User Input and Interactions on Microsoft Research ESL Assistant Claudia Leacock, Butler Hill Group Michael Gamon, Microsoft Research Chris Brockett, Microsoft.

User Input and Interactions on Microsoft Research ESL Assistant

Claudia Leacock, Butler Hill GroupMichael Gamon, Microsoft ResearchChris Brockett, Microsoft Research

... andWilliam B. Dolan, Jianfeng Gao, Dmitriy

Belenko, Lucy Vanderwende (Microsoft Research)

Alexandre Klementiev (University of Illinois at Urbana Champaign)

Outline

• Who is using it and how often• How users are interacting with the system• Does it help the users to improve their writing?

Most frequent errors made by East Asian non-native speakers

Noun Related: Articles (inclusion & choice), Noun Number, Noun of Noun• I think it’s *a/the best way to resolve issues like this.• Conversion always takes a lot of *efforts/effort.• Please send the *feedback of customer/customer feedback to me by mail.

Preposition Related: inclusion & choice• It seems ok and I did not pay much attention *on/to it. • I should *to ask/ask a rhetorical question.

Verb Related: Gerund/Infinitive Confusion, Auxiliary Verb Error, Verb Formation Errors (6), Cognate/ Verb confusion, Irregular Verbs• On Saturday, I with my classmate went *eating/to eat.• Hope you will *happy/be happy in Taiwan.• I *teached/taught him all the things I know.

Adjective Related: Adjective Confusion (4), Adjective Order• She is very *interesting/interested in the problem.• So *Korea/Korean Government is intensely fostering trade.

4

User Interface Deployed 6/2008

5

Page Views per Week

7/20/0

8

8/ 3/0

8

8/17/0

8

8/31/0

8

9/14/0

8

9/28/0

8

10/12/0

8

10/26/0

8

11/ 9/0

8

11/23/0

8

12/ 7/0

8

12/21/0

8

1/ 4/0

9

1/18/0

9

2/ 1/0

9

2/15/0

9

3/ 1/0

9

3/15/0

9

3/29/0

9

4/12/0

9

4/26/0

9

5/10/0

90

5,000

10,000

15,000

20,000

25,000

Page Views per Week

User Locations

1. China 59,276 17.9% 10. Japan 5,941 1.8%

2. United States 55,104 16.6% 11. Spain 5,924 1.8%

3. Taiwan 47,159 14.2% 12. United Kingdom 5,828 1.8%

4. Korea - South 18,730 5.6% 13. Russian Federation 5,454 1.6%

5. Hong Kong 14,259 4.3% 14. France 3,971 1.2%

6. Brazil 8,444 2.5% 15. Saudi Arabia 3,893 1.2%

7. Germany 8,219 2.5% 16. Mexico 3,878 1.2%

8. Canada 7,634 2.3% 17. Netherlands 3,330 1.0%

9. Italy 6,880 2.1% 18. Thailand 3,207 1.0%

Repeat Users

once only 2 times or more 3 times or more 4 times or more 5 times or more0

10

20

30

40

50

60

70

80

90

100

Return frequencype

rcen

tage

of t

otal

visi

ts

8

Frequent Users (4/21/09).

9

adj2%

verb10%

prep27%

noun61%

Frequent Users 854

Sessions 8,339

Session-Unique Sentences 66,765

Grammatical Error Flags 22,542

Collected Data (4/21/09)

Email49%

Non-technical25%

Technical17%

Other4%

Unrelated5%

Writing Domains: By Number of SentencesWriting Domains: By Number of Sentences

10

User Interaction 1:Responses to “Tell us what you think!”

Some users wrote: Other users wrote:

“This is awesome! It works really well.” “It didn’t work at all.”

“I found the tool very useful.” “I hate it.”

“Great tool in general – thank you!!!!!!!” “Terrible job.”

“I love the feature where it looks for a phrase in web pages.”

“The microsoft search results below confuses me.”

Bug reports: “When I first opened it, it wouldn’t let me type in any characters at all.”

“What wearies me is the message ‘Server is temporarily unavailable’.”

Suggestions: “There should be some indication that the check is done.”

“I would like a filter for business and personal use.”

Users Examine 83% of SuggestionsAccept

42%

Trigger web search but don't accept

28%

Look at suggestion but not trigger web

search31%

12

Conclusion: A significant number of users are inspecting the suggested rewrites and making a deliberate choice to accept it or not accept it.

Inspect >18.3K Flags to Accept 7.6K

Do users make the right choices?

Evaluated ~900 complete user sessions: 6K flags1. Calculate system performance for ALL

suggestions.2. Calculate performance for ONLY

suggestions that were accepted.3. Compare ratios of good and bad flags.

13

Evaluation CategoriesEvaluation SubEval Description

Good Correct Flag The correction fixes a problem in the user input.

Neutral

Both Good The suggestion is a legitimate alternative of a well-formed original input. Ex: I like working/to work.

MisdiagnosisThe original input contained an error but the suggested rewrite neither improves nor further degrades the user input. Ex: If you have fail machine on hand.

Non-ascii A non-ascii or text processing mark-up character is in the immediate context. (Only applies to user data)

Bad False Flag The suggestion resulted in an error or would otherwise lead to a degradation over the original user input.

14

Are users accepting good suggestions?All significant in the Wilcoxin’s signed-ranks test.

Noun-related Prep-related Verb-related Adj-related

15

good56%neut

28%

bad

16%

All Suggestions

good63%

neut

26%

bad11%

Accepted

good37%

neut

39%

bad24%

All Suggestions

good45%

neut

42%

bad13%

Accepted

good62%

neut32%

bad6%

All Suggestions

good72%

neut25%

bad3%

Accepted

good45%

neut 32%

bad23%

All Suggestions

good63%

neut28%

bad9%

Accepted

By Domain: All significant in the Wilcoxin’s signed-ranks test.

Email Non-technical Technical

16

good53%neutral

32%

bad15%

Suggestions

good62%

neutral28%

bad9%

Accepted

good56%

neutral32%

bad12%

Suggestions

good56%

neutral34%

bad10%

Accepted

good38%

neutral28%

bad34%

Suggestions

good51%

neutral29%

bad19%

Accepted

What do users do with neutral flags?

Both OK15%

Misdiagnosis78%

non-ascii7%

Neutral Categories

• I don't know that you knew or not , this early morning i got a from head office ... – suggestion: delete “from”I don't know that you knew or not , this early morning I heard from the head

office ...• Please play with the software and Friday I will be by to work with any questions

you may regarding it.– suggestion: regardingregardPlease play with the software and Friday I will be by to work with any questions

you may have regarding it.

18

Neutral Flags not accepted but sentence edited to produce no flag

From 1,349 sentences with neutral flags found 215 subsequently submitted “similar” strings with no error flag.Users not accept suggestion but did something ELSE to make the flag go away.

Users improve 40% of the time

19

Typed in suggestion44%

Revise and improve40%

Revise and not improve16%

Not Accept Suggestion but Revise Sentence

Identifying the location of an error can help the user.

Conclusions

• Traffic: There is an interest in ESL proofing tools• Even current state-of-the-art error correction can be

useful for ELLs:Users do not accept proposed corrections blindly – they are

selective in their behaviorUsers make informed choices – they can distinguish correct

suggestions from incorrect ones Sometimes just identifying the location of an error enables

the users to repair the problem themselves

20

New Interface

www.eslassistant.com