flow: a first-language-oriented writing assistant system

Post on 09-Jan-2016

34 Views

Category:

Documents

1 Downloads

Preview:

Click to see full reader

DESCRIPTION

FLOW: A First-Language-Oriented Writing Assistant System. Mei-Hua Chen*, Shih-Ting Huang+, Hung-Ting Hsieh*, Ting-Hui Kao+, Jason S. Chang+ * Institute of Information Systems and Applications + Department of Computer Science - PowerPoint PPT Presentation

TRANSCRIPT

Mei-Hua Chen*, Shih-Ting Huang+, Hung-Ting Hsieh*, Ting-Hui Kao+, Jason S. Chang+

* Institute of Information Systems and Applications + Department of Computer Science National Tsing Hua University HsinChu, Taiwan, R.O.C. 30013

ACL 2012

1

Feature• First-Language-Oriented• Translations• Paraphrases• N-grams (N=5)

2

Introduction composing stage We propose a method to ” 解決問題“ .

solve the problem tackle the problem revising stage We propose a method to solve the

problem

3

盡力

try our bestdo our best

Translation-based N-gram Prediction• {e1, e2, …em, f1, f2 …fn}1.predict the possible translations (Och and Ney, 2003)

bilingual phrase alignments

2. disambiguous (correct the alignment error)

ex. ...on ways to identify tackle 洗錢 money laundering money His forum entitled money laundry

4

Paraphrase Suggestion• {e1, e2,…ek}• pivot-based method proposed by Bannard

and Callison-Burch (2005).

5

Experiment• Training data: Hong Kong Parallel Text (2,220,570 Chinese-English sentence pairs)

• 10 Chinese sentences• two students to translate the Chinese

sentences to English sentences using FLOW

6

Result

7

•Paraphrase performance well•N-gram tends to produce shorter phrases

Keke Cai, Jiajun Bu, Chun Chen, Kangmiao LiuCollege of Computer Science, Zhejiang University

Hangzhou, 310027, China

ACL 2007

8

Sentence Retrieval • Limited information

• Application:• document summarization• question answering• novelty detection

9

Term Dependence• Query:{Everest, highest , mountain}• Q ={TS1, TS2, …, TSn}• Term combinations:{Everest highest,

highest mountain, Everest mountain}• further evaluated in each retrieved

sentence• Ex. Everest is the highest mountain

10

MINIPAR• a dependency parser• Ex. Everest is the highest mountain• :{Everest highest, highest mountain,

Everest mountain}

11

Distance=(3+1+2)/3

Association Strength

12

: Size of

D( ) :

Discussion• Query:{ Everest, highest , mountain}• TS1:{ Everest, highest , mountain} TS2:{ highest , mountain}

AS(TS1, S1)= 0.5^(1/3)*0.5^2=0.1984AS(TS2, S2)= 0.5^(1/2)*0.5^1=0.35355

• Dependency distance tend to small set pairs

13

Experiments• Testing data: TREC novelty track 2003 and

2004• Average precision of each different retrieval

models

14

15

Colin Bannard , Chris Callison-BurchSchool of Informatics

University of Edinburgh2 Buccleuch Place

Edinburgh, EH8 9LW

ACL 2005

16

Parallel Corpora• Monolingual

• Bilingual (German-English)

17

Bilingual Parallel Corpora• much more commonly available resource• one language can be identified using a

phrase in another language as a pivot.

18

German is pivot, use it to find English phrase

Paraphrases• Application multidocument summarization machine translation question answering

19

Aligning phrase pairs• statistical machine translation• phrase alignment• Och and Ney(2003)

20

Giza++

Assigning probabilities

21

: original English phrase

: candidate English phrase

: foreign language phrase

Experimental Design1• 46 English phrases (occurred multiple times in the first 50,000 sentences)

• Corpus:

German-English section of the Europarl corpus (1,036,000 German-English sentence pairs)

• Manually aligned• 289 evaluation sets (each contain 2~10)

• Judgment: (meaning and grammar)

two native English speakers• Precision: 0.605

22

23

24

Experimental Design2• evaluated the accuracy of top ranked

paraphrases• conditions 1. manual alignments 2. automatic alignments 3. automatic alignments & multiple corpora in different languages (French-English, Spanish-English, Italian-

English) (4,000,000 sentence pairs)

4. re-ranking 5. limited to the same sense

25

26

IgnoreGrammar

trigram language model

27

Hao Xu† Jingdong Wang‡ Xian-Sheng Hua‡ Shipeng Li‡ †MOE-MS KeyLab of MCC, University of Science and Technology of China, Hefei, 230026, P. R. China ‡Microsoft Research Asia, Beijing 100190, P. R. China

SIGIR 2010

28

29

30

Image search schemes

Flowchart

31

Visual Instance Transformation• text-based image search (Top 50)• affinity propagation (AP) clustering

algorithm

• sort the obtained centers in a descending order of their groups sizes

32

• snoopy

33

Visual Instance

Side view

Front view

Spatial Intention Estimation• position• influence scope• Use 2D Gaussian distribution

34

Layout Sensitive Relevance Evaluation• Sum up the relevance score for each

concept• Appearance consistency -the count of common visual words

• Spatial consistency -desired spatial distribution of the concept k

-spatial distribution of visual instance v in the image

35

Quantitative Search Performance

36

Visual Results (1)

37

38

Visual Results (2)

User Study• participants : 20 college students

• To the question “have you ever had any image search intention concerning the concept layout?”

• 20% of respondents replied with “yes” and 50% of respondents replied with “no, but probably in

the future”.

39

Michael Bendersky , W. Bruce CroftDept. of Computer Science

Univ. of Massachusetts AmherstAmherst, MA

SIGIR 2012

40

Feature

41

a more accurate modeling of the dependencies between the query terms Query concepts n-grams, term proximities, noun phrases, named entities

verbose natural language queries (grammatical complexity)

42

Example

Provide information on the use of dogs worldwide for law enforcement purposes.

sequential dependence model (dog, “law enforcement”) (information, “law enforcement”)

43

Hypergraph structure

Query: “ international art crime “

44

Evaluation

top related