Maria Holmqvist
Sweden. PhD student
Word alignment applications: bilingual dictionaries, computer aided translation, statistical machine translation (most likely answer)
Corpus (training sample?) - hopefully millions of sentences w/ translations.
Extract phrase translations from word-aligned texts
Estimate translation probabilities (of different interpretations)
This is unsupervised- does not require human intervention for training.
Gold standard evaluation: % correct links in sample.
Alignment with parallel phrases (Maria's method)
- uses manual alignments to improve alignment quality.
-- find common alignment patterns, match them to patterns in new text.
- generalize phrases with parts of speech
-- replace nouns with noun tag, verb with verb tag- i.e. creating grammar translators!
Limit number of part of speechtags.
Questions to answer with the research:
How does phrase-based compare with statistical alignment?
How can they be combined for good results?
Ongoing work:
comparison of the two methods
comparison of wars to combine different alignments
machine translation experiments
Q: What kind of work has Bing done on getting diverse data? Your data has been academic. How do you verify your data?
A: Verify by comparing to Gold Standard of "manually aligned"
A: On the web, can get comparable text but not parallel text. So it's hard to get and work with "real" data. But work is being done to collect more diverse data.
Q: Have you looked at what the statistical method gets wrong?
A: Infrequent words. I'm doing a deeper analysis. Problem: in Swedish, it's a compound language, so there will be lots of previously unseen words in a given chunk of text.
Ewa Musial
University at Albany, SUNY
Modeling and testing of services-oriented applications
Existing testing methods: single-tier testing (presentation tier, application tier, service tier, data tier)
New approach: choreography/orchestration testing- "web flow in application layer"
- about "Business process execution applications"
- they don't question the incompleteness of current techniques.
Problem: lack of techniques to test ~~
Test model:
~
test study 1~
test study 2:
fault detection by users. 36 test cases, 31 faults.
results:
results show that approach is effective for improving reliability of industrial projects
develop complete tools for automated testing
Q: Just java?
A: yes, but not always
Q: (from guy!) Can you characterize the general faults that you were finding?
A: Yes. People tend to test easy things. If people move, the database won't keep up.
Q: How did you make the complicated dependency graph? Can't all that change?
A: It can change. Sometimes it needs to be adjusted. During testing, the reliances change...
Fariba Khan
U of Illinois- Urbana, Champaign
Assuring network service with bandwidth and integrity-based fairness
Also, DDOS attacks.
Fair-queue- head of line blocking
IP spoofing- can't use IPs as a solution
Graph- testing packet success rate under different attack circumstances. Attack, no defense and Attack, FQ, spoofing were both ~4/100 packets getting through.
Ingress filtering is neither complete not verifiable. IP of a filtered domain can be spoofed either in the same domain or in an unfiltered domain.
Design: integrity levels, integrity token, queue.
Spoofing index table (?)
MAC integrity token
(My worry: would this make communications more tracable? Where does this interact with trueanon and cyberprivacy?)
Q: How did you develop a relationship with an ISP so they will use your algorithm?
A: Haven't yet.
Now: turn in your presentation evaluations!