---
title: "Dialogue systems — papers"
description: "Dialogue systems: 255 papers of the Elamur Research library, the most cited first."
canonical: "https://research.elamur.ai/free/en/topics/dialogue-systems/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/dialogue-systems/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Dialogue systems

# 🗨️ Dialogue systems

255 papers · page 1 of 11 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Talking with machines: chatbots, task assistants, and the structure of conversation.

A survey · about 19 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

A dialogue system is a computer program that takes part in a conversation: it reads or hears what a person says, works out what was meant in the light of everything said before, decides what to do, and answers in words. The field covers two broad families. Task-oriented systems help a user reach a concrete goal, such as booking a table, finding a train or querying a database. Open-domain systems, usually called chatbots, aim at free social conversation with no fixed task. Around both sits a third line of work that is less about building agents than about understanding conversation itself: how turns are organised, what each utterance does, how speakers refer back to earlier talk, and how emotion and stance move between participants.

The papers here span all of these. Some are corpus and annotation studies of human conversation, from telephone calls and business meetings to online chats; some propose architectures for understanding, tracking and generating dialogue; and a growing share treat dialogue as the main interface of large pretrained language models. The topic therefore sits where linguistics and engineering meet. It inherits ideas from [speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/) and [discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/), and in recent years it has been reshaped by [language models and pretraining](https://md.research.elamur.ai/free/en/topics/language-models/).

What follows describes the scope of the area, its history, its main problems and methods, the datasets that have defined progress, and the landmark works, people and places represented in this library.

## What it is

### Scope and core notions

Dialogue differs from ordinary text in several ways that shape every method in the field. It is produced by more than one speaker, so each contribution must be read relative to who said it and what it responds to. It unfolds over time, so a system must remember and update what has been established. It is full of short, context-dependent fragments: "yes", "the cheaper one", "what about Tuesday?" — utterances that are meaningless in isolation. And in speech it carries prosody, overlaps, hesitations and repairs.

A few notions recur throughout the literature:

- **Dialogue acts** (also called speech acts or communicative functions) label what an utterance does: asking, informing, requesting, agreeing, acknowledging. Tag sets range from four coarse classes to several dozen fine ones.
- **Dialogue state** is the system's running record of what the user wants, typically a set of slot–value pairs (cuisine = Italian, area = centre) together with a probability for each.
- **Dialogue policy** is the rule or learned function that chooses the next system action given the state.
- **Grounding and common ground** refer to the process by which participants establish mutual understanding, through acknowledgements, confirmations and clarification requests.
- **Response generation and response selection** are the two ways a system can produce its turn: by writing it word by word, or by choosing it from a pool of candidates.

### The classic pipeline

The textbook architecture of a spoken dialogue system is a pipeline: automatic speech recognition, natural language understanding (intent detection and slot filling), dialogue state tracking, a dialogue manager that applies a policy, natural language generation, and speech synthesis. Much of the research in this library targets one box of that pipeline. A large part of the neural era has been an attempt to replace the boxes with a single trainable model, and the most recent work builds the whole system on top of a large pretrained language model.

### How it differs from its neighbours

Dialogue systems overlap with [question answering](https://md.research.elamur.ai/free/en/topics/question-answering/), but a conversational QA system must interpret each question against previous turns. They overlap with [text generation](https://md.research.elamur.ai/free/en/topics/text-generation/), but a reply must fit a specific context and partner rather than a fixed input. They overlap with [speech and spoken language](https://md.research.elamur.ai/free/en/topics/speech/) processing, but are concerned with the interaction above the level of recognising words. The strongest link in this library is to [speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/): dialogue act annotation is the point where philosophical work on what utterances do becomes a practical labelling scheme for corpora.

## History and milestones

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 2000–2022; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [2000 — Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech](https://md.research.elamur.ai/free/en/papers/prag-2000-stolcke-dialogue-act-modeling-automatic-tagging/) | 1,447 |
| [2011 — Data-Driven Response Generation in Social Media](https://md.research.elamur.ai/free/en/papers/ritter-2011-data-driven-response-generation-social/) | 692 |
| [2015 — A Neural Conversational Model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/) | 1,759 |
| [2015 — Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems](https://md.research.elamur.ai/free/en/papers/wen-2015-semantically-conditioned-lstm-based-natural/) | 1,192 |
| [2016 — Deep Reinforcement Learning for Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2016-deep-reinforcement-learning-dialogue-generation/) | 1,479 |
| [2017 — DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset](https://md.research.elamur.ai/free/en/papers/li-2017-dailydialog-manually-labelled-multi-turn/) | 1,195 |
| [2017 — Adversarial Learning for Neural Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2017-adversarial-learning-neural-dialogue-generation/) | 1,044 |
| [2017 — Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders](https://md.research.elamur.ai/free/en/papers/zhao-2017-learning-discourse-level-diversity-neural/) | 986 |
| [2018 — Personalizing Dialogue Agents: I have a dog, do you have pets too?](https://md.research.elamur.ai/free/en/papers/zhang-2018-personalizing-dialogue-agents-i-have/) | 1,839 |
| [2019 — CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) | 1,541 |
| [2021 — Recipes for Building an Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/roller-2021-recipes-building-open-domain-chatbot/) | 1,206 |
| [2022 — LaMDA: Language Models for Dialog Applications](https://md.research.elamur.ai/free/en/papers/thoppilan-2022-lamda-language-models-dialog-applications/) | 1,314 |

### Early systems and the symbolic era

The idea of conversing with a computer is as old as artificial intelligence itself. Turing framed machine intelligence as a conversational test in 1950; Weizenbaum's ELIZA (1966) imitated a therapist with pattern-matching rules; Winograd's SHRDLU (around 1970) held typed dialogues about a world of blocks. Through the 1970s and 1980s, research on dialogue was largely symbolic: plan-based models of intention, and theories of discourse structure that tied utterances to the speakers' goals. The oldest item in this library is a 1982 description of the conversational environment of the ARIANE machine translation system, a reminder that "conversational" first meant interactive human–machine work rather than chat.

### Corpora, annotation and evaluation (1990s)

The 1990s brought a turn to data. Large recorded corpora of spontaneous talk — the HCRC Map Task, Switchboard telephone conversations, the Verbmobil negotiation dialogues — made it possible to study real conversation quantitatively. Annotation schemes such as DAMSL defined dialogue acts in a way that coders could apply consistently, and the kappa statistic became the standard test of whether they did. In this library, [The Reliability of a Dialogue Structure Coding Scheme](https://md.research.elamur.ai/free/en/papers/carletta-1997-reliability-dialogue-structure-coding-scheme/) and [Utilizing Statistical Dialogue Act Processing in Verbmobil](https://md.research.elamur.ai/free/en/papers/reithinger-1995-utilizing-statistical-dialogue-act-processing/) represent this period. At the same time, spoken dialogue systems for travel and timetable information were being deployed, and the need to compare them led to evaluation frameworks such as [PARADISE](https://md.research.elamur.ai/free/en/papers/walker-1997-paradise-framework-evaluating-spoken-dialogue/).

