School of Electronic Engineering and Computer Science


Cognitive Science Research Group

Chatterbox Labs Ltd (originally Chatterbox Analytics Ltd) is a spinout company founded in 2011, providing intelligent software to allow brands to understand their customers and competitive landscape via social media.

It has become a stand-alone company generating revenue, with the potential to grow significantly in 2013. Chatterbox developed from human interaction research in the Cognitive Science research group (CogSci). CogSci’s expertise in experimental psychology, computational linguistics and human-computer interaction has led to the development of models and techniques for understanding and generating human-human and human-computer dialogue. One recent area of focus has been online interaction on social networking systems, and this has led to this commercial impact.

Our Impact

The research has had an economic impact via the creation of a spinout, Chatterbox, which offers a sentiment analysis API service in 8 languages. Chatterbox was incorporated as a limited company in November 2011, and now employs a technical and commercial team in London. The company now has:

  • Grant funding and revenue generation from API sales, ensured a profit in 2012.[I1,I2].
  • A sentiment analysis web module is being used by commercial partners [I1,I2,I4]
  • A public API in free and paid form, covering 8 languages with over 400 users, generating monthly revenue: [I3,I9].
  • Initial investment from Telefonica and grant funding [I1,I2,I4,I5].

Using Chatterbox sentiment analysis API allows clients to understand what people are saying in social conversations and has therefore become a major priority for: brands needing to respond to customer complaints or comments; marketing agencies trying to understand what a sector of the population likes or doesn't like; advertising agencies needing to characterise audience reaction to a campaign; broadcasters attempting to understand people's opinions and topics of discussion around a recent programme; and more.

Chatterbox’s technologies go beyond conventional social media analytics, which looks at crude indicators such as volume of traffic (or “buzz”), levels of individual connectedness, and general positivity of vocabulary, by applying much more sophisticated forms of analysis combined with a structural model of online conversations to enable focus on key interactions and participants.

Chatterbox’s agents automatically classify the components of conversations according to emotion and sentiment, allowing conversations as a whole to be characterised, providing customers with access to the users who are most important to them in terms of engaging their audience/customers and their opinions thereon. The use of robust computational methods from Chatterbox enables these tools to update dynamically, adjusting to new vocabulary and slang; and to be applied quickly to different languages (currently available in 8 including Mandarin). The purpose of emotional sentiment analysis is to exploit textual datasets.

Underpinning Research


The impact has resulted from the combination of two distinct but complementary strands of research within CogSci: theoretical and empirical research on multi-party interaction; and computational research into automatic language and dialogue processing.

The underpinning research is based on capabilities developed through a significant programme of research into the dynamics and structure of group interaction at Queen Mary. Projects led by Healey, Bryan-Kinns, McOwan and Kempson (Kings College London (KCL) and QMUL Visiting Professor), have advanced our understanding of how humans interact in groups from several perspectives. Healey's work in experimental psychology and psycholinguistics, and Kempson's in formal linguistics, have examined natural human interaction and communication via verbal, graphical and gestural modes. Bryan-Kinns has investigated how creative, engaging interfaces can enhance group interaction between remote and co-located individuals.

Analysing Non-Verbal Communication Using Motion Capture from MAT, QMUL on Vimeo.

McOwan has researched how robotic agents can interact within a human social setting, discovering features of human interaction which can be detected and/or simulated. Integrated together, these research paths have produced a body of understanding into the nature of multi-party human interaction: which features and modalities are important, which ones can be modelled to what extent, and which ones signify or encourage engagement and successful communication. Recently, these approaches and methods have been applied to online, text-based dialogue between people:

Healey and White's analysis of chat-room dialogues have shed light on how people engage via these emerging, novel modalities [R6]; while Poslad's work on mobile social networks has contributed the important element of understanding how they can influence other people's behaviour. In particular, CogSci's DiET experimental toolkit and paradigm has provided the only framework for experimenting directly with online text dialogue. By manipulating dialogue features (words, utterances) in real time, hypotheses can be tested and verified directly. This has led to significant findings as to how people use particular kinds of language (ellipsis, clarification etc.) and how we can understand what they mean by what they say [R2,R5].

The second strand is provided by Purver's research into computational linguistics and dialogue systems. Early work outside QM developed expertise in combining models of dialogue structure with robust machine learning methods to cope with noisy, real-world conversation data (e.g. [R3,R9]), focusing on relatively structured settings – e.g. the topic, decision and action-item detection agents for the first working automatic business meeting assistant system on the DARPA CALO project [R4].

From 2008 onwards, the two strands then came together in the CogSci research projects DynDial [G5], RISER [G3] and PPAT [G4]. These produced new models of overlapping, multi-party dialogue with its incremental, collaborative constructions, both theoretical (e.g. [R2]) and computational (e.g. [R7]). By using scalable, lightly supervised machine learning methods, the resulting techniques are now being applied to a range of fields including language acquisition [R7], dialogue analysis for mental health [R8], and social media [R1]. The latter has led to new, practical models attracting continuing research funding [G1,G2] and providing the basis for direct commercial impact via the spinout Chatterbox that is the subject of this impact statement.




