Related Term Suggestion using Cooking Recipe Document Structure and its Application to Interactive Query Expansion

Size: px

Start display at page:

Download "Related Term Suggestion using Cooking Recipe Document Structure and its Application to Interactive Query Expansion"

Johnathan Gregory
5 years ago
Views:

1 Related Term Suggestion using Cooking Recipe Document Structure and its Application to Interactive Query Expansion 1 Michiko Yasukawa Faculty of Science and Technology, Gunma University Abstract: This paper introduces a method for related term suggestion for cooking recipe search. Our method recognizes the document structure of recipe data and allows users to specify which word context is to be searched. Users can also specify the length of related terms and the relation between the search term and the related term. The suggestion system uses cross-searching to obtain related terms with the given length of word n-grams that are also consistent with both the user s search intent and the word meaning in the context of the given recipe documents. Two salient characteristics of the proposed method are (1) document structure recognition and (2) word n-gram length. Results of experiments conducted using standard text collection demonstrate that searches conducted with these two characteristics outperform baseline searches significantly. An exploratory search using query expansion with the proposed method is also described in this report. 1 [1] Google Web [2] Flamenco [2][3] yasukawa@gunma-u.ac.jp [4][5] n-gram

2 2:. 1:. NTCIR-11 RECIPE 2 [6] (1) ttl : title; (1 ) (2) ing : ingredient; (1 ) (3) prp : preparation; (1 ) (4) att : attribute; ( ) (att) Druck[7] [8] [9] 3 [2] (disorderly groupings) [3] (soup) (noodle) (chicken) (oriental) 3 4-gram n-gram 1 1 n-gram

1:. (prefix) (infix) (suffix) (indirect) (vietnamese) (mexican) n-gram ( ) 1 n-gram ( ) n-gram 2 2 2 ( 2 ) Document Frequency; DF ( ) 2 3: n-gram.

3 1:. (prefix) (infix) (suffix) (indirect) (vietnamese) (mexican) n-gram ( ) 1 n-gram ( ) n-gram ( 2 ) Document Frequency; DF ( ) 2 3: n-gram. ( ) ( ) ( ) 2 3 n-gram n-gram n n-gram n 1 4 n-gram (ttl) (att) n-gram ( 1 ) n-gram ( 4 ) 1/2/3/4-gram ( 4 ) 1/2/3/4-gram ( 16 ) (DF ) 2 1: 2: Algorithm1 Algorithm2-14 -

Algorithm 1 Boolean-search to calculate S(d i ). 1: S(d i ) 0.0 2: for t j d i do 3: if t j is included in Q OR then 4: S(d i) S(d i) + 1.

4 Algorithm 1 Boolean-search to calculate S(d i ). 1: S(d i ) 0.0 2: for t j d i do 3: if t j is included in Q OR then 4: S(d i) S(d i) : end if 6: end for 7: if any t j Q AND is not included in d i then 8: S(d i ) 0.0 9: end if 10: if any t j Q NOT is included in d i then 11: S(d i) : end if Algorithm 2 Cross-search for related terms. 1: for t j INDEX TARGET do 2: DF(t j ) 0.0 3: end for 4: for d i INDEX TARGET do 5: calculate S(d i) in INDEX SOURCE with query Q 6: if S(d i ) 1.0 then 7: for t j d i in INDEX TARGET do 8: DF(t j) DF(t j) : end for 10: end if 11: end for (carrot) (onion) gram 1 2 beef stew ( ) vegetable soup ( ) ( ) ( ) ( ) ( ) 2 AND : term1 AND term2 AND/OR/NOT : term1 +term2 term3 AND AND 4 AND AND/OR/NOT AND/NOT +/ OR (term1 OR term2) AND (term3 OR term4) 4:. 4 : : n-gram 2 NTCIR-11 RECIPE 4.1 NTCIR-11 [10] 3 Yummly[11] 500 Yummly Recipe Data v1 JSON 101,783 Yummly 6,258 5 ( run ) run0: OR run1: run0 OR run2: AND/NOT run1 run3: AND/NOT n-gram run2 run4: run3 run5: OR n-gram run4 (neg) (pos)

