Learning Prototypical Goal Activities for Locations Tianyu Jiang, - PowerPoint PPT Presentation

Learning Prototypical Goal Activities for Locations Tianyu Jiang, Ellen Riloff University of Utah July 17, 2018

In Infer erenc ences es We e Are e Mak aking ing John went to the restaurant. What did he do? • “sat down” • “read the menu” • “ate food” • “left a tip” 2

Pr Proto totypical Activities People go to __________ to __________ . libraries study churches pray hospitals see a doctor, have surgery ACL learn others’ work, meet people, give talk We refer to an activity that represents a common reason why people typically go to a location as a prototypical goal activity (goal-act). 3

Ho How w is is it it us useful? ul? Conversational systems, question answering, semantic disambiguation... Did you go I just went to (a) She went to the kitchen and swimming? the beach. got chicken. retrieve (b) She went to the supermarket and got chicken. purchase Notice different inferences of “ got ”! 4

Wha What is is a lo locatio tion? • Geographic Coordinate: (37.81 � S, 144.96 � E) • Political Region: China, Melbourne • Institution: school, hospital • Landscape: mountain, forest • Organization: Walmart, Sunoco 5

Ar Are these locations? doctor bedroom Wikipedia They act as locations in the phrase “ go to X ”. For example, “go to doctor” => “go to the doctor’s office”. 6

Fr Fram amework Corpus (Loc, Act) Labeled Data Activity Profile Y Location Similarity W Activity Similarity A Until Convergence Ranked list of Activities for Locations 7

Loc Location on an and Activity tivity Extr trac actio tion Dataset: 2011 Spinn3r Weblog Subset (Burton et al., 2011) Pattern 1: go to X to Y extract (Loc, Act) pairs Pattern 2: Y in/at X extract more Acts 8

Hig High h Fr Freq eq No Not Gu Guarantee “ “Go Goal” • (clinic, have appointment): not goal • (university, study law): too specific • (Disneyland, visit): too general 9

Fr Fram amework Corpus (Loc, Act) Labeled Data Activity Profile Y Location Similarity W Activity Similarity A Until Convergence Ranked list of Activities for Locations 10

Ac Activity Profile Activity profile matrix Y , where Y i,j represents the strength of the j th activity a j being a goal-act for the i th location l i . ... a 1 = buy book a 2 = eat burger a m = pray l 1 = McDonald’s 5 300 1 l 2 = Burger King 2 500 2 l 3 = bookstore 400 20 4 ... ln = church 5 10 700 An illustration of the activity profile matrix Y . 11

Ac Activity Profile Learning Intuitively, we assume that similar locations share similar activity profiles, which motivates the objective function over Y : Initialization y i 0 is a mix of co-occurrence data and labeled data. 12

Fr Fram amework Corpus (Loc, Act) Labeled Data Activity Profile Y Location Similarity W Activity Similarity A Until Convergence Ranked list of Activities for Locations 13

Ac Activity Similarity Matrix co-occurrence word matching embedding 14

Go Gold ld Stan andar ard Da Data We use Amazon Mechanical Turk to ask workers to provide ONE primary activity that is the reason why a person would go to the listed locations. People go to LOC to ___ ___ VERB NOUN We got answers for 200 locations from each of the 10 workers. 15

Gold Go ld Stan andar ard Da Data Location Gold Goal-Acts Toys R Us buy toys (9), browse gifts sink wash hands (7), wash dishes (3) airport catch flight (7), board planes, take airplane, take trips bookstore buy books (6), browse books (2), browse bestsellers, read book lake go fishing (3), go swimming (3), drive boat (2), ride boat, see scenery chiropractor get treatment (3), adjust backs (3), alleviate pain (2), get adjustment, get aligned Chinatown buy goods (2), buy duck, buy souvenirs, eat dim sum, eat rice, eat won- tons, find Chinese, speak Chinese, visit restaurants Goal-acts provided by human annotators. 16

Go Gold ld Stan andar ard Da Data Percentage of locations that have at least one goal-act assigned by multiple annotators. 96% 100% At least half of annotators 90% listed the same goal-act for 40% 78% 80% nearly 40% of the locations. 70% % of Locations 60% 53% 50% Only 1 location was assigned 39% 40% exactly the same goal-act by 1 30% 25% all annotators. 20% 15% 6% 10% 2% 0.5% 0% 2 3 4 5 6 7 8 9 10 # of Annotators Listing the Same Activity 17

Ev Evaluation Metrics Our systems produce a ranked list of hypothesized goal-acts for a location. Mean Reciprocal Rank(MRR) is used to judge the quality of the top 10 activities for each location. 18

Expe Experimental Resul sults MRR E MRR P 0.02 0.09 EMBED PMI 0.20 0.33 FREQ 0.23 0.34 AP 0.28 0.38 AP+A L 0.28 0.40 0.23 0.33 AP+A O AP+A E 0.25 0.36 AP+A L+E 0.29 0.42 Scores for MRR. 19

Expe Experimental Resul sults TOP1 TOP2 TOP3 0.05 0.08 0.12 EMBED PMI 0.25 0.36 0.41 FREQ 0.23 0.32 0.40 AP 0.29 0.41 0.47 AP+A L 0.32 0.44 0.49 0.24 0.35 0.43 AP+A O AP+A E 0.28 0.40 0.47 AP+A L+E 0.35 0.44 0.52 Scores for Top K results. 20

Expe Experimental Resul sults AP + A L+E Top 3 Location Gold Activity List PMI Top 3 buy book (6) buy book buy copy browse book (2) purchase book purchase book bookstore browse bestseller see book buy book read book get drug (4) find medicine buy pill fill prescription (3) pharmacy get prescription fill prescription get prescription (2) pick up prescription pick up prescription buy medicine get degree (4) gain education study law university gain education (5) further education study psychology watch sport gain knowledge pursue study buy grocery (8) buy item check out deal Meijer buy cream go shopping have shopping obtain grocery get item post today 21 Examples of Top 3 hypothesized prototypical goal activities.

Expe Experimental Resul sults AP + A L+E Top 3 Location Gold Activity List PMI Top 3 buy grocery (6) make money have demand buy fresh, buy goods eat out increase competition market buy shirt, find produce eat lunch lead player make call (4), NOT LOC (2) play game put number answer call, call friend phone browse website have number have conversation view website put card stop ring Examples of Top 3 hypothesized prototypical goal activities. 22

Con Conclusion ons • We introduced the problem of learning prototypical goal activities for locations. • Human annotations showed that people do associate prototypical goal-acts with locations. • Future: • More data collection. • Take advantage of more contextual information and other external knowledge. 23

Th Than ank yo you! Qu Ques estions? 24