Computational processing of language has been investigated in many domains, such as news and biomedical text and speech and many popular tasks have applied to the email domain. Additionally, new tasks have been inspired by properties of email. In this sec-tion we survey some popular NLP applicasec-tions in email: email understanding, document classification and email summarization.
9.2.1 Email Understanding
Several studies have looked at modeling the intent and meaning of email messages. The most popular example is the classification of emails into Email Speech Acts. This tax-onomy is based on Speech Act Theory [3, 211] and characteristics of email. Speech acts include request for information and propose a meeting. Carvalho and colleagues [35, 50]
use a variety of techniques to classify email into speech acts, including collective classi-fication techniques based on multiple emails in a conversation. Speech acts in email have been used in other applications, such as Khoussainov and Kushmerick [130] who define an iterative algorithm whereby speech acts improve task identification, and the identified tasks are then used to improve speech acts. Leusky [145] defined a similar speech acts classification problem to identify a user’s role based on speech act patterns.
There are other classification systems based on email intentions. Lockerd and Selker’s [152] DriftCatcher, described above, classifies emails according to eight type of interac-tions. Goldstein and Sabin [204] learned speech acts and genre classification of email.
Lampert and colleagues have studied the nature of requests and commitments in email.
They defined a Verbal Response Modes (VRM) taxonomy of speech acts, which catego-rizes emails along two dimensions: literal meaning and pragmatic meaning [141]. They presented a careful definition of requests and commitments in email to guide annotations and measure the confidence of classifications [142]. They built an automated system to classify workplace emails as requests for actions or commitments to act [143]. Other work
by Carvalho has focused on modeling the intent of the email sender, such as determining appropriate recipients for an email [36] and alerting the user when an email is sent to an inappropriate recipient (leak detection) [34].
Another perspective has been to modify email to convey additional meta-data about the semantic contents of the message. The Semantic Email project at the University of Wash-ington defines several email patterns that can be annotated with semantic information, allowing for richer interfaces for these processes [163,164]. Examples include scheduling meetings, RSVPing for events and first come first served giveaways.
9.2.2 Document Classification
Document classification, a common task in natural language processing, has many appli-cations in email. The most well known example of document classification is that of spam filtering, for which naive Bayes models are the most popular [205]. However, there are many other classification settings, many of which are unique to email.
One popular application has been automated foldering, the task of automatically pre-dicting the appropriate folder for a given email message. One of the first such applications of this learning task to a real email client was by Segal and Kephart [214]. Crawford et al. [60] showed predicted labels to users in the inbox. Many others have evaluated differ-ent learning algorithms on email foldering, such as rule induction and k-nearest neighbors [187], RIPPER (a rule induction classifier) [49], Naive-Bayes [191, 199], SVMs [132], pattern detection from graphs [1] and other models [98]. Bekkerman et al. [10] conducted a comparison of numerous techniques on the Enron corpus.
Another document classification task is call center classification, whereby incoming emails to support centers are automatically routed to necessary specialists or replied to automatically [137]. Nenkova and Bagga [178] first identify emails to a call center that seek a response, then filter these by stage of the conversation. Busemann et al. [28] used features from shallow text processing, such as chunking, to classify call center emails.
Brutlag and Meek [25] identify some challenges of document classification in the
email domain. In particular, they recognize the need to adapt to changing behavior, handle heterogeneity of classes and speed/memory requirements needed for a desktop system.
9.2.3 Email Summarization
Automatic document summarization is an important application in natural language pro-cessing. As email volume has increased, so has the desire to apply summarization tech-niques to the email domain. However, email poses significant new challenges for summa-rization systems, such as threads instead of single documents, multiple authors, informal writing styles and different types of emails (announcements, discussions, question-answer exchanges, etc.)
There are also different types of summaries that can be generated for email: informa-tive and indicainforma-tive. An informainforma-tive summary replaces the contents of the message and summarizes the important points, allowing users to quickly read a long email thread or recall previous threads in preparation for meetings. In contrast, an indicative summary for a document indicates the contents without providing the information presented in the doc-ument. Other types of summaries include task oriented summaries, where the summary is a list of tasks mentioned in the email [52]. In this section, we focus on summaries of the first two types.
Informative Summaries
Sentence extraction is a popular approach to informative summarization. There have been several applications of sentence extraction summarization techniques to creating informa-tive summaries for email threads. Lam et al. [140] built an early system for evaluating the effectiveness of various summarization techniques. Fasil is a full summarization sys-tem used to generate summaries for mediums with limited interaction capabilities, such as automated phone systems [67]. However, the two dominant approaches to email summa-rization have relied on detecting question-answer pairs and using clue words.
The work of McKeown and colleagues use both standard summarization techniques
and question-answer extraction to summarize email threads. They assume that the first email in the thread introduces the topic of discussion and that subsequent emails address this topic without changing subjects. Based on this structure, they use centroids in vector space to identify key sentences in the initial email and related sentences in followup mes-sages [248]. Additional features relevant to this task are explored by Rambow et al. [195].
Sentences from each email are then tiled together to form a summary.
Shrestha and McKeown [220] extend this approach to identifying and extracting question-answer pairs directly, a goal similar to that of Murakoshi et al. [175]. Machine learning classification is used to identify which sentences in the original email are questions and which sentences in subsequent emails answer the questions. These extracted question-answer pairs and the extracted sentences from discussions are interleaved to form a single summary of the email thread [165].
Carenini et al. offer a second approach to email thread summarization based on con-versational elements in email [30]. They observe that followup emails will quote previous messages when responding, often layering the response with the original quotation. From this structure they construct a quotation graph linking fragments of emails with other mes-sages where they are quoted. The summary contains the most cohesive fragments, se-lected by word repetition in fragments. Followup work compared this approach favorably to graph methods based on semantic similarity [31].
Indicative Summaries
The purpose of an indicative summary is to give the reader an idea about what is discussed in a document without replacing the need to read the document. It should aid the user in deciding whether or not the document is relevant to the current task. Nenkova and Bagga [179] identify sentences in the first email of the thread that contain the main issue by finding the shortest sentence in the message that has the largest overlap of nouns, verbs, adjectives or adverbs with the message’s subject. Muresan et al. [176] extract high quality noun phrases to capture the gist of an email. A third approach is that of Newman [182,183]
for understanding the contents of an email collection. Newman clusters groups of topically related messages and then extracts a summary common to the whole group.