Keywords

1 Motivation

With an increasing output quality of text-to-text NLG models, the attention of the field is turning towards the ultimate goal, to enable human-like natural language interactions. Even outside of the dialog-system area, the generated language is produced to fulfill specific communication goals [113], hence should be tailored to the specific audience [43, 100]. Human speakers naturally use a conceptual model of the recipient in order to achieve their communication goal more efficiently, for example adjust the style or level of complexity [60, 101, 126, 150]. It is therefore reasonable to assume that such user models improve the quality of NLG systems through better adaptivity and robustness [36, 82], and to personalize the system outcomes based on the available relevant information about the user. Research in this area is driven by insights from numerous disciplines, from psychology across linguistics to human-computer interaction, while the industry focus on customer-driven solutions powers the personalization of conversational assistants[8, 13, 14, 22]. As a result, research contributions are scattered across diverse venues. Our aim is to help to limit duplicate research activities, and to organize user-centric efforts within the NLG community. The possibilities of personalizing generated text towards the user range across multiple dimensions and numerous levels of depth, from factual knowledge over preferences and opinions to stylistic discourse adjustments. We use for all these user adjustment variations an umbrella term user-centric natural language generation. We provide a comprehensive overview of recent approaches and propose a categorization of ongoing research directions.

2 Related Surveys

Related to our work, [118] conduct a survey of datasets for dialogue systems, yet noting that “personalization of dialogue systems as an important task, which so far has not received much attention”. [32] surveys user profiling datasets, however, without an NLG focus. Given various input types in NLG (e.g., tables [99], RDF triple [44], meaning representation [31]), we narrow our focus to user-centric text-to-text generation when referring to user-centric NLG in this work.

3 User-Centric NLG

Generally, NLGFootnote 1 is a process that produces textual content (a sequence of consecutive words) based on a chosen structured or unstructured input. In the ideal case, such textual content shall be syntactically and semantically plausible, resembling human-written text [45, 46]. NLG encompasses a wide range of application tasks [43], such as neural machine translation, text summarization, text simplification, paraphrasing with style transfer, human-machine dialog systems, video captioning, narrative generation, or creative writing [43].

Fig. 1.
figure 1

User-centric natural language generation

Full size image

3.1 When Is NLG User-Centric?

Given a text generation problem transforming an input x to an output y, we refer to it as user-centric natural language generation system when the output y of the NLG model is conditioned by information \(I_u\) available about the user u. In other words, the information \(I_u\) is leveraged to alter the projection of an input x to the output space. As illustrated in Fig. 1, this available user information can be of various kinds depending on specific application domains, which we categorize as follows.

3.2 User and Application Domain

In this paper we interpret the user in the term “user-centric” as the recipient of the generated text. Note that the previous work on personalized NLG, which we review here, sometimes takes an author-centric rather than recipient-centric view, for example dialog system works often refer to personalization as modeling of the chatbot persona [63]. The specific role of a user is dependent on a particular application domain, which also typically characterizes the type of the input. Below are some common application domains and the user and input examples.

  • Conversational agents [69]: User is the human participant of the conversation. Input are typically the preceding utterances.

  • Personalized machine translation [92, 94, 107, 130]: User is the requester of the translation. Input is the text to be translated. Prior work mainly studied how particular personal trait of the author such as gender gets manifested in the original texts and in translations. [94] introduced a personalized machine translation system where users’ preferred translation is predicted first based on similar other users.

  • Writing Assistants: User is the final editor of the generated text, typically also the author of the input. Most current automated response generations such as Smart Reply in emails [65] are conducted in a generic way, not an user-specific manner.

  • Personalized text simplification [9, 72, 89]: User is the reader, input is the text to be simplified.

Depending on whether a user is the recipient or the actor of a text, user-centric systems adapt themselves accordingly in terms of how to incorporate personalized or user-specific information into the modeling process.

Diverse Understanding of Personalization. As shown in Table 1, the interpretation of what personalized NLG means varies largely. Many systems optimize for speaker persona consistency or traits, while others operate with recipient’s preferences. However, only a few studies considered recipients in their models [28]. We therefore argue, that in order to “solve” user-centric NLG, we must state more explicitly who our users are, what user needs we assume from them, and more importantly, how these user needs are reflected in our system design.

3.3 User Information

As shown in Fig. 1, we categorize user information into the following categories: (1) factual knowledge, which includes (1a) user trivia and (1b) preferences, and (2) stylistic modifiers, which encompass (2a) situational, (2b) personal, and (2c) emotional choices.

