, and place names that correspond to entities in the real world. Previous work in integration of heterogeneous databases has assumed that local name constants can be mapped into an appropriate global domain by normalization. However, in many cases, this assumption does not hold; determining if two name constants should be considered identical can require detailed knowledge of the world, the purpose of the user's query, or both. In this paper, we reject the assumption that global domains can be easily constructed, and assume instead that the names are given in natural language text. We then propose a logic called WHIRL which reasons explicitly about the similarity of local names, as measured using the vector-space model commonly adopted in statistical information retrieval. We describe an efficient implementation of WHIRL and evaluate it experimentally on data extracted from the World Wide Web. We show that WHIRL is much faster than naive inference methods, even for short queries. We also show that inferences made by WHIRL are surprisingly accurate, equaling the accuracy of hand-coded normalization routines on one benchmark problem, and outperforming exact matching with a plausible global domain on a second.

Most databases contain “name constants” like course numbers, personal names, and place names that correspond to entities in the real world. Previous work in integration of heterogeneous databases has assumed that local name constants can be mapped into an appropriate global domain by normalization. Here we assume instead that the names are given in natural language text. We then propose a logic for database integration called WHIRL which reasons explicitly about the similarity of local names, as measured using the vector-space model commonly adopted in statistical information retrieval. An implemented data integration system based on WHIRL has been used to successfully integrate information from several dozen Web sites in two domains.

Message-passing libraries such as the Parallel Virtual Machine (PVM) and Message Passing Interface (MPI) provide a common Application Programming Interface (API) to implement parallel programs across multiple computers. Such libraries provide a means to program a collection of normally independent computers to work co-operatively on a single computation. However, for programs written in C and Fortran these collections of machines may provide a heterogenous set of computer architectures, requiring a different executable for each type of architecture. The Java language offers a potentially machine-independent method of distributing the same code to perform the computations on different computer architectures. The communication performance between processors running Java programs is a crucial issue for this type of application. This paper compares the performance between tradition PVM implemented in C code, Java code interfaced to the traditional PVM libraries (JavaPVM), and Java code that performs functions equivalent to the traditional PVM library (JPVM). The Java implementations are slower, but performance improvements are possible. © 1998 John Wiley & Sons, Ltd.

Article Free Access Share on Context-sensitive learning methods for text categorization Authors: William W. Cohen AT&T Research, 600 Mountain Avenue, Murray Hill, NJ AT&T Research, 600 Mountain Avenue, Murray Hill, NJView Profile , Yoram Singer AT&T Research, 600 Mountain Avenue, Murray Hill, NJ AT&T Research, 600 Mountain Avenue, Murray Hill, NJView Profile Authors Info & Claims SIGIR '96: Proceedings of the 19th annual international ACM SIGIR conference on Research and development in information retrievalAugust 1996 Pages 307–315https://doi.org/10.1145/243199.243278Online:18 August 1996Publication History 151citation592DownloadsMetricsTotal Citations151Total Downloads592Last 12 Months12Last 6 weeks6 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF

We present algorithms that learn certain classes of function-free recursive logic programs in polynomial time from equivalence queries. In particular, we show that a single k-ary recursive constant-depth determinate clause is learnable. Two-clause programs consisting of one learnable recursive clause and one constant-depth determinate non-recursive clause are also learnable, if an additional ``basecase'' oracle is assumed. These results immediately imply the pac-learnability of these classes. Although these classes of learnable recursive programs are very constrained, it is shown in a companion paper that they are maximally general, in that generalizing either class in any natural way leads to a computationally difficult learning problem. Thus, taken together with its companion paper, this paper establishes a boundary of efficient learnability for recursive logic programs.

In a companion paper it was shown that the class of constant-depth determinate k-ary recursive clauses is efficiently learnable. In this paper we present negative results showing that any natural generalization of this class is hard to learn in Valiant's model of pac-learnability. In particular, we show that the following program classes are cryptographically hard to learn: programs with an unbounded number of constant-depth linear recursive clauses; programs with one constant-depth determinate clause containing an unbounded number of recursive calls; and programs with one linear recursive clause of constant locality. These results immediately imply the non-learnability of any more general class of programs. We also show that learning a constant-depth determinate program with either two linear recursive clauses or one linear recursive clause and one non-recursive clause is as hard as learning boolean DNF. Together with positive results from the companion paper, these negative results establish a boundary of efficient learnability for recursive function-free clauses.

This paper considers the learnability of subsets of first-order logic. Piror work has established two boundaries of learnability: Haussler [1989] has shown that conjunctions in first-order logic cannot be learned in the Valiant model, even if the form of the conjunction is highly restricted; on the other hand, Valiant [1984] has shown that propositional conjunctions are learnable. In this paper, we study the learnability of the restricted first-order logics known as description logics. Description logics are also subsets of predicate calculus, but are expressed using a different syntax, allowing a different set of syntactic restrictions to be explored. In this paper, we first define a simple description logic, summarize some results on its expressive power, and then analyze its learnability. It is shown that the full logic cannot be tractably learned; however, syntactic restrictions that enable tractable learning exist. The learnability results hold even if the alphabets of primitive classes and roles (over which descriptions are constructed) are infinite; our positive result thus generalizes not only the result of Valiant [1984] on learning monomials to learning concepts in our (conjunctive) first order language, but also the result of Blum [1990] on learning monomials over infinite attribute spaces.

Much research in machine learning has been focused on the problem of symbol‐level learning (SLL), or learning to improve the performance of a program given examples of its behavior on typical inputs. A common approach to symbol‐level learning is to use some sort of mechanism for saving and later reusing the solution paths used to solve previous search problems. Examples of such mechanisms are macro‐operator learning, explanation‐based learning, and chunking. However, experimental evidence that these mechanisms actually improve performance is inconclusive. This paper presents a formal framework for analysis of symbol‐level learning programs, and then uses this framework to investigate a series of solution‐path caching mechanisms which provably improve performance. The analysis of these mechanisms is illuminating in many respects; in particular, in order to obtain positive results, it is necessary to use a novel representation for a set of solution paths, and also to apply certain unusual optimizations to a set of solution paths. Several of the predictions made by the model have been confirmed by recently published experiments.

The automation of the synthesis and optimization of combinational logic can result in savings in design time, significant improvements of the circuitry, and guarantee functional correctness. Synthesis quality is often measured in terms of the area of the circuit on the chip, which fails to take into account the timing constraints that might be imposed on the logic. This paper describes SOCRATES, a synthesis system capable of generating combinational logic in a given technology under user-defined timing constraints. We believe this system is the first to perform optimized, delay-constrained, multilevel synthesis into standard cell libraries. Applied to a large number of examples, the system has successfully traded off area versus delay and performs optimized, delay-constrained, multilevel synthesis into standard cell libraries.