Computational models of language have the exciting potential to help writers generate and express their ideas. Current approaches typically provide their outputs to writers in a way that writers can (and often do) appropriate as their owngiving the system more control than necessary over the final outcome of the writing. We present early explorations of two new types of interactions with generative language models; both share the design goal of keeping the writer in ultimate control while providing generative assistance. One interaction enables new kinds of structural manipulation of alreadydrafted sentences; it keeps the writer in semantic control by conditioning the output to be a paraphrase of human-provided input. The other interaction enables new kinds of idea exploration by ofering questions rather than snippets to writers; it keeps the writer in semantic control by providing its ideas in an open-ended form. We present the results of our early experiments on the feasibility and suitability of these types of interactions.
Human writers can use computational tools to be more eficient, creative, and efective. As the capability and accuracy of language models has been improving rapidly in recent years, we are excited by the idea that these technologies might help writers in new ways. Recent developments have used language modeling technology to improve text entry (e.g., improving speed and accuracy) and feedback about grammar and style. We envision technology that will go beyond these capabilities to help writers explore ideas and alternatives—to rapidly hone both what to say and how to say it.
A common approach for using language models to help
writers is to have the model generate text, often with a
seed, constraint, or objective; the writer then uses the
resulting text directly or as inspiration (see, for example, a
study from last year’s HAI-GEN workshop [
In this work, we present prototypes and feasibility studies of these two interactions. We first discuss the task of interactive sentence structure manipulation and present two interaction techniques and one NLP approach to power them. We then describe the task of providing open-ended topic ideas to writers and present evidence from an exploratory study that suggests that writers substantially prefer guidance in the form of questions to guidance in the form of examples.
technology to enable writers to manipulate individual sentences in their writing to shape their meaning and organization. This manipulation could entail changing both the words and their arrangement within the sentence.
Some current interfaces suggest edits that can be accepted or rejected, typically for grammatical error correction [4] or contextual spelling correction. Other in
It was earlier demonstrated by researchers that heart attacks can be caused by high stress. select
It was earlier demonstrated by researchers that heart attacks can be caused by high stress. drag drop earlier demonstrated researchers high stress heart attacks caused
It was earlier demonstrated by researchers that high stress can cause heart attacks. terfaces allow exploration of alternatives to single words could begin with a diferent construction such as prepousing a contextual thesaurus [5]. Still others provide sitional phrase or subordinate clause [11]. In many senindirect controls over the the system’s edits [6, 7, 8]. tences, such constructions could be reordered without
Instead, we draw inspiration from systems like GAN- afecting the general meaning of the sentence. So our dePaint [9] and collaborative summarization [10] that allow sign goal is to enable writers to easily rearrange phrases direct manipulation of creative outputs using high-level in sentences they have written. tools. This kind of interaction can be called Collaborative We have explored several potential interaction techSemantic Inference (CSI) [10]: the interaction provides niques for manipulating phrase order in sentences. Fighuman-understandable hooks into the model’s inference ure 1 shows a direct manipulation interaction: the writer process, enabling the human and the generative model to drags a selected phrase to a new location and the system collaborate in the process of editing an image or writing performs the necessary edits on the rest of the sentence a summary of an article. Our approach conceptually ex- (highlighted in green in the figure). However, in this tends the collaborative summarization system described interaction it is not obvious to the writer which manipin the CSI paper to manipulation types that are helpful ulations are likely to be successful, so we also explored in diferent kinds of writing tasks. For example, their interactions in which the system presents several plauwork described summarization tasks, in which the selec- sible reorderings that the writer can choose among and tion and paragraph-level organization of information is then refine using further selection operations (Figure 2). important; we focus instead on helping the writer craft the organization and expression of information within a 2.2. Approach given sentence. Moreover, in a summarization task much of the content is determined by the source documents, Round-trip machine translation via a pivot language is a but we focus on cases where the writer wants to be the common and efective approach for paraphrase generaultimate author of most of the content. tion (sometimes called sentence rewriting) [12, 13]. The pivot sentence captures much of the original meaning 2.1. Goal: Phrase Reordering of the source-language sentence without constraining the ordering or word choice. Thus, the model can retain We focus our work here on helping writers vary sen- the original meaning as much as possible even when the tence structure in informative writing tasks, specifically words it can generate are manipulated and restricted in on reordering phrases in a sentence while preserving the certain ways. We implemented an English-to-English overall meaning. The Oxford Essential Guide to Writing paraphrase generation model using Spanish as a pivot discusses the importance of both recurrence and variety language (because it was understood by members of our in sentences. Recurrence, which is the repetition of the team and it is related to English) and pretrained transsame sentence structure, can be used to highlight par- lation models.1 Since we can use these models directly allel ideas. However, the overuse of recurrence leads to monotony and a lack of focus. One method of adding variety to a piece of writing is to vary sentence openings.
