Proceedings of the Workshop on Uphill Battles in Language Processing: Scaling Early Achievements to Robust Methods最新文献

{"title":"C2D2E2: Using Call Centers to Motivate the Use of Dialog and Diarization in Entity Extraction","authors":"Kenneth Ward Church, Weizhong Zhu, Jason W. Pelecanos","doi":"10.18653/v1/W16-6008","DOIUrl":"https://doi.org/10.18653/v1/W16-6008","url":null,"abstract":"This paper introduces a deceptively simple entity extraction task intended to encourage more interdisciplinary collaboration between fields that don’t normally work together: diarization, dialog and entity extraction. Given a corpus of 1.4M call center calls, extract mentions of trouble ticket numbers. The task is challenging because first mentions need to be distinguished from confirmations to avoid undesirable repetitions. It is common for agents to say part of the ticket number, and customers confirm with a repetition. There are opportunities for dialog (given/new) and diarization (who said what) to help remove repetitions. New information is spoken slowly by one side of a conversation; confirmations are spoken more quickly by the other side of the conversation. 1 Extracting Ticket Numbers Much has been written on extracting entities from text (Etzioni et al., 2005), and even speech (Kubala et al., 1998), but less has been written in the context of dialog (Clark and Haviland, 1977) and diarization (Tranter and Reynolds, 2006; Anguera et al., 2012; Shum, 2011). This paper describes a ticket extraction task illustrated in Table 1. The challenge is to extract a 7 byte ticket number, “902MDYK,” from the dialog. Confirmations ought to improve communication, but steps need to be taken to avoid undesirable repetition in extracted entities. Dialog theory suggests it should be possible to distinguish first mentions (bold) from confirmations (italics) based on prosodic cues such as pitch, energy and duration. t0 t1 S1 S2 278.16 281.07 I do have the new hardware case number for you when you’re ready 282.60 282.85 okay 284.19 284.80 nine 285.03 285.86 zero 286.22 286.74 two 290.82 291.30 nine 292.87 293.95 zero two 297.87 298.24 okay 299.30 300.49 M. as in Mike 301.97 303.56 D. as in delta 304.89 306.31 Y. as in Yankee 307.50 308.81 K. as in kilo 310.14 310.57 okay 310.77 311.70 nine zero two 311.73 312.49 M. D. 312.53 313.18 Y. T. 313.75 314.21 correct 314.21 317.28 and thank you for calling IBM is there anything else I can assist you with Table 1: A ticket dialog: 7 bytes (902MDYK) at 1.4 bps. First mentions (bold) are slower than confirmations (italics). phone matches calls ticket matches (edit dist) 66% 238 0 59% 82 1 55% 40 2 4.1% 4033 3+ Table 2: Phone numbers are used to confirm ticket matches. Good ticket matches (top row) are confirmed more often than poor matches (bottom row). Poor matches are more common because ticket numbers are relatively rare, and most calls don’t","PeriodicalId":274608,"journal":{"name":"Proceedings of the Workshop on Uphill Battles in Language Processing:\n Scaling Early Achievements to Robust Methods","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1900-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126753826","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}