### Statistical dialogue modelling (2000s)

Around 2000, statistical methods took over. Dialogue act tagging was cast as sequence labelling with hidden Markov models and n-gram act grammars, most influentially in the Switchboard work of Stolcke, Jurafsky, Shriberg and colleagues. Dialogue management was framed as a decision problem: first as a Markov decision process trained with reinforcement learning, then, in the Cambridge school led by Steve Young, as a partially observable MDP (POMDP) that keeps a belief distribution over the user's goal to cope with recognition errors. Because such policies need many dialogues to learn, user simulation became a research topic in its own right. Meeting corpora such as ICSI extended the study of dialogue acts to multi-party talk.

### Standards and multidimensional annotation (late 2000s–2010s)

A parallel line, centred on Tilburg and Saarland, generalised dialogue act annotation into a multidimensional framework, DIT++, in which one utterance can serve several functions at once (for example, answering a question while also managing turn-taking). This work became the international standard ISO 24617-2, released in 2012 and revised in later years. Shared tasks on dialogue state tracking (DSTC, from 2013) gave the task-oriented community common benchmarks.

### The neural turn (2015 onward)

Papers on this topic by year 255 dated papers from 1982 to 2026; the busiest year is 2026 with 46, and 67% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 46 |
| 2025 | 4 |
| 2024 | 10 |
| 2023 | 11 |
| 2022 | 15 |
| 2021 | 24 |
| 2020 | 26 |
| 2019 | 20 |
| 2018 | 14 |
| 2017 | 17 |
| 2016 | 10 |
| 2015 | 8 |
| 2014 | 3 |
| 2013 | 4 |
| 2012 | 3 |
| 2011 | 5 |
| 2010 | 4 |
| 2009 | 2 |
| 2007 | 4 |
| 2005 | 1 |
| 2004 | 6 |
| 2003 | 1 |
| 2002 | 2 |
| 2001 | 1 |
| 2000 | 3 |
| 1999 | 3 |
| 1998 | 2 |
| 1997 | 2 |
| 1995 | 2 |
| 1988 | 1 |
| 1982 | 1 |

In 2015 several groups showed that sequence-to-sequence neural networks, trained on large logs of conversation, could produce replies end to end with no hand-written rules. Within a year the field had found the central weakness of these models — dull, all-purpose answers in the manner of "I don't know" — and a stream of work tried to fix it with mutual-information decoding, reinforcement learning, adversarial training and latent variables. Retrieval-based chatbots that select responses from candidates developed in parallel. Large public datasets (Ubuntu, PersonaChat, DailyDialog, MultiWOZ) and open platforms such as ParlAI made comparison possible. The years from 2015 to 2019 are the period in which the most cited papers of this library cluster.

### Pretrained models and large language models (2019 onward)

From 2019, dialogue models were built by pretraining Transformers on huge conversational corpora, chiefly Reddit: DialoGPT, Meena, PLATO and the BlenderBot series. Attention shifted from fluency to qualities that scale alone did not deliver — safety, factual grounding, consistency of persona, empathy. LaMDA and Sparrow added tool use, retrieval and fine-grained human feedback; reinforcement learning from human feedback became the standard route to the general-purpose chat assistants that appeared from late 2022. The most recent papers here concern speech-native dialogue models, evaluation of large models over many turns, and whether such models actually understand dialogue structure.

## Main problems and tasks

### Understanding

- **Dialogue act recognition**: assigning communicative functions to utterances, using words, prosody and the preceding acts. It is one of the most heavily studied tasks in this library, with work on telephone speech, meetings, live chats, Twitter, email threads, Wikipedia talk pages and Mastodon.
- **Intent detection and slot filling**: the language understanding step of a task-oriented assistant. Nested requests have pushed this toward tree-shaped and executable representations, a link to [semantic parsing](https://md.research.elamur.ai/free/en/topics/semantic-parsing/).
- **Reference in dialogue**: resolving pronouns, ellipsis and discourse deixis ("that", meaning a whole earlier proposal), a link to [coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/). The CODI-CRAC shared tasks of 2021 and 2022 were devoted to anaphora in conversation.
- **Discourse structure of conversation**: which earlier turn each turn responds to, and with what relation, especially in multi-party chats where threads interleave.
- **Emotion and stance**: emotion recognition in conversation, agreement and disagreement, sarcasm and nastiness, and early signs that a conversation will go badly.

### Managing the conversation

- **Dialogue state tracking**: maintaining the user's goal across turns despite recognition errors, corrections and changes of mind.
- **Policy learning**: choosing what to do next, often trained with reinforcement learning against simulated or real users.
- **Knowledge access**: querying databases and knowledge bases, and grounding answers in retrieved documents or the web.

### Responding

- **Natural language generation** from a dialogue act and its slot values, in task-oriented systems.
- **Open-domain response generation and selection**, with persistent problems of genericness, repetition, contradiction and hallucination.
- **Specialised conversational skills**: persona consistency, empathy and emotional support, persuasion, negotiation, tutoring.

### Evaluating

Evaluation is a long-standing difficulty. Task success and dialogue length can be measured for goal-directed systems, but open conversation has no single correct reply, and word-overlap metrics borrowed from [machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) correlate poorly with human judgments. The field has moved through user-satisfaction models, learned evaluators, reference-free metrics and structured human evaluation protocols. This connects the topic to [evaluation and benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/).

## Methods and schools

### Rule-based and plan-based systems

Early systems, and many deployed ones, rely on hand-written rules: finite-state dialogue flows, frame-filling forms, templates for generation. Plan-based theories modelled dialogue as the recognition of speakers' plans and intentions. These approaches are transparent and controllable, which is why the template generator and the scripted flow survived long into the statistical era; their weaknesses are cost of construction and brittleness outside the anticipated paths.

### Statistical and decision-theoretic approaches

The statistical school treated each component as a probabilistic model learned from data. For understanding, this meant HMMs, Bayesian networks, maximum entropy and CRF taggers over words and prosody. For management, it meant MDPs and POMDPs whose policies are optimised by reinforcement learning, with belief tracking to absorb uncertainty from speech recognition. Cambridge work on statistical spoken dialogue — user simulation, belief tracking, statistical generation — is the best-known example, and several of its papers are among the most cited here.

