Product Name Classification for Product Instance Distinction

Copyright 2012 by Hye-Jin Min and Jong C. Park 26th Pacific Asia Conference on Language,Information and Computation pages 289–298 Product Name Classification for Product Instance Distinction Hye-Jin Min Jong C. Park CS Department, KAIST CS Department, KAIST Daejeon, Republic of Korea Daejeon, Republic of Korea hjmin@nlp.kaist.ac.kr park@cs.kaist.ac.kr Abstract Product names with a temporal cue in a product review often refer to several product instances purchased at different times. Previous approaches to product entity recognition and temporal information analysis do not take into account such temporal cues and thus fail to distinguish different product instances. We propose to formulate the resolution of such product names as a classification problem by utilizing time expressions, event features and other temporal cues for a classifier in two stages, detecting the existence of such temporal cues and identifying the purchase time. The empirical results show that term-based features and existing event-based features together enhance the performance of product instance distinction. 1 Introduction Traditional work on product entity recognition has been conducted on competing products for comparative opinion mining from forum data (Jindal and Liu, 2006; Ding et al., 2009; Li et al., 2010), but not on the same type of products purchased at different times, thus failing to distinguish products at the instance level. The use of temporal information would help to make such distinction, but previous studies of temporal information have been made only for the detection and determination of temporal relations between time expressions and events, through the relevant shared tasks, or TempEval-1 and TempEval-2 tasks (Pustejovsky et al., 2003; Verhagen et al., 2009; Verhagen et al., 2010), but not for the distinction of products. There is evidence that temporal relations between product instances of the same type are found quite often in product reviews, to give rise to important differences in the respective opinion of the reviewer. Consider Examples (1) and (2)1 below, with two product names other Levis 501s and these new ones that refer to the product instances that the customer bought. While the former refers to the past purchase, the latter refers to the recent purchase. (1) My husband has [other Levis 501s]a and [these new ones]b are different in the weight of fabric (light), (…) We are not happy with these jeans. (2) a. I don’t wear boots and I wear these jeans when I ride my bike. b. Jeans were exactly like one purchased from Khols or Sears. c. I’m done with buying jeans online. Resolving such different product instances properly is found crucial to identifying long-term customers, among others, whose opinions count at least as important as those of human annotators for influential reviews (Min and Park, 2012). Moreover, it is also crucial to identifying such long-term customer’s sentiment change over several purchases of the same product (cf. Min and Park, 2012). First, we note that sortal anaphoric expressions such as these jeans in (1) indicate the presence of a temporal cue but may also refer to the whole product as shown in (2a). We also …

Content truncated to 500 words