Semantic Role Labeling in Conversational Chat using Deep Bi-Directional Long Short-Term Memory Networks with Attention Mechanism

Semantic Role Labeling in Conversational Chat using Deep Bi-Directional Long Short-Term Memory Networks with Attention Mechanism Valdi Rachman1, Rahmad Mahendra1, Alfan Farizki Wicaksono1, Ahmad Rizqi Meydiarso2, and Fariz Ikhwantri2 1Faculty of Computer Science, Universitas Indonesia, Depok, Indonesia 2Kata.ai, Jakarta, Indonesia valdi.rachman@gmail.com {rahmad.mahendra, alfan}@cs.ui.ac.id {rizqi, fariz}@kata.ai Abstract Semantic Role Labeling (SRL) has been ex- tensively studied, mostly for understanding English formal language. However, only a few reports exist for informal conversational text, especially for language being used in the chat- bot system. The challenges of informal texting language include a wide variety of slangs and abbreviations, short sentences, as well as dis- organized grammars. In this work, we propose an attention mechanism on top of Long Short- Term Memory Networks architecture for solv- ing SRL task on informal conversations. The task is evaluated on informal language used on an Indonesian chatting platform. Our pro- posed model achieves F1 score of 82.68%. 1 Introduction Semantic Role Labeling (SRL) is a task in Natu- ral Language Processing (NLP) which aims to au- tomatically assign semantic roles to each argument for each predicate in a given input sentence. As for a brief definition, given an input sentence, SRL sys- tem will give an output of ”Who did what to whom” with what as the predicate and who and whom being the argument of the predicate. SRL is an integral part of understanding natural language as it helps the machine to retrieve semantic information from the input. In practice, SRL has been widely used as one of the intermediate steps for many NLP tasks, some of which are information extraction (Surdeanu et al., 2003; Emanuele et al., 2013), machine trans- lation (Liu and Gildea, 2010; Lo et al., 2013), and question-answering (Shen and Lapata, 2007; Mos- chitti et al., 2003). In the chat bot industry, the bots need to under- stand semantic information of the user’s text in or- der to generate more personalized response. A chat bot system typically works in three steps, starting from understanding the incoming message from a user using Natural Language Understanding (NLU) system, tracking the dialogue using Dialogue State Tracking (DST), and finally generating a response using Natural Language Generation (NLG) system. SRL is a useful module under the NLU pipeline. To illustrate, suppose that the user sends a mes- sage to the bot for ordering the tickets. ("buy 4 ticket 4 US"). The SRL system is expected to extract the semantic information from the chat as follows. Predicate: buy Patient: tickets Beneficiary: us The system detects that the predicate is ”buy”, followed by ”tickets” as the patient, or the thing being bought. The beneficiary, or the recipient, is ”us”, the user. With these extracted semantic roles, the bot can ask the user further detail of request us- ing DST, such as the flight time, the airline, or user‘s budget. As we can see from the example, it is worth noting that the style of language used on chatting platform is different than those in formal text. In this …

Content truncated to 500 words