### Neural end-to-end models

The neural school replaced pipelines with encoder–decoder networks trained on raw conversations. Its main lines of development were:

- **Hierarchical context models**, which encode each utterance and then the sequence of utterances, as in HRED and its latent-variable successor VHRED.
- **Diversity-oriented training**, through mutual-information reranking, latent variables (variational autoencoders), reinforcement learning with designed rewards, and adversarial discriminators.
- **Retrieval and matching models**, which score candidate responses against the context, with increasingly fine matching between the candidate and each previous turn.
- **Memory and copying**, which let task-oriented models read knowledge-base entries and copy entity names into replies.
- **Single-model task-oriented systems**, which emit the belief state as text and then the reply, removing separately trained trackers.

### Pretraining and large language models

The current paradigm is to pretrain a large Transformer on conversation or general text and then adapt it: by fine-tuning on curated skill datasets, by conditioning on retrieved knowledge, and by learning from human preferences. Specialised encoders pretrained on dialogue (for task-oriented understanding, or for efficient response selection) sit alongside large generative models. In-context learning has brought even classical tasks such as state tracking into the prompting paradigm.

### Main debates

- **Pipeline versus end-to-end.** Modular systems are interpretable and controllable; end-to-end systems avoid error propagation and hand-engineering but are harder to constrain. Much recent work seeks hybrids that keep explicit intermediate structure.
- **Generation versus retrieval.** Retrieval guarantees fluent, vetted text but cannot say anything new; generation is flexible but can be dull or wrong. Retrieve-and-refine designs combine the two.
- **Scale versus targeted data.** Several landmark papers report that growing a model improves fluency but does little for safety, groundedness or consistency, which need specific data, rules or feedback.
- **What to measure.** The repeated finding that automatic metrics disagree with people keeps human evaluation central and keeps the design of new metrics an active area.
- **Linguistic structure.** A linguistic tradition argues that dialogue acts, discourse relations and common ground should be modelled explicitly; neural practice often leaves them implicit. Recent papers test whether large models recover such structure on their own.

## Datasets, resources and benchmarks

The history of the field can be read through its corpora. Among those well represented in this library:

| Resource | Kind of dialogue | What it supports |
| --- | --- | --- |
| Switchboard (with SWBD-DAMSL acts) | spontaneous telephone conversation | dialogue act tagging, prosody |
| HCRC Map Task | task-oriented spoken dialogue | dialogue structure annotation |
| ICSI Meeting Recorder (MRDA) | multi-party meetings | dialogue acts, agreement, adjacency pairs |
| ISO 24617-2 / DialogBank | multiple corpora | multidimensional dialogue act annotation |
| Ubuntu Dialogue Corpus | technical-support chat | response selection |
| PersonaChat / ConvAI2 | crowdsourced chit-chat with personas | persona consistency, engagingness |
| DailyDialog | everyday two-person dialogues | generation with act and emotion labels |
| MultiWOZ (2.0–2.4) | multi-domain task-oriented chat | dialogue state tracking |
| CoQA, DREAM, MuTual | conversation over texts | conversational reading comprehension |
| STAC | multi-party game chat | discourse structure of dialogue |
| DialogSum, MediaSum | everyday talk, interviews | dialogue summarization |

Other notable resources here include the Douban Conversation Corpus for Chinese response selection, Frames for multi-frame travel dialogues, Taskmaster-1, the TOP dataset of hierarchical task requests, CoSQL for conversational database querying, PersuasionForGood, ESConv for emotional support, OpenDialKG for knowledge-graph walks, Wizard of the Internet, ProsocialDialog, SODA, MathDial and MT-Bench-101. On the software side, ParlAI unified many dialogue tasks behind one interface, and ConvoKit offers tools for analysing conversations as data.

## Landmark papers in this library

Citations of the papers on this topic, by when they were published Papers published in 2015–2019 gather 55% of the 47,731 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 50 papers, 17 citations |
| 2020–2024 | 86 papers, 13,634 citations |
| 2015–2019 | 69 papers, 26,025 citations |
| 2010–2014 | 19 papers, 2,727 citations |
| 2005–2009 | 7 papers, 731 citations |
| 2000–2004 | 13 papers, 2,878 citations |
| 1995–1999 | 9 papers, 1,658 citations |
| 1985–1989 | 1 papers, 35 citations |
| 1980–1984 | 1 papers, 26 citations |

**[Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech](https://md.research.elamur.ai/free/en/papers/prag-2000-stolcke-dialogue-act-modeling-automatic-tagging/)** (2000). Stolcke and colleagues annotated a large part of Switchboard with 42 act labels and modelled a conversation as a hidden Markov model whose states are acts, combining word, prosodic and act-sequence evidence. It set the template for statistical act tagging and supplied the tag set and data that later neural classifiers were tested on for two decades.

**[PARADISE: A Framework for Evaluating Spoken Dialogue Agents](https://md.research.elamur.ai/free/en/papers/walker-1997-paradise-framework-evaluating-spoken-dialogue/)** (1997). Walker and co-authors proposed scoring a spoken agent by regressing user satisfaction on task success (measured with kappa over an attribute–value description of the task) and on dialogue costs. Separating what a task requires from how a strategy achieves it made different systems comparable, and the framework shaped spoken dialogue evaluation for years.

**[Agenda-Based User Simulation for Bootstrapping a POMDP Dialogue System](https://md.research.elamur.ai/free/en/papers/schatzmann-2007-agenda-based-user-simulation-bootstrapping/)** (2007). This Cambridge paper modelled a simulated user as a goal plus a stack of pending acts, with few enough parameters to set by hand. A policy trained purely against it worked with real people, showing how statistical dialogue managers could be built before any real data existed.

**[Data-Driven Response Generation in Social Media](https://md.research.elamur.ai/free/en/papers/ritter-2011-data-driven-response-generation-social/)** (2011). Ritter, Cherry and Dolan treated replying to a Twitter post as phrase-based translation from status to response, with fixes to stop the system parroting its input. It was the first widely noticed demonstration that conversational replies could be learned from raw data, and the direct ancestor of the neural generators that followed.

**[A Neural Conversational Model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/)** (2015). Vinyals and Le trained a plain LSTM sequence-to-sequence model on helpdesk logs and movie subtitles and showed that it could hold simple conversations and even troubleshoot. It became the baseline of neural dialogue generation and, with its examples of inconsistent answers, also an inventory of the problems to come.

**[Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models](https://md.research.elamur.ai/free/en/papers/serban-2015-building-end-end-dialogue-systems/)** (2015). Serban and colleagues adapted a hierarchical encoder–decoder to conversation, with separate networks for words within a turn and for the sequence of turns. Its observation that the most probable replies tend to be generic framed much of the following research, including the latent-variable VHRED model of [A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues](https://md.research.elamur.ai/free/en/papers/serban-2016-hierarchical-latent-variable-encoder-decoder/).

