Joel Tetreault Daniel Blanchard Aoife Cahill Nuance Communications Educational Testing Service Educational Testing Service
Task of automatically identifying a speaker’s first language based solely on the speaker’s writing in another language Applications: ◦ Authorship profiling (Estival et al., 2007) ◦ Education: more targeted feedback to language learners (Leacock et al., 2010)
No risk no fun I agree the statement "Successful people try new things and take risk".In my mind it is so, to. When you thing you like do new stuff you need a liddelbit the kick. That is the big point what I need. For exsample I like to go to a big city like New York. I was never in this town I dont no from the city. But I like go to the city. Thats fun I stay every time for proplems. I need eat a hood offer my head. The ather side I can go dow. I dont gat waht I need…Next exsample the wall street you put money in funds, well you this make a good job. Dont for get the risk look like lose money.
For example, if you take a look at an ordinary school, you have different teachers for every subject. Your calculus teacher is different than your literature teacher. Each teacher must specialize in a specific subject in order to convey suffiecient and proper information to the students. However, that doesn't mean that the teacher is narrow-minded and has a limited perspective in life because to specialize in one subject doesn't hinder you or stop you from exploring other subjects.
Lots of work in NLI but…it has been hard to compare different approaches: ICLEv2 (Granger et al, 2009): de facto train/test data is small and has NLI unfriendly idiosyncrasies 2.
No consensus on evaluation: Which L1’s / how many L1’s?
Goal to unify community and help field progress Provide a larger, more NLI-friendly corpus that improves upon ICLEv2 Common evaluation framework ◦ Everyone evaluates using same train/dev/test splits and same L1s Corpus and scripts to be made public to further promote the field
Prior Work Data Shared Task Overview Results NLI Shared Task in the Future
Treat NLI as a classification task Koppel et al. (2005): POS n-grams, content and function words, spelling and grammatical errors Syntactic features (Wong and Dras, 2011) Tree Substitution Grammars (Swanson and Charniak, 2012) Adaptor Grammars (Wong et al., 2012) Data Size Effects (Brooke and Hirst, 2012) Word n-grams (Bykh and Meurers, 2012): LMs and Ensemble Classifiers (Tetreault et al., 2012)
12,100 essays from the ETS Test of English as a Foreign Language (TOEFL) 11 L1s: ◦ Arabic, Chinese French, German, Hindi, Italian, Japanese, Korean, Spanish, Telugu, Turkish ◦ 900 train / 100 dev / 100 test Sampled for equal representation of L1s across topics as much as possible Includes 3-tier proficiency level Public release via LDC this summer?
Closed-Training: 11-way classification task using only TOEFL11-TRAIN and DEV Open-Training-1: use of any amount or type of training data excluding TOEFL11 Open-Training-2: use of any amount or type of training data combined with TOEFL11 * All sub-tasks use TOEFL11-TEST for the final evaluation set
Each team allowed to submit up to 5 different systems per task Teams submitted a CSV file for each system to NLI Organizers Evaluation script automatically compares each prediction file to gold standard and creates performance report and contingency tables
Bobicev Chonger CMU-Haifa Cologne-Nijmegen CoRAL Lab @ UAB CUNI (Charles University) Cywu Dartmouth Eurac HAUTCS ItaliaNLP Jarvis Kyle et al.
LIMSI LTRC IIIT Hyderabad Michigan MITRE “Carnie” MQ NAIST NRC Oslo NLI Toronto UKP Unibuc UNT UTD VTEX Tuebingen Ualberta
Closed Open-1 Open-2
# Teams Competing
29 3 4
116 13 15
See Table 3 of Report for full results No statistically significant differences between top 5 teams
Jarvis Oslo NLI Unibuc MITRE “Carnie” Tuebingen
JAR OSL BUC CAR TUE
Challenge : finding new data to cover each L1
ICLE FCE ICNALE Lang8 All L1s except ARA, HIN, TEL All L1s except ARA , HIN, TEL CHI, JPN, KOR essays only All L1s, but mostly Asian L1s Data sources for HIN & TEL: ◦ ICNALE Pakistani essays HIN (TUE team) ◦ Bilingual blogs (TOR & TUE team)
Machine Learning ◦ SVM overwhelmingly the most popular approach ◦ 4 teams also tried Ensemble classifiers ◦ String kernels (BUC) using character level n grams
Features ◦ N-grams: word, POS, character, function ◦ Syntactic Features: Dependencies, TSG, CF Productions, Adaptor Grammars ◦ Spelling Features 4 of top 5 teams used n-grams at least 4 grams, some went up to 9-grams 2 of top 10 teams used syntactic features
◦ ◦ ◦ Ideas to expand scope of task ◦ Use a new set of TOEFL essays for test Expand genres: blogs? Tweets? Number of L1s Do different L2 ItaliaNLP – preparing Italian NLI corpus with CNR Pisa Also a corpus of Finnish with L1 (Turku Uni) ◦ Add slavic languages Logistics ◦ Hold another shared task in 2014? Or 2015?
◦ Merge with PAN Shared Task?
Tell us your thoughts!
Derrick Higgins (ETS) ETS TOEFL Patrick Houghton (ETS) BEA8 Organizers All the NLI Participants!