Blog

This is a research Blog for my course INFSCI 3005. I will update my resaerch insights and simenars summary.

Improving Sentence Retrieval from Case Law for Statutory Interpretation

ISP Forum 8/31/2019

Jaromir Savelka gave a research talk about his recent paper 'Improving Sentence Retrieval from Case Law for Statutory Interpretation'. The model he develops aims to retrieval important sentences form court opinions that elaborate on the meaning of a vague statutory term. By condensing the information users need to read, this technique may help review long lists of cases in search of useful explanations. The data set used in this research is a set of real case law sentences from three different sources, and was labeled by human reviewers in terms of their usefulness for interpretation. Techniques used to solve the ranking problem includes measure the similarity between the sentence and the query, utilize the context of a sentence, expand queries, or assess the novelty of a sentence with respect to a statutory provision from which the interpreted term comes.

I did not gain much insights from this presentation due to the lack of background knowledge in natural language processing techniques and task scenarios. I hope the talk about human robot interaction in this Friday may be more help.


Teamwork with robots

CMU HCII Seminar 9/6/2019

Dr. Malte Jung gave a research talk titled 'Teamwork with robots' at CMU last week. His research basiclly focuses on the influence and effect of social robots on human team dynamics. Here attached the note and my thought about his talk.

Teamwork with robots

  • interpersonal dynamics — human-human team
    • the signature of bad team work through the interaction of personals
    • patterns distinguish success marriage/team work
    • Dissertation research:
      • bring people in a lab scenario for a team task,
      • encode the verbal and nonverbal interaction (hostile),
      • find same pattern, same performance trend
  • how robot could influence this process
    • pure industrial robot —> team with human
    • one robot- one human interaction —> multi robots - multi human interaction
    • robot design influence how human perceive it -> robot shapes the social dynamic of human human interaction
    • Study 1 how does the distribution of block s influence the social and task dynamic?
      • Findings: different in tower height and relationship with human co-worker
      • expectation -> not fulfill -> discomfort sign for both ps -> social repair behaviors
    • Other research 2018:
      • mistake and comfort behaviors
      • robot mediates the relation between 2 kids in order to solve the conflicts
    • Study 2
      • Balanced the participation of team members increase the performance
      • control group, encourage group, random moving task
      • implicit interaction still has its effect -> and does not distract

Takeaways:

Theory: Robot not only change the way we interact with it, but also shape the way we interact with each other

Method: Simple paradigms to let robot collaborate with human

Design: Do not need complex humanoid robot

Q&A

Long term effect of robot

What if participants are were of the function of robots

If the robot receive attention due to novelty or its fundamental features?


Language and Interaction in Minecraft

CMU RI VASC Seminar 9/30/2019

Arthur Szlam is a research scientist working at Facebook AI. Prior to joining Facebook, he was an assistant professor at CCNY in the math department. He earnt his PhD under R. Coifman at Yale. In the seminar he talked about the Minecraft assistant project he leads at Facebook AI. The purpose of this research program is to implement agents that can complete tasks specified by dialogue, and to learn from dialogue interactions eventually. In order to reach this goal, instead of superhuman performance on a single difficult task, they are interested in competency across a large number of simpler tasks, specified by humans. In such cases, many tasks will have been seen only a few times, or even never, requiring sample efficiency and flexibility.The platform they are working on is a sand-box game environment named Minecraft. The constraints of the Minecraft world (e.g. coarse 3-d voxel grid, simple physics) and the regularities in the head of the distribution of in-game tasks allow numerous hand-holds for NLU research. The tools and platform he introduced give me insights on using Minecraft in our own research about natural language interaction. Such an environment allowes players to interact with the agents and to record those interactions, which makes the implementation of intelligent agents and human subject experimentations easier.

Expanding the reach of AIED systems: Adapting to social learning processes

ISP AI Forum 9/27/2019

This talk is given by Dr. Erin Walker, a faculty member at Pitt CS department. The topic of this talk is mainly about her work on Artificial Intelligence in education (AIED) systems. Two projects are introduced, one is Postdigital Textbook which allows student to draw concept map when reading the textbook to organize their knowledge, the other is EMBRACE which refers to a reading augmented application helping Spanish speaking children to learn English. The digital textbook is a great attempt on computer assisted learning because it tracks the process of generating knowledge representation. Having the reaction time and behavior patterns of students, individualized intervention can be applied to facilitate their learning. However, I barely see the innovation of a visualized tool for reading comprehension (EMBRACE). The learning process of bilinguals such as native Spanish speakers learning English might be restricted by multiple factors like culture background and socioeconomic status. Involving family in this process is desirable but still need more consideration in the level of application.


What do you think of vaping? Machine learning methods for Twitter stance detection

ISP AI Forum 11/08/2019

Sanya Tabeja, an ISP student introduces her work on natural language processing of e-cigarette (vaping) related twitters. They collect data using the twitter API and filter out those with relevant keywords. The following pipeline is typical for a text classification task: exclude irrelevant tweets and commercial advertisements then focus on the data that is most informative in terms of public opinion. The random forest tree method outperforms other methods including Logistic regression, Naive Bayes and SVM in the classification task.

My thought about this talk is that the presenter introduces many technical details about model selection and parameter training during the above process. However, the boarder impact and generalizability of this work is missing during the presentation. Considering the special topics and platform, collected dataset might be biased in terms of the target population. Therefore, the conclusions draw from this research need further evaluation before apply to other domains such as public policy.