**[Deep Reinforcement Learning for Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2016-deep-reinforcement-learning-dialogue-generation/)** (2016). Li and co-authors let two copies of a neural chatbot converse and optimised them with policy gradients for rewards favouring replies that are easy to answer, non-repetitive and coherent. The work recast open-domain generation as long-horizon decision-making; its sequel, [Adversarial Learning for Neural Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2017-adversarial-learning-neural-dialogue-generation/), replaced hand-built rewards with a learned human-versus-machine discriminator.

**[Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems](https://md.research.elamur.ai/free/en/papers/wen-2015-semantically-conditioned-lstm-based-natural/)** (2015). Wen and colleagues added to an LSTM a gated vector of the dialogue act and slots still to be expressed, which the network gradually "uses up" while writing. It showed that generation for task-oriented systems could be learned from data instead of templates, and became the standard neural baseline for that step.

**[The Ubuntu Dialogue Corpus](https://md.research.elamur.ai/free/en/papers/lowe-2015-ubuntu-dialogue-corpus-large-dataset/)** (2015). Lowe and colleagues extracted two-person dialogues from Ubuntu support channels and defined next-response selection as an evaluation task needing no manual labels. Its scale made it the standard benchmark for retrieval chatbots, used for example by the [Sequential Matching Network](https://md.research.elamur.ai/free/en/papers/wu-2017-sequential-matching-network-new-architecture/) and the [Deep Attention Matching Network](https://md.research.elamur.ai/free/en/papers/zhou-2018-multi-turn-response-selection-chatbots/).

**[Personalizing Dialogue Agents: I have a dog, do you have pets too?](https://md.research.elamur.ai/free/en/papers/zhang-2018-personalizing-dialogue-agents-i-have/)** (2018). The most cited paper here introduced PersonaChat, in which crowdworkers chat while playing short written personas, and showed that conditioning a model on its own persona makes it more consistent. The dataset anchored the [ConvAI2 competition](https://md.research.elamur.ai/free/en/papers/dinan-2019-second-conversational-intelligence-challenge-convai2/), the [Dialogue NLI](https://md.research.elamur.ai/free/en/papers/welleck-2019-dialogue-natural-language-inference/) consistency work, and large-scale persona training on Reddit.

**[CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/)** (2019). Reddy, Chen and Manning built a dataset of question–answer conversations about passages from seven domains, where most follow-up questions depend on earlier turns. It joined dialogue with reading comprehension and remains a standard benchmark for context-dependent QA.

**[DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://md.research.elamur.ai/free/en/papers/zhang-2019-dialogpt-large-scale-generative-pre/)** (2019). Microsoft researchers pretrained GPT-2 on Reddit reply chains and released the weights, giving the community a strong pretrained conversational model. It began the line of large pretrained chatbots that includes Google's [Meena](https://md.research.elamur.ai/free/en/papers/adiwardana-2020-towards-human-like-open-domain/), Baidu's [PLATO](https://md.research.elamur.ai/free/en/papers/bao-2020-plato-pre-trained-dialogue-generation/) and Facebook's BlenderBot.

**[Recipes for Building an Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/roller-2021-recipes-building-open-domain-chatbot/)** (2021). The BlenderBot paper showed that fine-tuning on a blend of skill datasets — persona, empathy, knowledge — and careful choice of decoding matter as much as size for how people rate a chatbot. It was followed by internet-augmented generation and the deployed [BlenderBot 3](https://md.research.elamur.ai/free/en/papers/shuster-2022-blenderbot-3-deployed-conversational-agent/).

**[LaMDA: Language Models for Dialog Applications](https://md.research.elamur.ai/free/en/papers/thoppilan-2022-lamda-language-models-dialog-applications/)** (2022). Google's family of dialogue Transformers up to 137B parameters found that scaling improved conversational quality but not safety or factual grounding, and that modest amounts of annotated data plus calls to external tools improved both. Together with DeepMind's [Sparrow](https://md.research.elamur.ai/free/en/papers/glaese-2022-improving-alignment-dialogue-agents-targeted/), which broke helpful and harmless behaviour into explicit rules for human raters, it marks the transition to aligned conversational assistants.

**[Deal or No Deal? End-to-End Learning of Negotiation Dialogues](https://md.research.elamur.ai/free/en/papers/lewis-2017-deal-or-no-deal-end/)** (2017). Agents trained to split a set of items through free-form chat, first by imitating people and then by self-play and look-ahead rollouts, became better negotiators than their supervised start. The paper made negotiation a standard testbed for goal-driven dialogue that combines language with strategy.

## People

Authors with the most papers on this topic Jason Weston leads with 10 papers; the 12 authors shown wrote 70 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Jason Weston](https://md.research.elamur.ai/free/en/authors/jason-weston/) | 10 papers · 5,571 citations |
| [Volha Petukhova](https://md.research.elamur.ai/free/en/authors/volha-petukhova/) | 7 papers · 570 citations |
| [Emily Dinan](https://md.research.elamur.ai/free/en/authors/emily-dinan/) | 6 papers · 4,380 citations |
| [Jianfeng Gao](https://md.research.elamur.ai/free/en/authors/jianfeng-gao/) | 6 papers · 3,211 citations |
| [Iulian Serban](https://md.research.elamur.ai/free/en/authors/iulian-vlad-serban/) | 6 papers · 3,027 citations |
| [Marilyn A. Walker](https://md.research.elamur.ai/free/en/authors/marilyn-walker/) | 6 papers · 947 citations |
| [Harry Bunt](https://md.research.elamur.ai/free/en/authors/harry-bunt/) | 6 papers · 551 citations |
| [Ryan Lowe](https://md.research.elamur.ai/free/en/authors/ryan-lowe/) | 5 papers · 2,503 citations |
| [Joelle Pineau](https://md.research.elamur.ai/free/en/authors/joelle-pineau/) | 5 papers · 2,502 citations |
| [Michael Strube](https://md.research.elamur.ai/free/en/authors/michael-strube/) | 5 papers · 348 citations |
| [Dan Jurafsky](https://md.research.elamur.ai/free/en/authors/dan-jurafsky/) | 4 papers · 4,102 citations |
| [Michel Galley](https://md.research.elamur.ai/free/en/authors/michel-galley/) | 4 papers · 2,748 citations |

