Building a Lightweight Semantic Model for Unsupervised Information Extraction on Short Listings

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1081–1092, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Building a Lightweight Semantic Model for Unsupervised Information Extraction on Short Listings Doo Soon Kim Accenture Technology Lab 50 W San Fernando St., San Jose, CA, 95113 Kunal Verma Accenture Technology Lab 50 W San Fernando St., San Jose, CA, 95113 {doo.soon.kim, k.verma, peter.z.yeh}@accenture.com Peter Z. Yeh Accenture Technology Lab 50 W San Fernando St., San Jose, CA, 95113 Abstract Short listings such as classified ads or product listings abound on the web. If a computer can reliably extract information from them, it will greatly benefit a variety of applications. Short listings are, however, challenging to process due to their informal styles. In this paper, we present an unsupervised information extrac- tion system for short listings. Given a cor- pus of listings, the system builds a seman- tic model that represents typical objects and their attributes in the domain of the corpus, and then uses the model to extract informa- tion. Two key features in the system are a se- mantic parser that extracts objects and their at- tributes and a listing-focused clustering mod- ule that helps group together extracted tokens of same type. Our evaluation shows that the semantic model learned by these two modules is effective across multiple domains. 1 Introduction Short listings such as classified ads or product list- ings are prevalent on the web. These texts are gen- erally concise – around 10 words in length. Fig. 1 shows some example listings. Due to the recent ex- plosive growth of such listings, extracting informa- tion from them becomes crucial for tasks such as faceted search and reasoning. For example, con- sider an online shopping site on which information about merchandises for sale is posted. Detecting brands/styles/features that are frequently mentioned in the postings would allow a company to design a better marketing strategy. Most Information Extraction (IE) techniques de- veloped for formal texts, however, would be inap- plicable to listings because of their informal and id- iosyncratic styles. For example, typos, abbreviations and synonyms often appear and should be resolved (e.g., apartment/apt, bike/bicycle). Symbols could have the special meanings (e.g., x in 2x2 in Fig. 1 indicates number of bedrooms and bathrooms). To- kenization based only on space is insufficient (e.g., RawlingBaseball in Fig. 1). Multiwords such as granite top should also be detected. Applying off- the-shelf parsers is infeasible because of unusual phrasal forms in most listings, such as a long se- quence of nouns/adjectives (e.g., “New Paint Wood Floors New Windows Gated Complex”) To address these challenges, several approaches have applied machine learning algorithms (Ghani et al., 2006) (Putthividhya and Hu, 2011) or an external knowledge base (Michelson and Knoblock, 2005). These approaches, however, commonly require hu- man supervision to produce training data or to build a knowledge base. This is expensive, requiring re- peated manual effort whenever a new domain or a …

Content truncated to 500 words