Revisiting the Predictability of Language: Response Completion in Social Media

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1489–1499, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Revisiting the Predictability of Language: Response Completion in Social Media Bo Pang Sujith Ravi Yahoo! Research 4401 Great America Parkway Santa Clara, CA 95054, USA bopang42@gmail.com sujith ravi@yahoo.com Abstract The question “how predictable is English?” has long fascinated researchers. While prior work has focused on formal English typically used in news articles, we turn to texts gener- ated by users in online settings that are more informal in nature. We are motivated by a novel application scenario: given the difficulty of typing on mobile devices, can we help re- duce typing effort with message completion, especially in conversational settings? We pro- pose a method for automatic response comple- tion. Our approach models both the language used in responses and the specific context pro- vided by the original message. Our experi- mental results on a large-scale dataset show that both components help reduce typing ef- fort. We also perform an information-theoretic study in this setting and examine the entropy of user-generated content, especially in con- versational scenarios, to better understand pre- dictability of user generated English. 1 Introduction How predictable is language? As early as 1951, long before large quantities of texts (or the means to pro- cess them) were easily available, Shannon had raised this question and proceeded to answer it with a set of clever analytical estimations. He studied the pre- dictability of printed English, or “how well can the next letter of a text be predicted when the preced- ing N letters are known” (Shannon, 1951). This was quantified as the conditional entropy, which measures the amount of information conveyed from statistics over the preceding context. In this paper, we discuss a novel application setting which mirrors the predictability study as defined by Shannon. (a) Google (b) Amazon (c) Netflix Figure 1: Query completion as users type into the “Search using Google” box on a browser, as well as the search box in Amazon and Netflix. Text completion for user-generated texts: Con- sider a user who is chatting with her contact or post- ing to a social media site using a mobile device. If we can predict the next word given the preceding words that were already typed in, we can help reduce the typing cost by offering users suggestions of pos- sible completions of their partially typed messages (e.g., in a drop-down list). If the intended word is ranked reasonably high, the user can select the word instead of typing it. Assuming a lower cost associ- ated with selections, this could lead to less typing effort for the user. An interface like this would be quite familiar to Web users today. Providing suggestions of possi- ble completions to partially typed queries, which we will refer to as query completion,1 is a common fea- ture of search boxes (Figure 1). In spite of the similarity in the …

Content truncated to 500 words