- **[Jason Weston](https://md.research.elamur.ai/free/en/authors/jason-weston/)** (New York University) is the most prolific author on this topic in the library. His papers here run from the [ParlAI](https://md.research.elamur.ai/free/en/papers/miller-2017-parlai-dialog-research-software-platform/) platform and PersonaChat to the BlenderBot recipes, Dialogue NLI and adversarial safety data.
- **[Emily Dinan](https://md.research.elamur.ai/free/en/authors/emily-dinan/)** co-authored PersonaChat and the BlenderBot paper, organised ConvAI2, and led work on [dialogue safety](https://md.research.elamur.ai/free/en/papers/dinan-2019-build-it-break-it-fix/) and on [reducing gender bias](https://md.research.elamur.ai/free/en/papers/dinan-2020-queens-are-powerful-too-mitigating/) in generated dialogue.
- **[Kurt Shuster](https://md.research.elamur.ai/free/en/authors/kurt-shuster/)** worked on blending conversational skills, [internet-augmented generation](https://md.research.elamur.ai/free/en/papers/komeili-2022-internet-augmented-dialogue-generation/) and the deployed BlenderBot 3.
- **[Jianfeng Gao](https://md.research.elamur.ai/free/en/authors/jianfeng-gao/)** (Microsoft) appears on several foundational neural papers: context-sensitive response generation, reinforcement learning for dialogue, end-to-end information-access agents and DialoGPT.
- **[Michel Galley](https://md.research.elamur.ai/free/en/authors/michel-galley/)** (Microsoft) links two eras: [agreement and disagreement detection](https://md.research.elamur.ai/free/en/papers/galley-2004-identifying-agreement-disagreement-conversational-speech/) in meetings, and neural response generation through DialoGPT.
- **[Dan Jurafsky](https://md.research.elamur.ai/free/en/authors/dan-jurafsky/)** (Stanford University) contributed to statistical dialogue act modelling of Switchboard, including [cues for dialog acts](https://md.research.elamur.ai/free/en/papers/jurafsky-1998-lexical-prosodic-syntactic-cues-dialog/), and later to reinforcement and adversarial learning for neural chatbots.
- **[Elizabeth Shriberg](https://md.research.elamur.ai/free/en/authors/elizabeth-shriberg/)** brought prosody and meeting speech into the study of dialogue acts, with the Switchboard act model and the [ICSI MRDA corpus](https://md.research.elamur.ai/free/en/papers/shriberg-2004-icsi-meeting-recorder-dialog-act/).
- **[Marilyn A. Walker](https://md.research.elamur.ai/free/en/authors/marilyn-walker/)** (University of California, Santa Cruz) is known here for PARADISE and the [DATE tagging scheme](https://md.research.elamur.ai/free/en/papers/walker-2001-date-dialogue-act-tagging-scheme/) for evaluating spoken systems, and later for corpora of [sarcasm in dialogue](https://md.research.elamur.ai/free/en/papers/oraby-2016-creating-characterizing-diverse-corpus-sarcasm/).
- **[Harry Bunt](https://md.research.elamur.ai/free/en/authors/harry-bunt/)** (Tilburg University) led the multidimensional approach to dialogue acts that became [ISO 24617-2](https://md.research.elamur.ai/free/en/papers/bunt-2012-iso-24617-2-semantically-based/), from the [first proposal for a standard](https://md.research.elamur.ai/free/en/papers/bunt-2010-iso-standard-dialogue-act-annotation/) to its [second edition](https://md.research.elamur.ai/free/en/papers/bunt-2020-iso-standard-dialogue-act-annotation/).
- **[Volha Petukhova](https://md.research.elamur.ai/free/en/authors/volha-petukhova/)** (Saarland University) co-developed the ISO standard and worked on the [semantics of discourse markers](https://md.research.elamur.ai/free/en/papers/petukhova-2009-multidimensional-semantics-discourse-markers-spoken/) and multidimensional segmentation and classification.
- **[David Traum](https://md.research.elamur.ai/free/en/authors/david-traum/)** (University of Southern California), a long-standing theorist of grounding, co-authored the ISO dialogue act work; his recent papers here concern persuasion policies and personalised dialogue evaluation.
- **[Steve Young](https://md.research.elamur.ai/free/en/authors/steve-young/)** led the Cambridge statistical dialogue programme represented by user simulation, [word-based state tracking](https://md.research.elamur.ai/free/en/papers/henderson-2014-word-based-dialog-state-tracking/) and LSTM-based generation, with [Tsung-Hsien Wen](https://md.research.elamur.ai/free/en/authors/tsung-hsien-wen/), [Nikola Mrkšić](https://md.research.elamur.ai/free/en/authors/nikola-mrksic/) and [Pei-Hao Su](https://md.research.elamur.ai/free/en/authors/pei-hao-su/).
- **[Iulian Serban](https://md.research.elamur.ai/free/en/authors/iulian-vlad-serban/)**, **[Ryan Lowe](https://md.research.elamur.ai/free/en/authors/ryan-lowe/)** and **[Joelle Pineau](https://md.research.elamur.ai/free/en/authors/joelle-pineau/)** formed the Montreal group behind HRED and VHRED, the Ubuntu corpus and [ADEM](https://md.research.elamur.ai/free/en/papers/lowe-2017-towards-automatic-turing-test-learning/), an early learned evaluator for responses.
- **[Michael Strube](https://md.research.elamur.ai/free/en/authors/michael-strube/)** (Heidelberg Institute for Theoretical Studies) worked on reference in spoken dialogue, from [pronoun resolution](https://md.research.elamur.ai/free/en/papers/strube-2003-machine-learning-approach-pronoun-resolution/) and discourse deixis to the [CODI-CRAC 2021 shared task](https://md.research.elamur.ai/free/en/papers/khosla-2021-codi-crac-2021-shared-task/).
- **[Zhou Yu](https://md.research.elamur.ai/free/en/authors/zhou-yu/)** (Columbia University) appears with socially oriented dialogue: [Persuasion for Good](https://md.research.elamur.ai/free/en/papers/wang-2019-persuasion-good-personalized-persuasive-dialogue/), [emotional support](https://md.research.elamur.ai/free/en/papers/liu-2021-emotional-support-dialog-systems/) and the [MIDAS](https://md.research.elamur.ai/free/en/papers/yu-2021-midas-dialog-act-annotation-scheme/) act scheme for open-domain conversation.
- **[Julia Hirschberg](https://md.research.elamur.ai/free/en/authors/julia-hirschberg/)** (Columbia University) brings the prosody and pragmatics of spoken interaction, from meeting agreement to lexical entrainment and a re-aligned Switchboard act corpus.
- **[Giuseppe Riccardi](https://md.research.elamur.ai/free/en/authors/giuseppe-riccardi/)** (University of Trento) and **[Giuseppe Carenini](https://md.research.elamur.ai/free/en/authors/giuseppe-carenini/)** (University of British Columbia) represent annotation and discourse parsing of spoken and written conversations.