5 2:. boolean operation n-gram length search part results pos/neg operator or and/not or and/not MAP relret run0 n/a or word(1-gram) n/a all all run1 pos or word(1-gram) n/a all all *** 5775 run2 pos, neg or, and, not word(1-gram) word(1-gram) all all *** 5604 run3 pos, neg or, and, not word(1-gram) 1/2/3/4-gram all all *** 5523 run4 pos, neg or, and, not word(1-gram) 1/2/3/4-gram all, ttl, ing, prp, att ing *** 5996 run5 pos, neg or, and, not 1/2/3/4-gram 1/2/3/4-gram all, ttl, ing, prp, att ing *** 5999 *** (run0) 0.1% (p < 0.001) : free, less, no, with no, without : and, with : no bake, no cook, no crust, gluten free, dairy free, fat free, cholesterol free less n-gram 1-gram 2-gram 1-gram: eggless, flourless 2-gram: egg less, flour less ( : no) 1 ( flour ) no flour cream soup no flour GETA[12] Singhal [13] WT SMART Porter [14] : in, for, or run4 run5 (all) (ttl) (ing) (prp) (att) 5 (Data Fusion) (1) d i S c (d i ) S c (d i ) = a 0 S 0 (d i ) +a 1 S 1 (d i ) + a 2 S 2 (d i ) +a 3 S 3 (d i ) + a 4 S 4 (d i ) (1) (1) S 0 (d i ), S 1 (d i ), S 2 (d i ), S 3 (d i ), S 4 (d i ) (all) (ttl) (ing) (prp) 3: MAP t p. run0 run1 run2 run3 run4 run5 run0 n/a *** *** *** *** *** run1 n/a *** *** run2 n/a *** *** run3 n/a *** *** run4 n/a * run5 n/a *p < 0.05, ***p < (att) d i a 0 a 4 a 1 a 4 NTCIR-11 RECIPE OKSAT[15] 0.4, 0.4, 0.2, 0.2 (all) a 0 [16] a 0 = run0 run5 run AND/OR/NOT n-gram run1 run3 run 1 1 run4 run5 run4 run5 AND NOT n-gram OR n-gram 2 MAP MAP (Mean Average Precision) relret (relevant returned document) MAP ( 2) t run0 run1 run5 ( 0.1%) run0 run5 MAP p 3 run1 run3 MAP ( 2)

5: MAP(Mean Average Precision). n-gram 2 NTICR-11 (AND/OR/NOT) JSPS JP26330363 6: (relret). run4 run5 run1 run3 0.

Search User Interfaces. Cambridge University Press, 2009. [3] M. A. Hearst. Clustering versus faceted categories for information exploration. Communications of the ACM, Vol. 49, No. 4, pp.

6 5: MAP(Mean Average Precision). n-gram 2 NTICR-11 (AND/OR/NOT) JSPS JP : (relret). run4 run5 run1 run3 0.1% run4 run5 5% run MAP (relret) 5 6 run1 run3 MAP run2 run3 run1 OR run2 run3 (no egg whites) (NOT ) (with goat cheese) (with yellow cake mix) (AND ) n-gram 5 n-gram [1],,,.., [2] M. A. Hearst. Search User Interfaces. Cambridge University Press, [3] M. A. Hearst. Clustering versus faceted categories for information exploration. Communications of the ACM, Vol. 49, No. 4, pp , [4] M. Käki. Findex: search result categories help users when document ranking fails. In Proc. of SIGCHI 05, pp , [5] O. Zamir and O. Etzioni. Grouper: a dynamic clustering interface to web search results. In Proc. of WWW 99, pp , [6]., [7] G. Druck. Recipe attribute prediction using review text as supervision. In Proc. of CwC (IJCAI workshop), cwc2013_submission_6.pdf. [8],,,,.. IMQ, Vol. 113, No. 468, pp , [9].., Vol. 6, No. 7, pp , [10] M. Yasukawa, F. Diaz, G. Druck, and N. Tsukada. Overview of the NTCIR-11 cooking recipe search task. In Proc. of the 11th NTCIR, pp , [11] Yummly: Personalized recipe recommendations and search. [12] GETA. [13] A. Singhal, C. Buckley, and M. Mitra. Pivoted document length normalization. In Proc. of SIGIR, pp , [14] M.F. Porter. An algorithm for suffix stripping. Program, Vol. 14, No. 3, pp , [15] Takashi Sato, Shingo Aoki, and Yuta Morishita. OK- SAT at NTCIR-11 recipesearch. In Proc. of the 11th NTCIR, pp , [16] M. Yasukawa, J. S. Culpepper, and F. Scholer. Data fusion for Japanese term and character n-gram search. In Proc. of the 20th ADCS, pp. 10:1 10:4,

The Recipe Learner. The Recipe Learner. Doug Safreno, Yongxing Deng

The Recipe Learner. The Recipe Learner. Doug Safreno, Yongxing Deng The Recipe Learner Doug Safreno, Yongxing Deng 1 Introduction Cooking is difficult, particularly when not following a specific recipe Often times a cook wants to know how much of a common ingredient (eg