Download presentation
Presentation is loading. Please wait.
1
The MSR ESL Assistant: Detecting and correcting non-native errors in English
Michael Gamon, Chris Brockett, William B. Dolan, Jianfeng Gao, Dmitriy Belenko (Microsoft Research), Alexandre Klementiev (University of Illinois at Urbana Champaign), Claudia Leacock (Butler Hill Group)
2
Making NLP useful
3
Overview Motivation Part I: The system
Error statistics Different solutions for different errors Machine learned classifiers for preposition and determiner errors Adding a language model and web-based examples Part II: Evaluation on native and non-native data Part III: Usage and interactions
4
Motivation: The Story of the Disappearing and Reappearing Slide
750M people use English as a second or foreign language (vs. 375M as first language) 74% of use of English is between non-native speakers As many as 300M people study English in China
5
Error statistics Previous studies:
Articles and prepositions account for 20% - 50% of ESL errors Prepositions are difficult for learners with various L1 backgrounds
6
Error statistics NICT Japanese Learners of English corpus:
26.6% of errors are determiner related 10% of errors are preposition related CLEC Chinese Learners’ Corpus: 10% of errors determiner and number related 2% preposition related, 5% collocation errors (which often involve prepositional collocations)
7
Most frequent errors made by East Asian non-native speakers
Preposition presence and choice: Finally, the pollution on the world is serious. Definite and indefinite determiner presence and choice: We should think whether we have ability to do it well. Noun pluralization: So other works couldn't be done in adequate times. Gerund/infinitive confusion: So, money is also important in improve people's spirit. Auxiliary verb presence and choice: The fire will break out, it can do harmful to people. Over-regularized verb inflection: It was builded in 1995. Adjective/noun confusion: There was a wonderful women volleyball match between Chinese team and Cuba team. Word order (adjective sequences and nominal compounds): A pop British band called "Spice Girl" has sung a song.
8
Different errors – different solutions
Prepositions and articles: much contextual information needed Over-regularized verb morphology: local information is enough Noun number: local information (mass noun, quantifier etc) is enough Machine learned approaches for (1), simple heuristics for (2) and (3). Total number of error modules: 4 machine-learned modules, 19 heuristic models
9
Modeling preposition and determiner errors
What data? Domain Sentences Encarta encyclopedia 487,281 Reuters newswire 567,394 UN proceedings (Hansard) 500,000 Europarl Web scraped, using an algorithm similar to STRAND (Resnik and Smith 2003) Total 2,554,675
10
Modeling preposition and determiner errors
Preprocessing: tokenization, POStagging Heuristic algorithm (based on POS tags): find left edges of NPs (potential sites for prepositions and articles) For each potential site of a preposition or article: Target feature 1: preposition/article present or absent Target feature 2: choice of preposition/article (if present) Contextual features (POS tags to the left/right, tokens to the left/right) Maximum Entropy classifier
11
Modeling preposition and determiner errors
Training data: 2.5M sentences: Encarta, Reuters, UN, EU, web scraped Classifier Training cases Article presence/absence 11.9M Article choice 4.3M Preposition presence/absence 16.1M Preposition choice 6.5M
12
Adding a language model
LM accuracy alone is not sufficient: 58.36%
13
Adding web search Observation: Non-native speakers often use the web to validate word choice
14
Show suggestions and originals in context
15
Evaluation (1): native text (correct usage of prepositions and determiners)
Splitting the original training data into 70% training, 30% test Note: classification is split into two questions: Should there be a determiner/preposition? If yes, which one should it be? (Prepositions: limiting the set to 12 choices that are common in errors: about, as, at, by, for, from, in, like, of, on, since, to, with, "other“)
16
Articles: results on native text
Presence/absence Choice model Combined Accuracy 89.94% 89.66% 86.76% Baseline 64.04% (no article) 77.73% (definite) 58.91% Presence/absence model Precision Recall Presence 87.89% 83.54% Absence 91.01% 93.54% Choice model Precision Recall the 91.48% 95.60% a/an 81.77% 68.94%
17
Prepositions: results on native text
Presence/absence Choice model Combined Accuracy 88.57% 66.23% 76.77% Baseline 59.57% (no preposition) 27.07% (of) 42.00% Presence/absence Precision Recall Presence 86.76% 84.66% Absence 89.75% 91.23%
18
Results on individual prepositions
Choice model Precision Recall as 77.28% 62.77% on 68.17% 56.69% of 71.91% 87.54% about 60.17% 35.12% to 67.92% 64.48% by 63.37% 52.62% at 64.92% 52.85% in 61.81% 69.87% since 62.62% 20.67% with 63.45% 47.94% from 59.58% 38.36% other 56.97% 55.14% for 58.46% 47.91%
19
Evaluation(2): Human evaluation
Spellchecked Chinese Learners’ Corpus (CLEC) Test set scraped from the web User data
20
Spellchecked Chinese Learners’ Corpus (CLEC)
1 million words of English compositions collected from Chinese learners of English in China with differing levels of proficiency: senior secondary school students English-major university students non-English-major university students
21
Web scraped data collected by a vendor for MSR
Scraped from 489 personal web pages and blogs of non-native speakers/students of English, of Korean, Chinese, or Japanese L1 background 6746 sentences, 1k selected randomly for our evaluation Education level ranges from high school to graduate school, professionals are also included Gender balanced
22
Intermission: Pie charts
23
Prepositions
24
Articles
25
Broader categories CLEC Web scraped adj related verb related
noun related prep related CLEC Web scraped
26
Usage of the prototype and evaluation of user data
27
Page views per day Live Translator snafu Beijing Olympics
28
User location country visits percentage China 51,285 26.80%
United States 28,916 15.10% Taiwan 25,753 13.40% Korea - South 12,934 6.80% Hong Kong 8,826 4.60% Brazil 4,648 2.40% Canada 3,917 2.00% Germany 3,077 1.60% United Kingdom 2,928 1.50% Japan 2,581 1.30% Italy 2,579 Spain 2,557 Russian Federation 2,448 Saudi Arabia 2.021 1.10%
29
Users and Sessions
30
Repeat users (2) Highlight the 4 or more category
31
Return visits
32
Collected data
33
User interactions
34
84% of squiggles are examined by the user
This is for all frequent users that are being evaluated (eg, not include random sentences). Also not include Thesuarus
35
Are users accepting the right suggestions?
suggested accepted
36
In summary Large market for ESL proofing tools
Detecting and correcting non-native errors is a non-trivial and interesting research problem We may already be at a point where the technology starts to be useful
37
Some open questions How does the accuracy of POStagging influence the accuracy of the overall system? How can we best leverage the user behavior as a supervision signal?
38
Some ideas Using web result counts directly as an LM approximation
Using web result counts as (part of a) supervision signal for ML Combining more sources of evidence: LMs trained on different data sets etc Build one single model, including LM scores Active learning to optimize thresholds
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.