Table 1. Overview table of example previous user-centric NLG works that fall into each user information type. Note that the majority of works focuses on modeling the speaker persona rather than personalizing towards a representation of a recipient.
Full size table

  • (1) Factual Knowledge. Incorporating factual knowledge specific to a given user is essential in increasing user engagement. Information concerning user trivia (1a) can include personal data such as user’s name or location, or user attributes such as occupation. For instance, [141] include user facts such as “i have four kids”, although factual knowledge is not introduced in a structured way. [19] utilized product user categories to generate personalized product descriptions. [91] uses reinforcement learning to reward chatbot persona consistency using fact lists with information like “my dad is a priest”. User facts for personalization include also expertise level in tutoring systems [53, 101]. In addition to user trivia, including user preferences (1b) (“i hate Mexican food” or “i like to ski” [141]), such as opinions, values, and beliefs [114] has been of importance for dialog systems, as it leads to producing more coherent and interesting conversations.

  • (2) Stylistic Modifiers. Stylistic variation can be characterized as a variation in phonological, lexical, syntactic, or discourse realisation of a particular semantic content, due to user’s characteristics [15, 45]. To date, most of the style adaptation work in the NLG area focused on the situational stylistic modifiers (2a), perceiving language use as a function of intended audience, genre, or situation. For example, professional/colloquial [35], personal/impersonal, formal/informal [18, 96, 102, 103, 110, 136, 143] or polite/impolite [25, 38, 85, 95, 116]. Recently, unsupervised style transfer has gained popularity [70].

Comparably less research has been conducted in the emotional and personal modifiers, such as empathy or demographics. Personal stylistic modifiers (2b) in our scheme include user attributes, i.e. both conscious and unconscious traits intrinsic to the text author’s individual identity [59]. A common property of these traits is that while their description is typically clear, such as teenager, Scottish, or extrovert, their surface realization is less well-defined [7]. Note that this is different from employing these attributes as user trivia in a factual way. The two main subgroups of personal modifiers are sociodemographic traits and personality. NLG words explore mostly gendered paraphrasing and gender-conditioned generation [104, 105, 112, 127]. Personality has been employed mostly in the dialog area, mainly on the agent side [50, 95, 97]. In an early work on personality-driven NLG, the system of [87] estimates generation parameters for stylistic features based on the input of big five personality traits [24]. For example, an introverted style may include more hedging and be less verbose. While the big five model is the most widely accepted in terms of validity, its assessments are challenging to obtain [123]. Some works thus resort to other personality labels [41, 132], or combinations of sociodemographic traits and personal interests [146]. Modeling personality of the recipient of the generated text is rare in recent NLG systems, although it has been shown to affect e.g. argument persuasiveness [33, 83] and capability of learning from a dialog [26]. For example [53] proposed to use a multi-dimensional user model including hearer’s emotional state and interest in the discussion, [26] represented users’ stylistic preference for verboseness and their discourse understanding ability, and [11] inferred user’s psychological states from their actions to update the model of a user’s beliefs and goals. [55] uses LIWC keywords to infer both instructor’s and recipient’s personality traits to study dialog adaptation.

Emotional stylistic modifiers (2c) encompass the broad range of research in the area of affective NLG [26]. In the early works, manually prepared rules are applied to deliberately select the desired emotional responses [124], and pattern-based models are used to generate text to express emotions [66]. There is a broad range of features beyond affective adjectives that can have emotional impact, such as an increased use of redundancy, first-person pronouns, and adverbs [27]. [47] introduce neural language models which allows to customize the degree of emotional content in generated sentences through an additional design parameter (happy, angry, sad, anxious, neutral). They note that it is difficult to produce emotions in a natural and coherent way due to the required balance of grammatically and emotional expressiveness. [4] show three novel ways to incorporate emotional aspects into encoder-decoder neural conversation models: word embeddings augmented with affective dictionaries, affect-based loss functions, and affectively diverse beam search for decoding. In their work on emotional chatting machines, [148] demonstrates that simply embedding emotion information in existing neural models cannot produce desirable emotional responses but just ambiguous general expressions with common words. They proposes a mechanism, which, after embedding emotion categories, captures the change of implicit internal emotion states, and boosts the probability of explicit emotion expressions with an external vocabulary. [125] observe, in line with [27], that one doesn’t need to explicitly use strong emotional words to express emotional states, but one can implicitly combine neutral words in distinct ways to increase the intensity of the emotional experiences. They develop two NLG models for emotionally-charged text, explicit and implicit. The ability to produce language controlled for emotions is closely tied to the goal of building empathetic social chatbots [28, 40, 41, 111, 121]. To date, these mainly leverage emotional embeddings similar to those described above to generate responses expected by the learned dialog policies. [78] point out the responses themselves don’t need to be emotional, but mainly understanding, and propose a model based on empathetic listeners.

