{"id":748432,"date":"2021-05-25T17:47:27","date_gmt":"2021-05-26T00:47:27","guid":{"rendered":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/?post_type=msr-blog-post&p=748432"},"modified":"2021-05-25T17:47:27","modified_gmt":"2021-05-26T00:47:27","slug":"when-does-text-prediction-benefit-from-additional-context-an-exploration-of-contextual-signals-for-chat-and-email-messages","status":"publish","type":"msr-blog-post","link":"https:\/\/cm-edgetun.pages.dev\/en-us\/research\/articles\/when-does-text-prediction-benefit-from-additional-context-an-exploration-of-contextual-signals-for-chat-and-email-messages\/","title":{"rendered":"When does text prediction benefit from additional context? An exploration of contextual signals for chat and email messages"},"content":{"rendered":"

By Stojan Trajanovski<\/a>, Chad Atalla<\/a>, Kunho Kim, Vipul Agarwal, Milad Shokouhi<\/a>, and Chris Quirk<\/a><\/b><\/p>\n

Email and chat communication tools are increasingly important for completing daily professional and personal tasks. Given the recent pandemic and shift to remote work, this usage has surged. The number of daily active users in Microsoft Teams, the largest business communication and chat platform, has increased from 20 million<\/a> pre-pandemic time in 2019 to more than 115 million<\/a> and 145 million in October 2020 and April 2021, respectively. On the other hand, email continues to be the crucial driver for formal communication showing ever increasing usage. Providing real-time suggestions for word or phrase auto-completions is known as text prediction. The efficiency of these communications is enhanced by suggesting highly accurate text predictions with low latency. Text prediction services have been deployed across popular communication tools and platforms such as Microsoft Outlook Text Predictions (opens in new tab)<\/span><\/a> or GMail Smart Compose (opens in new tab)<\/span><\/a> [1]<\/a>.<\/p>\n

Modern text prediction algorithms are based on large language models and generally rely on the prefix of a message (characters typed until cursor position) to create predictions. We study to what extent additional contextual signals<\/b><\/em> improve text predictions in chat and email messages in two of the largest commercial communication platforms Microsoft Teams and Outlook.
\nWe examine several signals accompanying the main message: composition time, subject,<\/em> and previous messages<\/em> (see the Table below).<\/p>\n\n\n\n\n\n\n\n
Contextual signal<\/strong><\/td>\nDetails<\/strong><\/td>\n<\/tr>\n
Composition time<\/em><\/td>\nIt is a contextual signal which can provide added value for text prediction, enabling suggestions with relevant date-time words, like “weekend”, “tonight”.<\/td>\n<\/tr>\n
Subject<\/em><\/td>\nMessage subjects often contain the purpose or summarized information of a message. In the email scenario, we use subject<\/em> as context. In the chat scenario, we use the chat window name<\/em> as a proxy for subject.<\/td>\n<\/tr>\n
Previous email<\/em><\/td>\nPrevious messages can provide valuable background information which influences the text of the current message being composed.\u00a0 In the email case, we create pairs of messages and replies.<\/td>\n<\/tr>\n
Previous chat messages<\/em><\/td>\nPrior message contextualization for chat scenario is much more complex. Chat conversations typically consist of many small messages sent in quick succession.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

We combine and encode these signals with the message body into a single “contextualized” string for the language model, using special tokens<\/em> to separate from the other signals (Figure 1).<\/p>\n

\"Context

Figure 1. Context extraction and encoding.<\/em><\/p><\/div>\n

We segment chat histories by message blocks<\/em> and time windows<\/em>. A series of uninterrupted messages sent by one sender is considered as a single message block<\/em>. Messages sent within the past N<\/em> minutes are within a time window<\/em>, which enforces recency as a proxy for relevance. We define three previous message context aggregation modes in the chat scenario (visualized in Figure 2), mimicking prior email context:<\/p>\n