Recent Advances in Deep Learning Based Dialogue Systems: A Systematic Survey

In plain words

Chatbots come in two broad kinds. Some are built to finish a job, such as booking a table or answering a bank question. Others just talk, with no goal beyond keeping the conversation going. This survey maps the research on both kinds, focusing on systems built with deep learning.

It covers the neural parts that these systems are assembled from, starting with older recurrent networks and moving to attention and the Transformer. It then explains how a job-focused bot is usually split into understanding, tracking what the user wants, deciding what to do, and writing the reply, and how newer work tries to merge those steps. For chatbots that just talk, it covers the problems that keep showing up, such as bland replies, forgetting who the bot is supposed to be, and missing knowledge. It ends with how these systems are judged and which public datasets exist, and with a short list of directions the authors expect to matter next.

TL;DR

This is a literature survey, not a new method, and it runs no experiments of its own. Its value is the map: two system types (task-oriented and open-domain), a list of neural building blocks with their equations, the open problems for each type, the evaluation methods, and two large dataset tables. Use it to find which family of models and which dataset fits a question. Check the original papers before you implement any single method from it.

Problem

Earlier dialogue surveys are described by the authors as outdated or not comprehensive, and most of them lack a multi-angle analysis §1. A newcomer needs one place that sorts the field by model type and by system type, names the open problems, and lists how systems are evaluated and which data exist §1. The survey aims to fill that gap for deep-learning systems and says it focuses on work from recent years §1.

Method

The survey has no algorithm. Its structure is a taxonomy, and the card records that taxonomy so that a reader can find the right section.

Two angles. Work is organised by model type (the neural building blocks, §2) and by system type (task-oriented, §3, and open-domain, §4). Evaluation (§5) and datasets (§6) follow, and the closing section names trends (§7) §1 fig. 1.

Task-oriented systems (§3). The modular design has four modules: natural language understanding (NLU), dialogue state tracking (DST), dialogue policy learning, and natural language generation (NLG). Together, DST and policy form the dialogue manager. The system also queries an external knowledge base §3. Some recent modular systems skip the NLU module and pass the raw message on, a design that aims to reduce error propagation between modules §3. The survey then covers end-to-end designs in two forms: a pipeline whose modules are made differentiable, and a single model that does knowledge lookup and generation together §3.5.

# Modular task-oriented turn  [§3] [fig. 13]
# Input:  user_message (str), history (list), kb (knowledge base)
# Output: response (str)

def respond(user_message, history, kb):
    # 1. Turn the message into domain, intent and slots  [§3.1] [tab. 2]
    nlu_output = understand(user_message)
    # 2. Update the belief state from the whole history  [§3.2]
    state = track(history, nlu_output)
    # 3. Query the knowledge base with the state  [§3]
    results = kb.query(state)
    # 4. Choose the next dialogue action  [§3.3]
    action = policy(state, results)
    # 5. Turn the action into surface text  [§3.4]
    return generate(action)

Open-domain systems (§4). These are split into generative, retrieval-based and ensemble systems §4 intro. The survey organises their open problems as ten topics: context awareness, response coherence, response diversity, speaker consistency and persona, empathy, controllable generation, conversation topic, knowledge grounding, interactive training, and visual dialogue §4.1–§4.10.

Neural building blocks (§2). Each block gets a short description and its key equations. The equations are copied from the original papers and are not derived here.

• Convolutional networks, used as feature extractors on top of an encoder rather than as the encoder itself §2.1 eq. 1 fig. 2.
• Recurrent networks: Jordan- and Elman-type cells §2.2.1 eq. 2–eq. 5 fig. 3; LSTM §2.2.2 eq. 6–eq. 11; GRU §2.2.3 eq. 12–eq. 15; bidirectional RNNs §2.2.4 eq. 16–eq. 18; and the encoder-decoder sequence-to-sequence model §2.2.5 eq. 19–eq. 20.
• Hierarchical recurrent encoder-decoder (HRED), with a token-level and a turn-level recurrent network; the variational version (VHRED) adds a latent variable §2.3 fig. 4.
• Memory networks, with a memory of facts read by soft selection §2.4 eq. 21–eq. 23 fig. 5.
• Attention and the Transformer: additive attention §2.5.1 eq. 24–eq. 29 fig. 6; the Transformer encoder-decoder with scaled dot-product self-attention, multi-head attention and sinusoidal positional encoding §2.5.2 eq. 30–eq. 33 fig. 7.
• Pointer network and CopyNet, which copy tokens from the input §2.6 eq. 34–eq. 39 fig. 8 fig. 9.
• Deep reinforcement learning: Deep Q-networks §2.7.1 eq. 40–eq. 43, REINFORCE and actor-critic §2.7.2 eq. 44–eq. 46; and GANs §2.7.3 eq. 47–eq. 48 fig. 10 fig. 11.
• Knowledge-graph embedding and knowledge-augmented networks §2.8 fig. 12.

