Crowd-sourcing has been one of the primary ways to curate conversational data, specially for certain scenarios like grounding in knowledge. In this setting, using online platforms like AMT, non-expert participants are hired to converse with each other, following instructions which try to guide the outcome towards the desired format. The resulting data then is used for different parts of dialog modelling like knowledge selection and response selection/generation. In this work, we take a closer look into two of the most popular knowledge grounded dialog (KGD) datasets. Investigating potential biases and artefacts in knowledge selection labels, we observe that in many cases the ‘knowledge selection flow’ simply follows the order of presented knowledge pieces. In Wizard of Wikipedia (the most popular KGD dataset) we use simple content-agnostic models based on this bias to get significant knowledge selection performance. In Topical-Chat we see a similar correlation between the knowledge selection sequence and the order of entities and their segments, as provided to crowd-source workers. We believe that the observed results, question the significance and origin of the presumed dialog-level attributes like ‘knowledge flow’ in these crowd-sourced datasets.

We expect to interact with home assistants irrespective of our language. However, scaling the Natural Language Understanding pipeline to multiple languages while keeping the same level of accuracy remains a challenge. In this work, we leverage the inherent multilingual aspect of translation models for the task of multilingual intent classification and slot filling. Our experiments reveal that they work equally well with general-purpose multilingual text-to-text models. Furthermore, their accuracy can be further improved by artificially increasing the size of the training set. Unfortunately, increasing the training set also increases the overlap with the test set, leading to overestimating their true capabilities. As a result, we propose two new evaluation methods capable of accounting for an overlap between the training and test set.

Researchers often use games to analyze the abilities of Artificial Intelligence models. In this work, we use the game of Twenty Questions to study the world knowledge of language models. Despite its simplicity for humans, this game requires a broad knowledge of the world to answer yes/no questions. We evaluate several language models on this task and find that only the largest model has enough world knowledge to play it well, although it still has difficulties with the shape and size of objects. We also present a new method to improve the knowledge of smaller models by leveraging external information from the web. Finally, we release our dataset and Twentle, a website to interactively test the knowledge of language models by playing Twenty Questions.

What do language models know about our world? This question is hard to answer but important to get right. To this end, we introduce 20Q, a novel benchmark using the Twenty Questions game to evaluate world knowledge and common sense of language models. Thanks to our overlap-free benchmark, language models learn the game of Twenty Questions without learning relevant knowledge for the test set. We uncover two intuitive factors influencing the world knowledge of language models: the size of the model and the topic frequency in the pre-training data. Moreover, we show that in-context learning is inefficient for evaluating language models’ world knowledge — fine-tuning is necessary to show their true capabilities. Lastly, our results show room for improvement to enhance the world knowledge and common sense of large language models. A potential solution would be to up-sample unfrequent topics in the pre-training of language models.

Generative conversational agents are known to suffer from problems like inconsistency and hallucination, and a big challenge in studying these issues remains evaluation: they are not properly reflected in common text generation metrics like perplexity or BLEU, and alternative implicit methods like semantic similarity or NLI labels can be misguided when few specific tokens are decisive. In this work we propose ConsisTest; a factual consistency benchmark including both WH and Y/N questions based on PersonaChat, along with a hybrid evaluation pipeline which aims to get the best of symbolic and sub-symbolic methods. Using these and focusing on pretrained generative models like BART, we provide detailed statistics and analysis on how the model’s consistency is affected by variations in question and context.

Automatic evaluation of open-domain dialogs remains an unsolved problem. Existing methods do not correlate strongly with human annotations. In this paper, we present a new automated evaluation method based on the use of follow-ups. We measure the probability that a language model will continue the conversation with a fixed set of follow-ups (e.g. not really relevant here, what are you trying to say?). When compared against twelve existing methods, our new evaluation achieves the highest correlation with human evaluations.

FAQs are important resources to find information. However, especially if a FAQ concerns many question-answer pairs, it can be a difficult and time-consuming job to find the answer you are looking for. A FAQ chatbot can ease this process by automatically retrieving the relevant answer to a user’s question. We present VaccinChatNL, a Dutch FAQ corpus on the topic of COVID-19 vaccination. Starting with 50 question-answer pairs we built VaccinChat, a FAQ chatbot, which we used to gather more user questions that were also annotated with the appropriate or new answer classes. This iterative process of gathering user questions, annotating them, and retraining the model with the increased data set led to a corpus that now contains 12,883 user questions divided over 181 answers. We provide the first publicly available Dutch FAQ answering data set of this size with large groups of semantically equivalent human-paraphrased questions. Furthermore, our study shows that before fine-tuning a classifier, continued pre-training of Dutch language models with task- and/or domain-specific data improves classification results. In addition, we show that large groups of semantically similar questions are important for obtaining well-performing intent classification models.

Knowledge Grounded Conversation Models are usually based on a selection/retrieval module and a generation module, trained separately or simultaneously, with or without having access to a ‘gold’ knowledge option. With the introduction of large pre-trained generative models, the selection and generation part have become more and more entangled, shifting the focus towards enhancing knowledge incorporation (from multiple sources) instead of trying to pick the best knowledge option. These approaches however depend on knowledge labels and/or a separate dense retriever for their best performance. In this work we study the unsupervised selection abilities of pre-trained generative models (e.g. BART) and show that by adding a score-and-aggregate module between encoder and decoder, they are capable of learning to pick the proper knowledge through minimising the language modelling loss (i.e. without having access to knowledge labels). Trained as such, our model - K-Mine - shows competitive selection and generation performance against models that benefit from knowledge labels and/or separate dense retriever.

In this paper, we present the first multilingual FAQ dataset publicly available. We collected around 6M FAQ pairs from the web, in 21 different languages. Although this is significantly larger than existing FAQ retrieval datasets, it comes with its own challenges: duplication of content and uneven distribution of topics. We adopt a similar setup as Dense Passage Retrieval (DPR) and test various bi-encoders on this dataset. Our experiments reveal that a multilingual model based on XLM-RoBERTa achieves the best results, except for English. Lower resources languages seem to learn from one another as a multilingual model achieves a higher MRR than language-specific ones. Our qualitative analysis reveals the brittleness of the model on simple word changes. We publicly release our dataset, model, and training script.