## Institutions and places

Where the authors on this topic work now Carnegie Mellon University is the current home of authors of 15 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) | 15 papers |
| [Microsoft (United States)](https://md.research.elamur.ai/free/en/companies/microsoft/) | 15 papers |
| [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/) | 14 papers |
| [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/) | 11 papers |
| [Columbia University](https://md.research.elamur.ai/free/en/universities/columbia-university/) | 10 papers |
| [National University of Singapore](https://md.research.elamur.ai/free/en/universities/national-university-of-singapore/) | 10 papers |
| [Beihang University](https://md.research.elamur.ai/free/en/universities/beihang-university/) | 9 papers |
| [Georgia Institute of Technology](https://md.research.elamur.ai/free/en/universities/georgia-institute-of-technology/) | 9 papers |
| [Saarland University](https://md.research.elamur.ai/free/en/universities/saarland-university/) | 9 papers |
| [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/) | 9 papers |

The institutional map follows the field's two traditions. Industrial laboratories dominate the neural and pretrained era: [Microsoft](https://md.research.elamur.ai/free/en/companies/microsoft/), [Meta](https://md.research.elamur.ai/free/en/companies/meta/), [Google](https://md.research.elamur.ai/free/en/companies/google/) and [Amazon](https://md.research.elamur.ai/free/en/companies/amazon/) are all current homes of authors here. Among universities, [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/), long a centre of spoken dialogue research, leads, followed by [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/), [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/), [Columbia University](https://md.research.elamur.ai/free/en/universities/columbia-university/), the [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/), the [Georgia Institute of Technology](https://md.research.elamur.ai/free/en/universities/georgia-institute-of-technology/), the [University of Southern California](https://md.research.elamur.ai/free/en/universities/university-of-southern-california/) and the [University of Illinois Urbana-Champaign](https://md.research.elamur.ai/free/en/universities/university-of-illinois-urbana-champaign/).

In Europe, the [University of Cambridge](https://md.research.elamur.ai/free/en/universities/university-of-cambridge/) is the historic home of statistical spoken dialogue, and [Saarland University](https://md.research.elamur.ai/free/en/universities/saarland-university/) of the multidimensional annotation tradition; the [Centre National de la Recherche Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/) is linked to work on discourse structure in multi-party chat. In Asia, [Tsinghua University](https://md.research.elamur.ai/free/en/universities/tsinghua-university/), [Beihang University](https://md.research.elamur.ai/free/en/universities/beihang-university/), [Soochow University](https://md.research.elamur.ai/free/en/universities/soochow-university/), the [National University of Singapore](https://md.research.elamur.ai/free/en/universities/national-university-of-singapore/) and [The University of Tokyo](https://md.research.elamur.ai/free/en/universities/the-university-of-tokyo/) are prominent, reflecting large Chinese and Singaporean contributions to response selection, task-oriented modelling and dialogue discourse parsing.

Countries of the authors on this topic United States comes first: authors working there now wrote 127 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| United States | 127 papers |
| China | 70 papers |
| United Kingdom | 39 papers |
| Germany | 29 papers |
| Canada | 20 papers |
| Japan | 20 papers |
| Netherlands | 17 papers |
| Hong Kong | 16 papers |
| Singapore | 16 papers |
| France | 14 papers |

The United States is the leading country by far, followed by China, the United Kingdom and Germany. The venues mirror the field's double identity: the annual SIGDIAL meeting, devoted to discourse and dialogue, and LREC, where corpora and annotation schemes appear, stand beside the main ACL and EMNLP conferences, where most neural models were published.

Where the papers on this topic were published Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue carries the most, 20 of 255.

**The numbers**

| Venue | Papers |
| --- | --- |
| Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue | 20 papers |
| Proceedings of the Fifteenth Language Resources and Evaluation Conference | 19 papers |
| Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics | 8 papers |
| Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) | 7 papers |
| Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) | 6 papers |
| Findings of the Association for Computational Linguistics: EMNLP 2023 | 5 papers |
| Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing | 5 papers |
| Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 5 papers |
| Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics | 5 papers |
| Transactions of the Association for Computational Linguistics | 5 papers |

## Open questions

- **Evaluation.** There is still no automatic measure of open conversation that people trust. Learned and reference-free metrics such as [USR](https://md.research.elamur.ai/free/en/papers/mehri-2020-usr-unsupervised-reference-free-evaluation/) and [DialoGPT-based scoring](https://md.research.elamur.ai/free/en/papers/mehri-2020-unsupervised-evaluation-interactive-dialog-dialogpt/) help, but human judgment remains the standard, and human protocols themselves vary.
- **Factuality.** Conversational models invent facts. Work here asks whether [hallucination comes from the data or the models](https://md.research.elamur.ai/free/en/papers/dziri-2022-origin-hallucinations-conversational-models-is/) and how to [check factual consistency](https://md.research.elamur.ai/free/en/papers/honovich-2021-q-2-evaluating-factual-consistency/) against grounding knowledge.
- **Consistency and memory.** Keeping a stable persona and remembering what was said over long conversations remain difficult; [unlikelihood training](https://md.research.elamur.ai/free/en/papers/li-2020-don-t-say-that-making/) and [pragmatic self-consciousness](https://md.research.elamur.ai/free/en/papers/kim-2020-will-i-sound-like-me/) are among the remedies explored.
- **Safety and values.** Dialogue agents can be provoked into offensive or harmful output and can amplify bias. [Prosocial training data](https://md.research.elamur.ai/free/en/papers/kim-2022-prosocialdialog-prosocial-backbone-conversational-agents/), rule-based feedback and [cultural value surveys](https://md.research.elamur.ai/free/en/papers/cao-2024-bridging-cultural-nuances-dialogue-agents/) are partial answers.
- **Annotation quality.** Benchmarks such as MultiWOZ have needed repeated corrections ([2.2](https://md.research.elamur.ai/free/en/papers/zang-2020-multiwoz-2-2-dialogue-dataset/), [2.4](https://md.research.elamur.ai/free/en/papers/ye-2022-multiwoz-2-4-multi-domain/)), which raises questions about how much reported progress reflects label noise.
- **Structure in large models.** Do large models understand dialogue acts, discourse relations and implicature, or only imitate their surface? Recent papers [test LLMs on dialogue acts](https://md.research.elamur.ai/free/en/papers/qamar-2025-do-llms-understand-dialogues-case/) and datasets such as [GRICE](https://md.research.elamur.ai/free/en/papers/zheng-2021-grice-grammar-based-dataset-recovering/) probe conversational implicature.
- **Real-time spoken interaction.** Turn-taking, interruptions and incremental processing are natural in speech but awkward for text models; [incremental dialogue processing](https://md.research.elamur.ai/free/en/papers/schlangen-2011-general-abstract-model-incremental-dialogue/) and speech-native models such as [Moshi](https://md.research.elamur.ai/free/en/papers/defossez-2024-moshi-speech-text-foundation-model/) approach this from opposite ends.
- **Languages beyond English.** Most corpora are English; cross-lingual transfer for task-oriented understanding is still far from monolingual quality.

