SSP Forum: Senior Honors Students, Class of 2021

The
Symbolic Systems Forum
presents

Annual Presentation of Honors Projects
featuring
Senior Honors Students from the Class of 2021
Symbolic Systems Program

Monday, May 24, 2021
2:30-4:00 pm [NOTE longer running time]
Join from PC, Mac, Linux, iOS or Android (requires logging in to a Zoom account): https://stanford.zoom.us/j/99713091022?pwd=b0hna3VRUkFWWHl0TUEwcTRQYktwQT09

SCHEDULE:

2:30 Joel Ramirez Jr., "Investigating Distribution of Talk During Day-Long Naturalistic Recordings in English- and Spanish-Speaking Children" (Honors Advisor: Ann Fernald, Psychology; Second Reader: Virginia A. Marchman, Psychology)
Abstract: Children vary substantially in their rates of language growth (Fenson et al., 2007). Numerous studies in Western societies have shown that at least some of this variation is attributable to the overall amount of talk that children hear. Caregivers who verbally engage more with their children have children who demonstrate stronger later language skills (Hart and Risley, 1995; Weisleder & Fernald, 2013).  However, studies conducted with children living in non-Western societies (e.g., Casillas et al, 2019) have questioned the universal ¬¬nature of this relation, showing normative language development even when children hear very little talk. This suggests that features of the language environments that extend beyond the overall amount of talk may be relevant for learning. This study directly contrasts mean overall talk (i.e., Adult Word Count, AWC) with two new measures that sought to capture variation in the moment-to-moment distribution of talk across the day: (1) consistency (i.e., degree of difference between higher and lower periods of talk) and (2) density (i.e., number of consecutive periods of dense talk). Participants were children in English- (n=56) and Spanish-speaking (n=48) families with 18-24-month-old children. Caregiver talk was estimated using day-long naturalistic audio recordings in the home using LENA (Xu et al, 2009). Outcome measures were real-time language processing and vocabulary size. Results revealed striking variability among families in all three measures of talk. Children in English-speaking families who experienced higher overall AWC and more consistent talk were more efficient in real-time language processing.  However, AWC was the only unique predictor. Surprisingly, no links were observed to vocabulary, and no relations were observed in the Spanish-speaking families. Similar results were obtained after filtering the audio recordings for likely periods of child-directed speech. Future analyses should continue to explore measures of consistency and density by listening to all-day recordings.

2:40 Khuyen Le, "Is it language or is it culture? Re-examining cross-cultural similarity judgements using lexical co-occurrence" (Honors Advisor: Mike Frank, Psychology; Second Reader: Alex Carstensen, Psychology)
Abstract: Is “cow” more closely related to “grass” or “chicken”? Speakers of different languages judge similarity in this context differently, but why? One possibility is that cultures co-varying with these languages induce variation in conceptualizations of similarity. Specifically, East Asian cultures may promote reasoning about thematic similarity, by which cow and grass are more related whereas Western cultures may bias similarity judgements toward taxonomic relations, like cow-chicken. This difference in notions of similarity is the consensus interpretation for cross-cultural variation in this paradigm. We consider, and provide evidence for, an alternative possibility, by which notions of similarity are equivalent across contexts, but the statistics of the environment vary. On this account, similarity judgements are guided by co-occurrence in experience, and observing or hearing about cows and grass or cows and chickens more often could induce preferences for the relevant grouping, and account for apparent differences in notions of similarity across contexts.

2:50 Gloria Yi, "Adaptive Games for Education: Performance and Engagement in Preschoolers" (Honors Advisor: Mike Frank, Psychology; Second Reader: George Kachergis, Psychology)
Abstract: Active learning—allowing learners to direct their own studies, to an extent—is a key part of some educational theories and practices but is not well-understood in terms of its underlying mechanisms or effectiveness at different ages. Meanwhile, cognitive psychology studies have found that adults can benefit from being given self-direction in some types of tasks, and school-aged children can also benefit in simplified versions of these paradigms. Here I extend these past studies in two ways: studying younger children, whose executive function skills may still be developing, and using real-world stimuli for which children may have varying degrees of prior knowledge. In this longitudinal study, I explore whether an active learning advantage can be found in preschoolers (ages 3-4) using a tablet game that teaches letter and number recognition over a week. I compare performance and engagement in two conditions: active learners are able to control time allocation between letter and number games that adapt to their level of knowledge; control learners follow a standard curriculum with no adaptive game dynamics. Despite the challenges presented by the adaptive curriculum, I find that active learners performed just as well as their peers who followed a standard curriculum.

