Pragmatics: The Last Frontier
Pragmatics in Courier
When we started working on Courier’s conversational email summarizer, we knew that just using an extractive summarization approach wasn’t enough. At Codeq, we understand that the way people construct a discourse to communicate meaning is very complex because there are several levels of linguistic phenomena that interact transversally.
Computational modeling of language morphology and syntax, even though they are far from being “solved”, are well researched and understood phenomena in natural language processing (NLP). Computational distributional lexical and compositional semantics are areas of research that have recently received more attention since the popularization of word embeddings as an instrument to capture linguistic meaning.
Computational pragmatics, though, is a NLP area that is still widely unexplored.
But what is pragmatics? Paraphrasing van Dijk (2009), pragmatics is a hardly well-integrated field of research of linguistics that “is concerned with the use of language in social contexts” (Nordquist, 2017).
From a pragmatic point of view, transmission of meaning is a multifaceted phenomenon that “not only depends on structural and linguistic knowledge […], but also on the context of [each] utterance.” (Wikipedia contributors, 2017)
In Courier we have implemented modules that model pragmatics in conversational email discourse at many levels. Although far from a comprehensive pragmatic analysis, the implementation of all these diverse models has helped us better understand the content in emails and the communicative intent of email senders.
Discourse coherence
Since day one, at Codeq we knew that “textual coherence” was key in order to generate short summaries that still retain “the conceptual relationships that comprehenders use to construct a coherent mental representation accommodated by what is said in the discourse.” (Louwerse & Graesser, 2005: 217).
Courier’s conversational email summarization module features a two-staged hybrid strategy:
1) A Machine Learning (ML) layer that ranks sentences contained in emails by importance. Among other factors this ML layer uses the information extracted by a discourse modeling module that finds explicit discourse relations between sentences, that is, relations expressed using explicit discourse markers, such as, “for example”, “in addition”, “therefore”, etc.
2) A set of discourse coherence post-processing rules, inspired by the concept that a “piece of text is often not to be understood individually, but understood by linking it with other text units from its context” (Lin et al, 2012: 151), which ensure that the coherence of the generated summaries is not broken.
Speech acts
Drawing from John L. Austin and John R. Searle’s body of work on the Speech-act theory, in Courier we have implemented a module that analyzes sentences according to their illocutionary force, that is, their intention and their effect in the world.
We use a simplified version of the Stolcke et al. (2000) taxonomy that we apply to sentences in emails so that we can exploit this information in order to produce email summaries that are relevant from a pragmatic point of view.
This helps us know if a sentence expresses a commitment to perform an action, an issuance of a command or request, a question, a desire or need or just a plain statement.
Sarcasm
Sarcasm is an indirect act of speech “in which speakers convey their message in an implicit way.” (Tsur et al, 2010: 162)
Sarcasm detection is an important task of computational pragmatics, in particular, and natural language understanding (NLU) in general, especially suited as a complementary task for other NLU tasks, such as sentiment and/or emotion classification. Given its fundamental ambiguity, sarcasm is a common source of “misunderstanding in every day communication” (Davidov et al, 2010: 107)
The intrinsic difficulty of sarcasm detection, which is in many occasions even for humans very hard to recognize, arises from its indirect speech nature, as readers or listeners of sarcastic utterances need to use a combination of hints/cues to try to discern the true nature of those utterances. Shared common world knowledge (common ground), lexical and/or grammatical cues, extreme semantic intensifications, contextual and paralinguistic information (like tone of voice or facial gestures) are among the pieces of information that humans leverage to try to recognize if certain statements are in fact sarcastic or not.
Despite this difficulty and assuming that the following only account for a subset of sarcastic occurrences in communication, Kreuz & Caucci (idem) suggest, there are local factors, like punctuation, use of interjections, lexical repetition, etc, that can be used to identify at least a subset of sarcastic utterances.
In Courier we have implemented a module of analysis that relies only on local information in order to detect sarcasm in emotion-bearing sentences.
Last words
At Codeq we believe that attempting to capture pragmatic phenomena when performing automated text analysis tasks is crucial for the success of NLU projects.
In this short article, we have described the pragmatic perspectives from which Courier analyzes content from conversational emails to generate the best email summaries for our users.
Despite the intrinsic difficulties in 100% accurately detecting pragmatic phenomena, the NLP team at Codeq will keep working on these and possibly other computational pragmatics modules to keep pushing the boundaries of industrial NLP applications.
We hope you join us in expliring this last frontier of NLP. Get Courier!
References
Davidov, D., Tsur, O., & Rappoport, A. (2010). Semi-supervised recognition of sarcastic sentences in twitter and amazon. Proceedings of the Fourteenth Conference on Computational Natural Language Learning, Association for Computational Linguistics, 107–116
Austin, J.L. (1962). How to do things with words. Cambridge: Harvard University Press
Khodak, M., Saunshi, N., & Vodrahalli, K. (2017). A Large Self-Annotated Corpus for Sarcasm. CoRR, 1–5.
Kreuz, R. J., & Caucci, G. M. (2007). Lexical influences on the perception of sarcasm, 1–4.
Lin, Z., Ng, H. T., & Kan, M.-Y. (2012). A PDTB-styled end-to-end discourse parser. Natural Language Engineering, 20(02). pp 151–184.
Linguistic Society of America. “Meaning (Semantics and Pragmatics)”. Linguistic Society of America. https://www.linguisticsociety.org/resource/meaning-semantics-and-pragmatics (accessed 2017–10–03).
Nordquist, R, 2017. “What Is Pragmatics?”. ThoughtCo. https://www.thoughtco.com/pragmatics-language-1691654 (accessed 2017–10–03)
Searle, J.R (1969). Speech acts: an essay in the philosophy of language. Cambridge: Cambridge University Press.
Stolcke, A., Ries, K., Coccaro, N., Shriberg, E., Bates, R., Jurafsky, D., et al. (2000). Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational Linguistics, 26(3), 339–373.
Tsur, O., Davidov, D., & Rappoport, A. (2010). ICWSM–A great catchy name: Semi-supervised recognition of sarcastic sentences in online product reviews, 162–169.
van Dijk, T.A., 2009. Society and Discourse: How Social Contexts Influence Text and Talk. Cambridge University Press, Cambridge.
Wikipedia contributors. “Pragmatics”. Wikipedia, The Free Encyclopedia. https://en.wikipedia.org/w/index.php?title=Pragmatics&oldid=802453855 (accessed 2017–10–03).