## Neighbouring topics

Neighbouring topics: papers shared Speech act theory shares the most: 58 of the 255 papers here (23%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/) | 58 shared · 23% |
| [Discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/) | 21 shared · 8% |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 20 shared · 8% |
| [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) | 16 shared · 6% |
| [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) | 14 shared · 5% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 10 shared · 4% |
| [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) | 10 shared · 4% |
| [Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) | 9 shared · 4% |
| [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) | 7 shared · 3% |
| [Question answering](https://md.research.elamur.ai/free/en/topics/question-answering/) | 6 shared · 2% |

Dialogue systems share the most papers with [speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/), through dialogue act annotation, and with [discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/), through the structure of conversation and multi-party discourse parsing. [Language models and pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) supplies the backbone of modern chatbots. [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) meets dialogue in the resolution of anaphora, ellipsis and deixis; [sentiment and emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) in emotion recognition and empathetic response; [vision and language](https://md.research.elamur.ai/free/en/topics/multimodal/) in visually grounded and multimodal dialogue.

Further links run to [evaluation and benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/), [speech and spoken language](https://md.research.elamur.ai/free/en/topics/speech/), [reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/), [question answering](https://md.research.elamur.ai/free/en/topics/question-answering/), [summarization](https://md.research.elamur.ai/free/en/topics/summarization/) of meetings and chats, [bias, fairness and safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/), [multilingual and cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/), [text generation](https://md.research.elamur.ai/free/en/topics/text-generation/), [information retrieval](https://md.research.elamur.ai/free/en/topics/information-retrieval/), [machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/), [semantic parsing](https://md.research.elamur.ai/free/en/topics/semantic-parsing/), [implicature](https://md.research.elamur.ai/free/en/topics/implicature/) and [code and programming languages](https://md.research.elamur.ai/free/en/topics/code-and-programming/).

The shelves this topic stands on Linguistics & languages: 167 of the 255 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) | 167 papers |
| [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) | 76 papers |
| [Other](https://md.research.elamur.ai/free/en/shelves/other/) | 15 papers |
| [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) | 7 papers |
| [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) | 3 papers |
| [Education](https://md.research.elamur.ai/free/en/shelves/education/) | 1 paper |

Most papers on this topic stand on the [Linguistics and languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) shelf, and a large share on [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/); smaller groups belong to [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/), [AI and machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) and [Education](https://md.research.elamur.ai/free/en/shelves/education/).

## Key works in this library

- [Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech](https://md.research.elamur.ai/free/en/papers/prag-2000-stolcke-dialogue-act-modeling-automatic-tagging/) — HMM-based act tagging of Switchboard.
- [The Reliability of a Dialogue Structure Coding Scheme](https://md.research.elamur.ai/free/en/papers/carletta-1997-reliability-dialogue-structure-coding-scheme/) — moves, games and transactions in the Map Task.
- [PARADISE: A Framework for Evaluating Spoken Dialogue Agents](https://md.research.elamur.ai/free/en/papers/walker-1997-paradise-framework-evaluating-spoken-dialogue/) — user satisfaction from task success and costs.
- [The ICSI Meeting Recorder Dialog Act (MRDA) Corpus](https://md.research.elamur.ai/free/en/papers/shriberg-2004-icsi-meeting-recorder-dialog-act/) — act annotation of multi-party meetings.
- [Agenda-Based User Simulation for Bootstrapping a POMDP Dialogue System](https://md.research.elamur.ai/free/en/papers/schatzmann-2007-agenda-based-user-simulation-bootstrapping/) — simulated users for policy learning.
- [ISO 24617-2: A semantically-based standard for dialogue annotation](https://md.research.elamur.ai/free/en/papers/bunt-2012-iso-24617-2-semantically-based/) — the international act standard.
- [Data-Driven Response Generation in Social Media](https://md.research.elamur.ai/free/en/papers/ritter-2011-data-driven-response-generation-social/) — replies as statistical translation.
- [Word-Based Dialog State Tracking with Recurrent Neural Networks](https://md.research.elamur.ai/free/en/papers/henderson-2014-word-based-dialog-state-tracking/) — state tracking straight from recognition output.
- [A Neural Conversational Model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/) — the seq2seq chatbot baseline.
- [Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems](https://md.research.elamur.ai/free/en/papers/wen-2015-semantically-conditioned-lstm-based-natural/) — learned generation for task systems.
- [The Ubuntu Dialogue Corpus: A Large Dataset for Research in Unstructured Multi-Turn Dialogue Systems](https://md.research.elamur.ai/free/en/papers/lowe-2015-ubuntu-dialogue-corpus-large-dataset/) — large-scale response selection.
- [Deep Reinforcement Learning for Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2016-deep-reinforcement-learning-dialogue-generation/) — self-play with conversational rewards.
- [Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders](https://md.research.elamur.ai/free/en/papers/zhao-2017-learning-discourse-level-diversity-neural/) — latent intents against bland replies.
- [DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset](https://md.research.elamur.ai/free/en/papers/li-2017-dailydialog-manually-labelled-multi-turn/) — everyday dialogues with act and emotion labels.
- [Sequicity: Simplifying Task-oriented Dialogue Systems with Single Sequence-to-Sequence Architectures](https://md.research.elamur.ai/free/en/papers/lei-2018-sequicity-simplifying-task-oriented-dialogue/) — belief state written as text.
- [Mem2Seq: Effectively Incorporating Knowledge Bases into End-to-End Task-Oriented Dialog Systems](https://md.research.elamur.ai/free/en/papers/madotto-2018-mem2seq-effectively-incorporating-knowledge-bases/) — memory networks with copying.
- [Personalizing Dialogue Agents: I have a dog, do you have pets too?](https://md.research.elamur.ai/free/en/papers/zhang-2018-personalizing-dialogue-agents-i-have/) — PersonaChat.
- [CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) — QA over many turns.
- [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://md.research.elamur.ai/free/en/papers/zhang-2019-dialogpt-large-scale-generative-pre/) — GPT-2 pretrained on Reddit threads.
- [TOD-BERT: Pre-trained Natural Language Understanding for Task-Oriented Dialogue](https://md.research.elamur.ai/free/en/papers/wu-2020-tod-bert-pre-trained-natural/) — an encoder adapted to task dialogue.
- [Towards a Human-like Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/adiwardana-2020-towards-human-like-open-domain/) — Meena and the SSA metric.
- [Recipes for Building an Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/roller-2021-recipes-building-open-domain-chatbot/) — BlenderBot.
- [Towards Emotional Support Dialog Systems](https://md.research.elamur.ai/free/en/papers/liu-2021-emotional-support-dialog-systems/) — the ESConv corpus and support strategies.
- [LaMDA: Language Models for Dialog Applications](https://md.research.elamur.ai/free/en/papers/thoppilan-2022-lamda-language-models-dialog-applications/) — safety and grounding at scale.
- [Improving alignment of dialogue agents via targeted human judgements](https://md.research.elamur.ai/free/en/papers/glaese-2022-improving-alignment-dialogue-agents-targeted/) — Sparrow and rule-based feedback.

