Mark Dredze

Confidence-weighted linear classification

We introduce confidence-weighted linear classifiers, which add parameter confidence information to linear classifiers. Online learners in this setting update both classifier parameters and the estimate of their confidence. The particular online algorithms we study here maintain a Gaussian distribution over parameter vectors and update the mean and covariance of the distribution with each instance. Empirical evaluation on a range of NLP tasks show that our algorithm improves over other state of the art online and batch methods, learns faster in the online setting, and lends itself to better classifier combination after parallel training.

Adaptive Regularization of Weight Vectors

We present AROW, an online learning algorithm for binary and multiclass problems that combines large margin training, confidence weighting, and the capacity to handle non-separable data. AROW performs adaptive regularization of the prediction function upon seeing each new instance, allowing it to perform especially well in the presence of label noise. We derive mistake bounds for the binary and multiclass settings that are similar in form to the second order perceptron bound. Our bounds do not assume separability. We also relate our algorithm to recent confidence-weighted online learning techniques. Empirical evaluations show that AROW achieves state-of-the-art performance on a wide range of binary and multiclass tasks, as well as robustness in the face of non-separable data.

Improving Lexical Embeddings with Semantic Knowledge

Word embeddings learned on unlabeled data are a popular tool in semantics, but may not capture the desired semantics. We propose a new learning objective that incorporates both a neural language model objective (Mikolov et al., 2013) and prior knowledge from semantic resources to learn improved lexical semantic embeddings. We demonstrate that our embeddings improve over those learned solely on raw text in three settings: language modeling, measuring semantic similarity, and predicting human judgements.

How Social Media Will Change Public Health

Recent work in machine learning and natural language processing has studied the health content of tweets and demonstrated the potential for extracting useful public health information from their aggregation. This article examines the types of health topics discussed on Twitter, and how tweets can both augment existing public health capabilities and enable new ones. The author also discusses key challenges that researchers must address to deliver high-quality tools to the public health community.

Quantifying Mental Health Signals in Twitter

The ubiquity of social media provides a rich opportunity to enhance the data available to mental health clinicians and researchers, enabling a better-informed and better-equipped mental health field. We present analysis of mental health phenomena in publicly available Twitter data, demonstrating how rigorous application of simple natural language processing methods can yield insight into specific disorders as well as mental health writ large, along with evidence that as-of-yet undiscovered linguistic signals relevant to mental health exist in social media. We present a novel method for gathering data for a range of mental illnesses quickly and cheaply, then focus on analysis of four in particular: post-traumatic stress disorder (PTSD), depression, bipolar disorder, and seasonal affective disorder (SAD). We intend for these proof-of-concept results to inform the necessary ethical discussion regarding the balance between the utility of such data and the privacy of mental health related information.

Combining Search, Social Media, and Traditional Data Sources to Improve Influenza Surveillance

We present a machine learning-based methodology capable of providing real-time (“nowcast”) and forecast estimates of influenza activity in the US by leveraging data from multiple data sources including: Google searches, Twitter microblogs, nearly real-time hospital visit records, and data from a participatory surveillance system. Our main contribution consists of combining multiple influenza-like illnesses (ILI) activity estimates, generated independently with each data source, into a single prediction of ILI utilizing machine learning ensemble approaches. Our methodology exploits the information in each data source and produces accurate weekly ILI predictions for up to four weeks ahead of the release of CDC’s ILI reports. We evaluate the predictive ability of our ensemble approach during the 2013–2014 (retrospective) and 2014–2015 (live) flu seasons for each of the four weekly time horizons. Our ensemble approach demonstrates several advantages: (1) our ensemble method’s predictions outperform every prediction using each data source independently, (2) our methodology can produce predictions one week ahead of GFT’s real-time estimates with comparable accuracy, and (3) our two and three week forecast estimates have comparable accuracy to real-time predictions using an autoregressive model. Moreover, our results show that considerable insight is gained from incorporating disparate data streams, in the form of social media and crowd sourced data, into influenza predictions in all time horizons.

Automatically classifying emails into activities

Email-based activity management systems promise to give users better tools for managing increasing volumes of email, by organizing email according to a users activities. Current activity management systems do not automatically classify incoming messages by the activity to which they belong, instead relying on simple heuristics (such as message threads), or asking the user to manually classify incoming messages as belonging to an activity. This paper presents several algorithms for automatically recognizing emails as part of an ongoing activity. Our baseline methods are the use of message reply-to threads to determine activity membership and a naïve Bayes classifier. Our SimSubset and SimOverlap algorithms compare the people involved in an activity against the recipients of each incoming message. Our SimContent algorithm uses IRR (a variant of latent semantic indexing) to classify emails into activities using similarity based on message contents. An empirical evaluation shows that each of these methods provide a significant improvement to the baseline methods. In addition, we show that a combined approach that votes the predictions of the individual methods performs better than each individual method alone.

Entity Linking: Finding Extracted Entities in a Knowledge Base

In the menagerie of tasks for information extraction, entity linking is a new beast that has drawn a lot of attention from NLP practitioners and researchers recently. Entity Linking, also referred to as record linkage or entity resolution, involves aligning a textual mention of a named-entity to an appropriate entry in a knowledge base, which may or may not contain the entity. This has manifold applications ranging from linking patient health records to maintaining personal credit files, prevention of identity crimes, and supporting law enforcement. We discuss the key challenges present in this task and we present a high-performing system that links entities using max-margin ranking. We also summarize recent work in this area and describe several open research problems.

Generating summary keywords for emails using topics

Email summary keywords, used to concisely represent the gist of an email, can help users manage and prioritize large numbers of messages. We develop an unsupervised learning framework for selecting summary keywords from emails using latent representations of the underlying topics in a users mailbox. This approach selects words that describe each message in the context of existing topics rather than simply selecting keywords based on a single message in isolation. We present and compare four methods for selecting summary keywords based on two well-known models for inferring latent topics: latent semantic analysis and latent Dirichlet allocation. The quality of the summary keywords is assessed by generating summaries for emails from twelve users in the Enron corpus. The summary keywords are then used in place of entire messages in two proxy tasks: automated foldering and recipient prediction. We also evaluate the extent to which summary keywords enhance the information already available in a typical email user interface by repeating the same tasks using email subject lines.

Discovering health topics in social media using topic models.

By aggregating self-reported health statuses across millions of users, we seek to characterize the variety of health information discussed in Twitter. We describe a topic modeling framework for discovering health topics in Twitter, a social media website. This is an exploratory approach with the goal of understanding what health topics are commonly discussed in social media. This paper describes in detail a statistical topic model created for this purpose, the Ailment Topic Aspect Model (ATAM), as well as our system for filtering general Twitter data based on health keywords and supervised classification. We show how ATAM and other topic models can automatically infer health topics in 144 million Twitter messages from 2011 to 2013. ATAM discovered 13 coherent clusters of Twitter messages, some of which correlate with seasonal influenza (r = 0.689) and allergies (r = 0.810) temporal surveillance data, as well as exercise (r = .534) and obesity (r = −.631) related geographic survey data in the United States. These results demonstrate that it is possible to automatically discover topics that attain statistically significant correlations with ground truth data, despite using minimal human supervision and no historical data to train the model, in contrast to prior work. Additionally, these results demonstrate that a single general-purpose model can identify many different health topics in social media.