Implicit User Modeling. With the rise of deep learning models and the accompanying learned latent representations, boundaries between the user information categories sometimes get blurred, as the knowledge extracted about the user often isn’t explicitly interpreted. This line of work uses high-dimensional vectors to refer to different aspects associated with users, implicitly grouping users with similar features (whether factual or stylistic) into similar areas of the vector space. Neural user embeddings in the context of dialog modeling have been introduced by [74], which capture latent speaker persona vectors based on speaker ID. This approach has been further probed and enhances by many others [63], e.g. by pretraining speaker embeddings on larger datasets [69, 137, 146, 147], incorporating user traits into the decoding stage [145], or via mutual attention [88].

4 Data for User-Centric NLG

We identify five main types of datasets that can be leveraged for user-centric NLG, and provide their overview in Table 2. These types include: (1) Attribute-annotated datasets for user profiling, such as in [109], (2) style transfer and attribute transfer paraphrasing datasets such as [110], (3) attribute-annotated machine translation datasets such as [130], (4) persona-annotated dialog datasets such as [141], and (5) large conversational or other human-generated datasets with speaker ID, which allow for unsupervised speaker representation training.

In addition, as [12] point out, the challenge in the big data era is not to find human generated dialogues, but to employ them appropriately for social dialogue generation. Any existing social media dialogues can be combined with a suite of tools for sentiment analysis, topic identification, summarization, paraphrase, and rephrasing, to bootstrap a socially-apt NLG system.

Table 2. Available datasets usable for user-centric NLG
Full size table

5 User-Centric Generation Models

Already [150] discuss how natural language systems consult user models in order to improve their understanding of users’ requirement and to generate appropriate and relevant responses. Generally, current user-centric generation models can be divided into rule-based, ranking-based and generation-based models.

Rule-based user models often utilize a pre-defined mapping between user types and topics [34], or hand-crafted user and context features [1]. The recent Alexa Prize social-bots also utilized a pre-defined mapping between personality types and topics [34], or hand-crafted user and context features [1].

Ranking-based models [2, 90, 141] focus on the task of response selection from a pool of candidates. Such response selection relies heavily on learning the matching between the given user post and any response from the pool, such as the deep structured similarity models [56] or the deep attention matching network [149]. [80] proposed to address the personalized response ranking task by incorporating user profiles into the conversation model. Generation-based models attempt to generate response directly from any given input questions. Most widely used models are built upon sequence-to-sequence models, and the recent transformer-based language models pretrained with large corpora [144].

With the development of large scale social media data [69, 117, 119, 128, 145], several personalized response generation models have been proposed. [21] introduced a neural model to learn a dynamically updated speaker embedding in a conversational context. They initialized speaker embedding in an unsupervised way by using context-sensitive language generation as an objective, and fine-tuned it at each turn in a dialog to capture changes over time and improve the speaker representation with added data. [74] introduced the Speaker Model that encoded user-id information into an additional vector and fed it into the decoder to capture the identity of the speakers. In addition to using user id to capture personal information, [141] proposed a profile memory network for encoding persona sentences. Recently, there are a few works using meta-learning and reinforcement learning to enhance mutual persona perception [68, 79, 88]. Generative models can produce novel responses, but they might suffer from grammar errors, repetitive, hallucination, and even uncontrollable outputs, all of which might degrade the performance of user-centric generation. For instance, under personalized dialog settings, [141] claimed that ranking-based models performed better than generative models, suggesting that building user-centric generative models is more challenging.

Hybrid models attempt to combine the strengths of the generative and rank paradigms [138] in a two-stage fashion, i.e., retrieving similar or template responses first, and then using these to help generate new responses. Hybrid models shed light on how to build user-centric NLG models as the first stage can be used to retrieve relevant knowledge/responses and the second stage can fine-tune the retrieved ones to be user-specific.

6 Evaluations

Current automatic evaluation metrics for response generation can be broadly categorized into three classes: content-driven, quality-based and user-centric. Content relatedness measures capture the distance of the generated response from its corresponding ground-truth, with representative metrics such as BLEU [98], NIST [30], and METEOR [71]. Speaker sensitive responses evaluation model [5] enhances the relatedness score with a context-response classifier. From a quality perspective, the fluency and diversity matter, assessed via perplexity [20] and distinct diversity [73]. From a user-centric perspective, we need to evaluate the style matching or fact adherence that compare the generated responses’ language to the user’s own language. Existing example metrics include the stylistic alignment [93, 129] at the surface, lexical and syntactic level, model-driven metrics such as Hits@1/N, calculating how accurate the generated response can be automatically classified to its corresponding user or user group [29, 93], and the average negative log-likelihood of generated text to user-specific language model, e.g. for poet’s lyrics [131].