## All 255 papers on this topic

**1** [2](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/2/) … [11](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/11/) [next →](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/2/)

- [Personalizing Dialogue Agents: I have a dog, do you have pets too?](https://md.research.elamur.ai/free/en/papers/zhang-2018-personalizing-dialogue-agents-i-have/) — Saizheng Zhang, Emily Dinan, Jack Urbanek, et al. · 2018 · 1 839 citations
- [A Neural Conversational Model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/) — Oriol Vinyals, Quoc Le · 2015 · 1 759 citations
- [CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) — Siva Reddy, Danqi Chen, Christopher D. Manning · 2019 · 1 541 citations
- [Deep Reinforcement Learning for Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2016-deep-reinforcement-learning-dialogue-generation/) — Jiwei Li, Will Monroe, Alan Ritter, et al. · 2016 · 1 479 citations
- [Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech](https://md.research.elamur.ai/free/en/papers/prag-2000-stolcke-dialogue-act-modeling-automatic-tagging/) — Andreas Stolcke, Klaus Ries, Noah Coccaro, et al. · 2000 · 1 447 citations
- [LaMDA: Language Models for Dialog Applications](https://md.research.elamur.ai/free/en/papers/thoppilan-2022-lamda-language-models-dialog-applications/) — Romal Thoppilan, Daniel De Freitas, Jamie Hall, et al. · 2022 · 1 314 citations
- [Recipes for Building an Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/roller-2021-recipes-building-open-domain-chatbot/) — Stephen Roller, Emily Dinan, Naman Goyal, et al. · 2021 · 1 206 citations
- [DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset](https://md.research.elamur.ai/free/en/papers/li-2017-dailydialog-manually-labelled-multi-turn/) — Yanran Li, Hui Su, Xiaoyu Shen, et al. · 2017 · 1 195 citations
- [Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems](https://md.research.elamur.ai/free/en/papers/wen-2015-semantically-conditioned-lstm-based-natural/) — Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, et al. · 2015 · 1 192 citations
- [Adversarial Learning for Neural Dialogue Generation](https://md.research.elamur.ai/free/en/papers/li-2017-adversarial-learning-neural-dialogue-generation/) — Jiwei Li, Will Monroe, Tianlin Shi, et al. · 2017 · 1 044 citations
- [Learning Discourse-level Diversity for Neural Dialog Models using Conditional Variational Autoencoders](https://md.research.elamur.ai/free/en/papers/zhao-2017-learning-discourse-level-diversity-neural/) — Tiancheng Zhao, Ran Zhao, Maxine Eskenazi · 2017 · 986 citations
- [Data-Driven Response Generation in Social Media](https://md.research.elamur.ai/free/en/papers/ritter-2011-data-driven-response-generation-social/) — Alan Ritter, Colin Cherry, William B. Dolan · 2011 · 692 citations
- [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://md.research.elamur.ai/free/en/papers/zhang-2019-dialogpt-large-scale-generative-pre/) — Yizhe Zhang, Siqi Sun, Michel Galley, et al. · 2019 · 684 citations
- [PARADISE: A Framework for Evaluating Spoken Dialogue Agents](https://md.research.elamur.ai/free/en/papers/walker-1997-paradise-framework-evaluating-spoken-dialogue/) — Marilyn A. Walker, Diane J. Litman, Candace A. Kamm, et al. · 1997 · 657 citations
- [Sequential Matching Network: A New Architecture for Multi-turn Response Selection in Retrieval-Based Chatbots](https://md.research.elamur.ai/free/en/papers/wu-2017-sequential-matching-network-new-architecture/) — Yu Wu, Wei Wu, Chen Xing, et al. · 2017 · 624 citations
- [Towards a Human-like Open-Domain Chatbot](https://md.research.elamur.ai/free/en/papers/adiwardana-2020-towards-human-like-open-domain/) — Daniel Adiwardana, Minh-Thang Luong, David R. So, et al. · 2020 · 573 citations
- [Conversational Memory Network for Emotion Recognition in Dyadic Dialogue Videos](https://md.research.elamur.ai/free/en/papers/hazarika-2018-conversational-memory-network-emotion-recognition/) — Devamanyu Hazarika, Soujanya Poria, Amir Zadeh, et al. · 2018 · 539 citations
- [The Second Conversational Intelligence Challenge (ConvAI2)](https://md.research.elamur.ai/free/en/papers/dinan-2019-second-conversational-intelligence-challenge-convai2/) — Emily Dinan, Varvara Logacheva, Valentin Malykh, et al. · 2019 · 525 citations
- [Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models](https://md.research.elamur.ai/free/en/papers/serban-2015-building-end-end-dialogue-systems/) — Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio, et al. · 2015 · 524 citations
- [ICON: Interactive Conversational Memory Network for Multimodal Emotion Detection](https://md.research.elamur.ai/free/en/papers/hazarika-2018-icon-interactive-conversational-memory-network/) — Devamanyu Hazarika, Soujanya Poria, Rada Mihalcea, et al. · 2018 · 523 citations
- [A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues](https://md.research.elamur.ai/free/en/papers/serban-2016-hierarchical-latent-variable-encoder-decoder/) — Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, et al. · 2016 · 518 citations
- [How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation](https://md.research.elamur.ai/free/en/papers/liu-2016-how-not-evaluate-your-dialogue/) — Chia-Wei Liu, Ryan Lowe, Iulian Serban, et al. · 2016 · 500 citations
- [Towards an Automatic Turing Test: Learning to Evaluate Dialogue Responses](https://md.research.elamur.ai/free/en/papers/lowe-2017-towards-automatic-turing-test-learning/) — Ryan Lowe, Michael Noseworthy, Iulian V. Serban, et al. · 2017 · 486 citations
- [ParlAI: A Dialog Research Software Platform](https://md.research.elamur.ai/free/en/papers/miller-2017-parlai-dialog-research-software-platform/) — Alexander Miller, Will Feng, Dhruv Batra, et al. · 2017 · 483 citations
- [The Reliability of a Dialogue Structure Coding Scheme](https://md.research.elamur.ai/free/en/papers/carletta-1997-reliability-dialogue-structure-coding-scheme/) — Jean Carletta, Amy Isard, Stephen Isard, et al. · 1997 · 480 citations

**1** [2](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/2/) … [11](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/11/) [next →](https://md.research.elamur.ai/free/en/topics/dialogue-systems/page/2/)

- [← previous topic Dependency parsing](https://md.research.elamur.ai/free/en/topics/dependency-parsing/)
- [next topic → Discourse analysis](https://md.research.elamur.ai/free/en/topics/discourse-analysis/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)