|
bwaber
|
read my profile
sign my guestbook
Name: Ben
Country: United States
State: Massachusetts
Metro: Boston
Gender: Male
Interests:
AI, ASL, Anime, Computer Vision, Frisbee, HCI, Hiking, Japanese, Programming, Games, Reading, Technology
Currently Reading:
Expertise in Context
Elements of Data Compression
Advances in Automatic Text Summarization
Causality: Models, Reasoning, and Inference
Foundations of Language : Brain, Meaning, Grammar, Evolution
Blogs I Read:
Message: message me
Website: visit my website
Member Since:
7/21/2005
|
|
SubscriptionsSites I Read
|
|
zPrime
|
|
| Jonathan Gips, Alex (Sandy) Pentland. "Mapping Human Networks", IEEE International Conference on Pervasive Computing and Communications, Pisa, Italy, March 2006.
The authors introduce a system that can infer group membership using "bookmarks" generated by users clicking badges that they wear during a meeting at the Media Lab. They use average voice amplitude, differences of voice activity across time, proximity to other users, and many other features to perform these inferences. This is a very interesting paper with regards to social interaction, but reading the paper I listed in my previous post should give you the gist of what they go into in this paper. Still, if you're really interested in social interaction, it's a good read.
PAPER RATING: 4 | | |
| My ASL Database release in on hold for the next few days, but I PROMISE that I'll put up the link soon, since some German researchers are going to use it for ASL research and hopefully test it out too. Anyway, I also just read a paper by Pentland, which was extremely interesting. Here's the info:
Alex (Sandy) Pentland. "Are We One? On the Nature of Human Intelligence," to appear in the 5th International Conference on Development and Learning, Bloomington Illinois. May 31 - June 2, 2006.
This paper discusses and refutes the traditional cognitive science stance of viewing humans' interactions as completely individualistic and entirely intentional. Using the structure of the social network of subjects in the Media Lab and MIT Business school and data collected over a long period of time, the author finds that behavior can be very accurately predicted using models that are non-linguistic in nature and taking into account signaling and herding behaviors. Essentially, what this paper boils down to is stating that we are very social animals, and very predictable in the context of social interaction. We are also much closer to our ape relatives in interaction than we'd like to admit. This is an excellent paper with broad applicability across many fields such as computer science, social psychology, and sociology. This paper is definitely worth a thorough read. Hence the paper rating below.
PAPER RATING: 5 | | |
| The online database access interface for annotated American Sign Language data is complete. We're still working with the design, so I'll wait until later next week to put up the link, but it's an incredibly useful tool. Most datasets are monolithic, and it takes a great deal of effort to wade through the data and extract what you really want. This database has:
1. Manually annotated, ground truth linguistic and computer vision data
2. Segmented and unsegmented videos of individual utterances and signs
3. Intuitive searching tools that allow you to search according to linguistic data using a query language that I developed (the Pair-Based Query Language, or PBL - see http://lang.bu.edu/asl/queryPages/PBL_Report.pdf for the technical report)
4. Many useful data visualization tools and statistics
5. Simple data browsing features
Stay tuned... | | |
| I've read some interesting papers on affect sensing and affect generation recently, so I thought I'd conglomerate them into one post. But before that, I'm going to formalize a paper rating system, so that you can determine if it's worth your time.
PAPER RATING SYSTEM:
1. WORTHLESS - don't bother reading
2. MARGINAL - only of interest if you're entrenched in the field
3. FAIR - of broader interest, but slightly limited in scope to related fields
4. NOVEL - definitely worth reading due to originality and wide applicability, although those in highly unrelated fields may not be interested
5. SEMINAL WORK - no excuse not to read
---
ET Mueller, MG Dyer: 1985, Towards a computational theory of human daydreaming. Proceedings of the Seventh Annual Conference of the Cognitive Science Society, Hillsdale, NJ: Lawrence Erlbaum, pp. 120-129
This paper presents a very interesting computational model of daydreaming. It was somewhat lacking in experimental rigor, however, although admittedly what constitutes a "correct" daydream and what doesn't is a large endeavor in itself. Still, it's worth noting for its attempt to capture what is a very human trait, and suceeding at least on some level.
PAPER RATING: 3
---
K Binsted, G Ritchie: 1994, An implemented model of punning riddles. Proceedings of the American Association of Artificial Intelligence, AAAI Press.
The authors present a very interesting examination of humor. In particular, puns. This is a much more computationally tractable model of humor than humor in its full generality, although the authors make some quite bold statements about humor in general (that I tended to disagree with). The authors use a lexicon as well as a homonym database to generate puns that scored fairly poorly on tests (which were flawed), but some of which were quite humorous. The extension of this method to general humor is also doubtful, since it relies on handmade templates and schemas to create jokes. In addition, the system testing was weak, but the method may be validated in future tests. Here's one joke that their program generated :
What kind of emotion has bits? A love byte.
PAPER RATING: 3
---
H Liu, P Maes: 2005, The aesthetiscope: visualizing aesthetic readings of text in color space. Proceedings of AI in Music and Art, FLAIRS 2005, AAAI Press, pp. 74-79
This paper, more of an art/AI paper, uses ConceptNet and computational models of 5 modes of perception (chosen in a somewhat ad-hoc fashion, since much of the psychological research that the authors cited were over 50 years old, and in one case over 150 years old) to visualize the aesthetic nature of poems, stories, and songs. The implemented system was also placed in the MIT Media Lab's "Living Room of the Future", where the aesthetics of its literary/audio input was visualized as pusling, constantly changing colored squares that also displayed a small trace of the program's reasons for choosing certain colors. This is a very interesting paper that, while not complete text understanding, seems to understand text at a local level and incorporates other methods (such as the What Would They Think? system, see my previous post on this paper) to model cultural attitudes to support the cultural perceptual modality. This is a very interesting paper, and its methods could definitely lead to further advances in text understanding.
PAPER RATING: 4 | | |
| So, as a follow up to my previous post, the talk was fairly interesting. I would like to see some more work on how to pick out different topics so that phrasal polarity (i.e. positive/negative/neutral subjective phrases) could be more accurately determined (although the results in the paper are significant, it only represnts a few percentage point improvedment over the naive classifier, so I'm curious how valuable this method really is, given the increased computational expense of 5000 rounds of boosting required for learning on a limited corpus). Here's the paper info:
Wilson, Theresa, Wiebe, Janyce, and Hoffmann, Paul (2005). Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis. HLT-EMNLP-2005
While it's not an incredibly complicated method, it is an important look at examining polarity at a phrasal level, rather than the normal document level. In addition, it allows for shift of polarity within sentences, since this is a common occurance. It's worth at least a quick glance at the methods, if only to see where the current research is heading.
| | |
|
