Dialogue is notoriously hard to evaluate. Past approaches have used human evaluation.
Dialogue act classification is the task of classifying an utterance with respect to the fuction it serves in a dialogue, i.e. the act the speaker is performing. Dialogue acts are a type of speech acts (for Speech Act Theory, see Austin (1975) and Searle (1969)).
The Switchboard-1 corpus is a telephone speech corpus, consisting of about 2,400 two-sided telephone conversation among 543 speakers with about 70 provided conversation topics. The dataset includes the audio files and the transcription files, as well as information about the speakers and the calls.
The Switchboard Dialogue Act Corpus (SwDA) [download] extends the Switchboard-1 corpus with tags from the SWBD-DAMSL tagset, which is an augmentation to the Discourse Annotation and Markup System of Labeling (DAMSL) tagset. The 220 tags were reduced to 42 tags by clustering in order to improve the language model on the Switchboard corpus. A subset of the Switchboard-1 corpus consisting of 1155 conversations was used. The resulting tags include dialogue acts like statement-non-opinion, acknowledge, statement-opinion, agree/accept, etc.
Annotated example:
Speaker: A, Dialogue Act: Yes-No-Question, Utterance: So do you go to college right now?
| Model | Accuracy | Paper / Source | Code |
|---|---|---|---|
| CRF-ASN (Chen et al., 2018) | 81.3 | Dialogue Act Recognition via CRF-Attentive Structured Network | |
| Bi-LSTM-CRF (Kumar et al., 2017) | 79.2 | Dialogue Act Sequence Labeling using Hierarchical encoder with CRF | Link |
| RNN with 3 utterances in context (Bothe et al., 2018) | 77.34 | A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks |
The MRDA corpus [download] consists of about 75 hours of speech from 75 naturally-occurring meetings among 53 speakers. The tagset used for labeling is a modified version of the SWBD-DAMSL tagset. It is annotated with three types of information: marking of the dialogue act segment boundaries, marking of the dialogue acts and marking of correspondences between dialogue acts.
Annotated example:
Time: 2804-2810, Speaker: c6, Dialogue Act: s^bd, Transcript: i mean these are just discriminative.
Multiple dialogue acts are separated by "^".
| Model | Accuracy | Paper / Source | Code |
|---|---|---|---|
| CRF-ASN (Chen et al., 2018) | 91.7 | Dialogue Act Recognition via CRF-Attentive Structured Network | |
| Bi-LSTM-CRF (Kumar et al., 2017) | 90.9 | Dialogue Act Sequence Labeling using Hierarchical encoder with CRF | Link |
Dialogue state tacking consists of determining at each turn of a dialogue the full representation of what the user wants at that point in the dialogue, which contains a goal constraint, a set of requested slots, and the user's dialogue act.
For goal-oriented dialogue, the dataset of the second Dialogue Systems Technology Challenges (DSTC2) is a common evaluation dataset. The DSTC2 focuses on the restaurant search domain. Models are evaluated based on accuracy on both individual and joint slot tracking.
| Model | Request | Area | Food | Price | Joint | Paper / Source |
|---|---|---|---|---|---|---|
| Zhong et al. (2018) | 97.5 | - | - | - | 74.5 | Global-locally Self-attentive Dialogue State Tracker |
| Liu et al. (2018) | - | 90 | 84 | 92 | 72 | Dialogue Learning with Human Teaching and Feedback in End-to-End Trainable Task-Oriented Dialogue Systems |
| Neural belief tracker (Mrkšić et al., 2017) | 96.5 | 90 | 84 | 94 | 73.4 | Neural Belief Tracker: Data-Driven Dialogue State Tracking |
| RNN (Henderson et al., 2014) | 95.7 | 92 | 86 | 86 | 69 | Robust dialog state tracking using delexicalised recurrent neural networks and unsupervised gate |
The WoZ 2.0 dataset is a newer dialogue state tracking dataset whose evaluation is detached from the noisy output of speech recognition systems. Similar to DSTC2, it covers the restaurant search domain and has identical evaluation.
| Model | Request | Joint | Paper / Source |
|---|---|---|---|
| Zhong et al. (2018) | 97.1 | 88.1 | Global-locally Self-attentive Dialogue State Tracker |
| Neural belief tracker (Mrkšić et al., 2017) | 96.5 | 84.4 | Neural Belief Tracker: Data-Driven Dialogue State Tracking |
| RNN (Henderson et al., 2014) | 87.1 | 70.8 | Robust dialog state tracking using delexicalised recurrent neural networks and unsupervised gate |
The main task retrieval-based chatbot is response selection, that aims to find correct responses from a pre-defined index.
The Ubuntu Corpus contains almost 1 million multi-turn dialogues from the Ubuntu Chat Logs. The task of Ubuntu Corpus is to select the correct response from 10 candidates (others are negatively sampled) by considering previous conversation history. You can find more details at here. The Evaluation metric is recall at position K in N candidates (Recall_N@K).
| Model | R_2@1 | R_10@1 | Paper / Source |
|---|---|---|---|
| DAM (Zhou et al. 2018) | 93.8 | 76.7 | Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network |
| SMN (Wu et al. 2017) | 92.3 | 72.3 | Sequential Matching Network: A New Architecture for Multi-turn Response Selection in Retrieval-Based Chatbots |
| Multi-View (Zhou et al. 2017) | 90.8 | 66.2 | Multi-view Response Selection for Human-Computer Conversation |
| Bi-LSTM (Kadlec et al. 2015) | 89.5 | 63.0 | Improved Deep Learning Baselines for Ubuntu Corpus Dialogs |