Evaluation towards open-ended conversations [64, 106] also use Grice’s Maxims of Conversation [49], i.e., evaluating whether the generated content violates Quantity that gives more or less information than requires, Quality that shares false information or things we do not have evidence, Relation that stays on the relevant topic, and Manner that requires communicating clearly without much disfluency. [67] further introduced a new diagnostic measure called relative utterance quantity (RUQ) to see if the model favors a generic response (e.g., ‘I don’t know’). over the reference it was trained on.

Despite various measures in automatically assessing the quality of responses generated, human evaluation still plays a key role in assessing user-centric NLG systems, as the correlation between automated and human quality judgments is very limited [81, 93]. Automatic metrics for evaluating user-centric NLG systems could then come in the form of an evaluation model learned from human data, e.g. collected from surveys, in order to provide human-like scores to proposed responses like BLEURT [115]. Recently, [54] argued that although human assessment remains the most trusted form of evaluation, the NLG community takes highly diverse approaches and different quality criteria, making it difficult to compare results and draw conclusions, with adverse implications for meta-evaluation and reproducibility. Their analyses on top of 165 NLG papers call for standard methods and terminology for NLG evaluation.

Human judgement for user-centric NLG requires significant efforts. User information such as styles, opinions or personalized knowledge is often scattered throughout the entire participation history in various formats such as posts, comments, likes or log-ins. It is impossible for annotators to go through these hundreds of activity records to infer whether the generated response fits the user well; furthermore, personalization is hardly reflected in a single message, but mostly inferred from a large collection of users’ activities. [123]. Moreover, users’ preferences and interests change over time either slowly or rapidly [48, 77], making it even harder to third-parties to judge and evaluate. As a result, direct and self-evaluation from users of the user-centric NLG systems deserves more attention.

7 Challenges and Opportunities

User-Centric Data Collection and Evaluation. Collecting large-scale personalized conversations or data for NLG systems is challenging, expensive and cumbersome. First, most datasets suffer from pre-defined or collected user profiles expressed in a limited number of statements. Second, crowdsourcing personalized datasets is likely to result in very artificial content, as the workers need to intentionally inject the received personalization instructions into every utterance, which does not align well with human daily conversations. Correspondingly, state-of-the-art models tend to perform the attribute transfer merely at the lexical level (e.g. inserting negative words for rude or “please” for polite), while the subtle understanding and modification of higher-level compositionality is still a challenge [39, 62, 148]. Even more problematic assumption of most user-centric generation systems is that users exhibit their traits, moods and beliefs uniformly in a conversation. However, humans do not always express personalized information everywhere, thus real world data is persona-sparse [147]. This calls for a nuanced modeling of when, where and to what extent personalization needs to be included for NLG systems [17, 37, 122].

Personalized Pretraining and Safeguards. Getting data is a key challenge when it comes to personalized pre-training [147], which requires extensive data even for each single user. The proliferation of personalization also brings in trust and privacy issues [23, 120]. How does user-centric generation relate to ethics and privacy as the personalization always involve using user specific data [51]? One key issue associated with personalized pretraining is that the extensive personal data needed by pretrained language models might include all sorts of dimensions about users, including sensitive and private information which should not be used by user-centric NLG systems [52, 108]. For instance, [16] demonstrated that an adversary can perform an extraction attack to recover individual training examples from pretrained language models. These extracted examples include names, phone numbers, email addresses, and even deleted content. Such privacy concerns might become more salient and severe when it comes to user-centric pretraining, as models can easily remember details and leak training data for potential malicious attacks.

Biases and Generalization. The creation of corpora for user-centric NLG might suffer from self-selection bias as people who decides to use certain platforms like Twitter or Reddit might be very different. The reporting bias further adds complexity to this space as people do not necessarily talk about things in the world in proportion to their persona or personality. Thus, NLG systems built upon available data might be skewed towards certain population (e.g., educational background, access to Internet, specific language uses). The crowdsourcing bias [57], i.e., the potential inherent bias of crowd workers who contribute to the tasks might introduce biased ground-truth data.

Gaps Between Users and Systems. We argue that the evaluation process should look into what dimension users expect to see and identify what users want from these generated texts. For example [135] points out the expectations from human and artificial participants of the conversation are not the same, and shall be modeled differently. We need metrics to capture any failures, and mechanisms to explain the decision-making process behind these user-centric NLG models, since the data-driven systems tend to imitate utterances from their training data [61, 133, 139]. This process is not directly controllable, which may lead to offensive responses [58]. Another challenge is how to disentangle personalization from the generic representation [39], such as using domain adaptation techniques to transfer generic models to specific user groups [142].

8 Conclusion

This work presents a comprehensive overview of recent user-centric text generation across diverse sets of research capturing multiple dimensions of personalizing systems. We categorize these previous research directions, and present the representative tasks and evaluation approaches, as well as challenges and opportunities to facilitate future work on user-centric NLG.