  • [R1] Purver, Battersby (2012). Experimenting with Distant Supervision for Emotion
    Classification. In Proceedings of the 13th Conference of the European Chapter of the
    Association for Computational Linguistics (EACL 2012), pages 482-491, Avignon, France. ISBN 978-1-937284-19-0. (9 citations).
  • [R2] Howes, Purver, Healey, Mills, Gregoromichelaki (2011). On Incrementality in Dialogue: Evidence from Compound Contributions. Dialogue & Discourse 2(1), pages 279-311. ISSN 2152-9620. (5 citations).
  • [R3] Purver (2011). Topic Segmentation. In Tur & de Mori (eds.), Spoken Language
    Understanding: Systems for Extracting Semantic Information from Speech, Wiley. ISBN 978-0-470-68824-3. (17 citations).
  • [R4] Tur, Stolcke, Voss, Peters, Hakkani-Tür, Dowding, Favre, Fernández, Frampton,
    Frandsen, Frederickson, Graciarena, Kintzing, Leveque, Mason, Niekrasz, Purver,
    Riedhammer, Shriberg, Tien, Vergyri, Yang (2010). The CALO Meeting Assistant System. IEEE Transactions on Audio, Speech and Language Processing 18(6), pages 1601-1611. ISSN 1558-7916. (41 citations).
  • [R5] Healey, Swoboda, Umata, King (2007). Graphical Language Games: Interactional constraints on representational form. Cognitive Science, 31, 285-309. (40 citations).
  • [R6] Healey, White, Eshghi, Reeves, Light (2007). Communication Spaces. Computer Supported Co-operative Work 17 (2-3): 169-193. (16 citations).
  • [R7] Eshghi, Purver, Hough (2013). Probabilistic Induction for an Incremental Semantic Grammar. In Proceedings of the 10th International Conference on Computational Semantics (IWCS), pages 107-118, Potsdam, Germany.
  • [R8] Howes, Purver, McCabe (2013). Using Conversation Topics for Predicting Therapy Outcomes in Schizophrenia. Biomedical Informatics Insights 6(Suppl. 1), pages 39-50. ISSN 1178-2226, DOI 10.4137/BII.S11661.
  • [R9] Purver, Körding, Griffiths, Tenenbaum (2006). Unsupervised Topic Modelling for Multi-Party Spoken Discourse. In Proceedings of COLING/ACL, pages 17-24, Sydney, Australia. ISBN 1-932432-65-5. (75 citations).




  • [G1] Wiggins, Purver (QMUL); Battersby (Chatterbox) – ConCreTe: Concept Creation
    Technologies, EU FP7 611733. October 2013 – October 2016. €3,215,963 total cost.
  • [G2] Purver (QMUL), Battersby (Chatterbox) – Cultural Mobility through Social Intelligence, AHRC/Creativeworks London. January 2013 – November 2013. £15,000.
  • [G3] Purver – RISER: Robust Incremental SEmantic Resources for Dialogue, EPSRC EP/J010383/1. January 2012 - February 2013. £120,288 FEC.
  • [G4] Purver, McCabe – PPAT: Predicting Patient Adherence to Treatment from Dialogue Transcripts, EPSRC Pump-Priming Scheme EP/J501360/1. January 2012 - June 2012. £23,571.
  • [G5] Healey, Kempson – DynDial: The Dynamics of Conversational Dialogue, ESRC RES-062- 23-0962. January 2008 – March 2011. £698,951 FEC.
  • [G6] McOwan – LIREC: Living with Robots and intEractive Companions. EU FP7 215554. March 2008 – August 2012. €10,992,017 total cost.
  • [G7] Healey – DiET: Dialogue Experimentation Toolkit, EPSRC EP/D057426/1. September 2006 – August 2009. £178,525 EPSRC contribution.
  • [G8] Kempson, Healey – Leverhulme Network DialogueMatters. 2006 – 2008.
  • [G9] Healey – MAGIC: Multimodality and Graphics in Interactive Communication, EPSRC/ESRC PACCIT initiative L328 25 3003. November 2000 – October 2003.
  • [G10] Poslad – TRIDEC: Collaborative, Complex and Critical Decision Support in Evolving Crises, EU FP7 258723. September 2010 – August 2013. €8,898,342 total cost.


Impact Corroboration


  • [I1] Senior Technology Transfer Manager, Queen Mary Innovation: company formation, funding and investment, financial position.
  • [I2] Chief Technology Officer, Chatterbox Labs Ltd: all details given.
  • [I3] Chief Executive Officer, Mashape: commercial sale and use of Chatterbox APIs.
  • [I4] Research Manager, Channel 4: TSB project, Chatterbox platform development and use in media sector.
  • [I5] Head of Marketing, Barbican Centre: Creativeworks London project, Chatterbox research use in arts sector.
  • [I6] Project Manager, Scottsdale Cultural Council: Beacon project, Chatterbox use in public art.
  • [I7] Chatterbox at Downing Street round table technology talks:
  • [I8] Consumer statistics at for Chatterbox’s Sentiment Analysis API:
  • [I9] PraxisUnico Impact Awards finalists 2013: