Noura Farra

Scoring Persuasive Essays Using Opinions and their Targets

In this work, we investigate whether the analysis of opinion expressions can help in scoring persuasive essays. For this, we develop systems that predict holistic essay scores based on features extracted from opinion expressions, topical elements, and their combinations. Experiments on test taker essays show that essay scores produced using opinion features are indeed correlated with human scores. Moreover, we find that combining opinions with their targets (what the opinions are about) produces the best result when compared to using only opinions or only topics.

Annotating Targets of Opinions in Arabic using Crowdsourcing

We present a method for annotating targets of opinions in Arabic in a two-stage process using the crowdsourcing tool Amazon Mechanical Turk. The first stage consists of identifying candidate targets “entities” in a given text. The second stage consists of identifying the opinion polarity (positive, negative, or neutral) expressed about a specific entity. We annotate a corpus of Arabic text using this method, selecting our data from online commentaries in different domains. Despite the complexity of the task, we find high agreement. We present detailed analysis.

The Columbia System in the QALB-2014 Shared Task on Arabic Error Correction

The QALB-2014 shared task focuses on correcting errors in texts written in Modern Standard Arabic. In this paper, we describe the Columbia University entry in the shared task. Our system consists of several components that rely on machinelearning techniques and linguistic knowledge. We submitted three versions of the system: these share several core elements but each version also includes additional components. We describe our underlying approach and the special aspects of the different versions of our submission. Our system ranked first out of nine participating teams.

A novel mobile wireless sensing system for realtime monitoring of posture and spine stress

Poor posture or extra stress on the spine has been shown to lead to a variety of spinal disorders including chronic back pain, and to incur numerous health costs to society. For this reason, workplace ergonomics is rapidly becoming indispensable in all major corporations. Making the individual continuously aware of poor posture may reduce out-of-posture tendencies and encourage healthy spinal habits. We have developed a novel wireless mobile sensing system which monitors spine stress in real-time by detecting poor back posture and strain on the back due to prolonged sitting or standing. The system provides a new method of measuring spine stress at both the back and the feet by integrating posture sensors with strain sensors. Posture and strain data is collected by means of a posture sensor at the neck and weight sensors at the feet. Data is transmitted wirelessly to a central processing station and real-time feedback is provided to the users mobile device when sustained bad posture is detected. Moreover, the position of the patient (sitting, standing, or walking) can be determined by analysis of the weight sensor data and is visualized in real-time, along with back posture, at the central station by means of a graphical animation. Finally, data from all sensors is stored in a database to enable post processing and data analysis, and a summary report of daily posture and physical activity is sent to the users email. The use of centralized processing allows for high performance data analysis and storage at the central station which enables tracking of the individuals progress. We demonstrate effectiveness of our system in simultaneously monitoring posture and position by testing in numerous situations.

Generalized Character-Level Spelling Error Correction

We present a generalized discriminative model for spelling error correction which targets character-level transformations. While operating at the character level, the model makes use of wordlevel and contextual information. In contrast to previous work, the proposed approach learns to correct a variety of error types without guidance of manuallyselected constraints or language-specific features. We apply the model to correct errors in Egyptian Arabic dialect text, achieving 65% reduction in word error rate over the input baseline, and improving over the earlier state-of-the-art system.

Energy-efficient mobile gesture recognition with computation offloading

Gesture recognition is a novel and compelling user input modality which allows users to interact quickly and naturally with their devices with less demand on their visual attention. Continuous gesture recognition places stringent demands on device power consumption, battery life and processing capability. In this work, we show that we can reduce the energy consumed during continuous gesture recognition on a mobile device with the delegation of the pre-processing stages, which filter out non-gesture segments, to a low power node that is separate from the main CPU. The main CPU can thus be kept in stop mode until a potential gesture is detected by the low power node, invoking the main processor to perform the computation-intensive gesture classification to detect which exact gesture has been performed by the user. We present details of the processing performance and power consumed at each step of the processing pipeline, showing the extent of power savings achieved. Experiments were conducted for detailed evaluation of the power consumption of the optimized gesture pipeline.

Cross-lingual sentiment transfer with limited resources

We describe two transfer approaches for building sentiment analysis systems without having gold labeled data in the target language. Unlike previous work that is focused on using only English as the source language and a small number of target languages, we use multiple source languages to learn a more robust sentiment transfer model for 16 languages from different language families. Our approaches explore the potential of using an annotation projection approach and a direct transfer approach using cross-lingual word representations and neural networks. Whereas most previous work relies on machine translation, we show that we can build cross-lingual sentiment analysis systems without machine translation or even high quality parallel data. We have conducted experiments assessing the availability of different resources such as in-domain parallel data, out-of-domain parallel data, and in-domain comparable data. Our experiments show that we can build a robust transfer system whose performance can in some cases approach that of a supervised system.