Equation numbers were counted in source order over all numbered displays in the LaTeX source. The source labels only a few of them (for example the convolution and the first recurrent equations), so the other numbers follow from that count rather than from printed labels.

Evaluation (§5). For task-oriented systems: task completion rate, task completion cost measured in turns, human evaluation via crowdsourcing or live deployment, and user simulators, either agenda-based or model-based §5.1. For open-domain systems: word-overlap metrics (BLEU, ROUGE, METEOR), neural metrics (for example RUBER, and discriminator-based scores), custom metrics for diversity, consistency and naturalness, and ranking-based human evaluation §5.2. The survey notes that word-overlap scores correlate poorly with human judgement in dialogue §5.2.

Datasets (§6). Table 3 lists task-oriented datasets, such as MultiWOZ, SGD, Schema, ATIS, SNIPS-NLU, E2E, Frames and Taskmaster tab. 3. Table 4 lists open-domain datasets, such as PERSONA-CHAT, Wizard of Wikipedia, CMU DoG, DailyDialog, EMPATHETIC DIALOGUES, Topical-Chat, OpenDialKG, VisDial and Image-Chat tab. 4. Each table gives a short description, the task, and the origin, which is a paper or a URL. Table 3 and Table 4 together are the dataset index of the survey §6.

Trends (§7). The authors name multimodal dialogue, blending task-oriented and open-domain systems into one model, exploring online dialogue corpora in real time, user modelling and simulation, and dialogue with a long-term goal §7.

Hyperparameters

The survey trains nothing, so it gives no training hyperparameters. The table records the numbers it states, so that an implementer can see which ones are architecture settings and which are background.

Results

Not applicable. The survey runs no experiments and reports no results of its own. Its numbers are the dataset counts in tab. 3 and tab. 4 and the market figures in §1, which the authors take from web sources. The survey does not rank systems on a common benchmark.

Paper vs code

No official code. The paper has no code or project-page link. Its only links are footnotes to market sources and to dataset pages in tab. 3 and tab. 4. I did not run a GitHub search, because a survey proposes no method to implement and the search would only return awesome-list repositories, which are not code for this paper.

Pitfalls

• Not an implementation spec. The equations and pseudocode above restate the original works in notation that differs between them. Before you build anything, read the cited original paper, not this survey.
• Memory-network similarity. §2.5.1 says the original memory network used cosine distance as its similarity model. The equation in §2.4 that the survey prints, eq. 21, uses a softmax over a dot product, u transposed times m_i. The text and the equation disagree; follow the equation and check the original.
• GAN notation. eq. 47 and eq. 48 use ω for the discriminator input in the text but w in one sentence. The text calls the pair of losses a bilevel loss without saying which network minimises which loss. Check the GAN paper before you pick signs.
• LSTM naming. §2.2.2 calls the candidate ĥ the short-term memory and s the long-term memory. The equations do not enforce that split, so the naming should not be taken as a design rule.
• BERT and GPT-2 block counts. §2.5.2 says both models consist of 12 transformer blocks. The survey does not name the model size behind that number, and it gives no other sizes.
• Unverified statistics. The claim that roughly 70 percent of articles name open-domain systems "chit-chat" §1, the market figures §1, and the average-turn counts in tab. 3 come from the authors' own survey or from web sources without a method or a link to a stable record. Treat them as the authors' claims.
• Dataset availability and terms. The survey does not report licences for the datasets in tab. 3 and tab. 4. It says that social-media APIs come with legal terms that restrict reuse §6, and at least one row gives no public source: the SubTle corpus row points to the authors for data tab. 4. Check each dataset's own licence and availability before you use it.
• Outdated and changing material. The authors note they will update the survey as new approaches appear [abstract footnote]. The source is version 5 (updated 2022-03-30). Newer work, including large pretrained language models for dialogue, is not covered by the text.
• Coverage is by the authors' own search. The survey calls itself the most comprehensive and up-to-date work of its kind, and in a footnote says it may still omit works [abstract footnote]. A method missing from it is not evidence that the method does not exist.

Lineage

• Builds on: none. It is a survey and reviews the prior surveys it names, which the authors describe as outdated or not comprehensive §1.
• Compared against: the earlier dialogue surveys cited in §1, and the Transformer-variant surveys cited in §2.5.2.
• Superseded by: not stated.