3:00 Dhara Kumari Yu, "Toward Few-Shot Response Generation for Conversational Agents" (Honors Advisor: Monica Lam, Computer Science; Second Reader: Chris Potts, Linguistics)
Abstract: Task-oriented conversational agents operate as a point of interaction between users and the Web, allowing people to perform actions such as making reservations, searching for information, and controlling smart home devices through a textual or voice-based interface. These systems typically rely on a natural language generation (NLG) module, separate from the natural language understanding and agent policy, to convert the agent's state into natural language that is easily understandable by users. In recent years, NLG research has shifted towards using neural models to generate more natural-sounding and human-like utterances. However, while neural NLG models can produce more natural outputs, they require a large amount of annotated data for each new domain, which is expensive to collect and is prone to annotation errors. Therefore, building models that can produce natural and accurate utterances only using a small training set - i.e., in a few-shot setting - is a critical step to deploying neural NLG in real-world settings. In my talk, I will show that using unlabeled conversational data, which is easy and inexpensive to collect, improves the quality of generated responses in few-shot settings. This approach of pretraining neural response generation models with unlabeled data outperforms previously proposed techniques, and enables response generation in unseen domains - bringing us closer to truly conversational agents.

3:10 Graham Todd, "Analyzing Parole Hearings with Natural Language Processing" (Honors Advisor: Chris Potts, Linguistics; Second Reader: Dan Jurafsky, Linguistics and Computer Science)
Abstract: The parole process is a vital component of America's criminal justice system which allows eligible candidates to be released from potentially indeterminate life sentences. In the state of California, thousands of parole hearings are conducted each year. Despite this immense volume, however, studies of the efficacy and fairness of the parole process are largely limited to close analyses of small selections of hearings. Recent advances in natural language processing have made it feasible to automatically examine vast quantities of text, but these methods are largely untested on the domain of parole hearings. On a novel dataset of roughly 35,000 parole hearings from the state of California, we perform linguistic experiments utilizing both curated lexicons and model-based predictions to characterize the language used in hearings when conditioning on a candidate's race / ethnicity and their type of legal representation. We find that these factors predict striking differences in language use by both parole commissioners and attorneys, and our results both confirm earlier studies and indicate new directions for future research. Most notably, we find across a variety of measures that privately retained attorneys are both more active in parole hearing and more likely to make use of the relevant legal precedents than their government-appointed counterparts. Our results, broadly speaking, demonstrate the promise of unsupervised methods when applied carefully to high-impact domains. However, we also note cases in which expert guidance is necessary to draw conclusions from unsupervised experiments. We argue that a combination of both approaches will be necessary to fully unlock the promise of data-driven methods.

3:20 Sofia Miranda Avila Jamesson, "Fallouts of Deportations in Texas: Analyzing the impact of large-scale workplace raids on academic performance" (Honors Advisor: Michelle Jackson, Sociology; Second Reader: Beka Guluma, Sociology)
Abstract: Workplace raids are one of the most visible and traumatic immigration enforcement tactics taken by ICE; they can result in the detention of hundreds of immigrants at one time, sending shocks through the community and leaving children suddenly without their family. Despite the humanitarian concerns surrounding these raids and their impact on the well-being of children, there is a limited amount of research on the topic. Using test scores from schools in proximity to a raid taking place in Allen, TX, I employ a series of quasi-experimental designs to determine the effects of the raid on Hispanic students' academic outcomes. I conduct two difference-in-differences analyses and a triple-difference analysis using two distinct control groups. I find statistically significant evidence that the raids negatively impacted Hispanic students’ performance on the Texas STAAR testing for both reading and math. Additionally, I find that schools closer to the raid experienced a larger decrease in test scores.

3:30 Julia Grace Jorgensen, "The Effects of Task and Stimulus on Neural Responses in the Visual Word Form Areas"  (Honors Advisor: Kalanit Grill-Spector, Psychology; Second Reader: Mareike Grotheer, Psychology)
Abstract: The Visual Word Form Areas (VWFAs) are two regions of the brain that have been found to be selective for processing words. My objective in this project is to determine the extent to which each VWFA is selective for the task of reading versus the visual presentation of text stimuli. I tested this hypothesis by using functional Magnetic Resonance Imaging to scan subjects’ brains while they performed reading task and color-recognition tasks on English words formed by either text or emojis. The fMRI results allowed us to differentiate between neural responses from the different tasks and stimuli. Preliminary fMRI results suggest that the VWFAs are not selective for text stimuli, but that one or the other (depending on the participant) could be selective for the task of reading.

3:40 Quinn Barry, "Why Did Big-City Machines Decline in the South?"  (Honors Advisor: Gary Cox, Political Science; Second Reader: Adam Bonica, Political Science)
Abstract: Will be posted later.

3:50 Mohammed Abdal Monium Osman, "Task-driven Modeling of Drosophila Heading Direction Circuits" (Honors Advisor: Thomas Clandinin, Neurobiology; Second Reader: Shaul Druckmann, Neurobiology)
Abstract: Will be posted later.

A NOTE ON THE RECORDING OF EVENTS:

If a decision has been made in advance to record an event and to make it available for later public viewing, the event announcement will usually state this. In many cases, however, decisions to record, and/or to make a recording available publicly, are not finalized before an event is announced. Availability decisions for recordings are often subject to what speakers prefer after an event has concluded, among other considerations that may include usage rights for material used in an event, as well as the need for, and practicality of, editing. When recordings are made publicly available, they will be linked within the original event announcement on the Symsys website in the days or weeks following an event. Unfortunately, we cannot follow up on individual requests for more information about whether and when a recording may become available if it is not yet posted publicly.