Instead of always starting a sentence with the subject, it
Helsinki-NLP/opus-mt-en-ROMANCE and https://huggingface. co/Helsinki-NLP/opus-mt-ROMANCE-en. They are seq2seq models with a BART-like architecture, pre-trained using Marian without fine-tuning, our approach is “plug-and-play.” pletion for each prefix. To help the writer understand the
All of the interactions we explored require obtaining a alternatives at a glance, it color-codes each clause; the set of high-quality paraphrases that are diverse in clause order of colors serves as a glanceable summary of how ordering. While in theory these diferent orderings would each alternative sentence has edited the original. all have reasonable probabilities under the conditional It shows a variety of types of alternatives. First, for generative model, unmodified beam search tends to find each prefix, it ofers alternative full completions. Also, a set of very similar outputs, with only minor variations for each generated word, it shows the 10 most likely in words towards the end of the sequence. Generating alternatives ofered by the language model (conditioned alternatives that meaningfully difered in clause order on left context only, because of the limitation of left-towould have required unreasonably large beam sizes. Al- right generation). Clicking on any option causes it to be though methods such as Diverse Beam Search [16] have chosen (regenerating the rest of the sentence if necessary) been developed to address this problem in general, we and the resulting sentence can continue to be edited. focused our implementation efort on approaches that could be directly controlled in interpretable ways. 2.3. Initial Evaluation
We started with a simplified reordering task: varying the opening clause of a sentence. Varying openings Although our system’s implementation is not yet sufican express many meaningful variations of a sentence, ciently refined for formal evaluation with users (e.g., its such as choosing between active and passive voice. We response latency is too high), an informal ofline evaluasolved this problem using a three-step approach. First, tion of its outputs shows promise for its ability to support the system identifies clauses that could be moved to be various semantic editing operations. openers, specifically noun phrases, prepositional phrases, To test our approach, we drew from a characterization and adverbial clauses, using spaCy [17]. Then it applies of paraphrases by Bhagat and Hovy, who classify paralanguage-specific heuristics to edit these phrases to be phrases and quasi-paraphrases into 25 categories and appropriate as sentence prefixes. For example, object provide simple examples of each [18]. We used one senpronouns like “me” are replaced with the corresponding tence from each of the example pairs from that study as subject pronouns (“I”), and the first word is capitalized. input and the other as a target. A test is successful if the Finally, the system uses beam search to generate the most target sentence can be displayed within one or two aclikely completions of each selected starting phrase. tions by a hypothetical user. In most cases, this means it is
A problem we faced was that some sentence comple- one of the 5 alternatives displayed after the user changes tions would duplicate or skip information from the source a word to one of the other top 10 predictions for that sentence. For example, sometimes the phrase selected as word. Of the 25 examples, there were 16 successes, all but a prefix would recur later in the sentence. To avoid such one occurring after just one interaction. The categories results, we postprocessed the beam search results to pe- for these successes are synonym substitution, change nalize discrepancies between the number of occurrences of voice, change of person, pronoun substitution, ellipof each content word in the original sentence and in the sis, function word variations, actor/action substitution, generated output.2 verb/“semantic-role noun” substitution (successful in 1 of The interface (Figure 2) displays the most likely com- 2 tests), manipulator/device substitution, general/specific substitution, part/whole substitution, verb/noun converNMT [14] on parallel corpora collected by the OPUS project [15] sion, noun/adjective conversion, change of tense, and and g2Wraceioduefinsleydschoanrteednwtwithortdhseapsutbolkicenbsynthote iHneslpsianCkyi’NsLstPopgwroourpd. change of aspect. list and not punctuation. A limitation of this evaluation approach is that the most of the paraphrases considered by Bhagat and Hovy consist of lexical substitutions rather than phrase reordering. Since our system’s automatically generated alternatives were designed to emphasize diferences in phrase order, few of the target paraphrases could be made by a single selection. So we also informally tested sentences pulled from other sources. We found that sentences from informative genres (such as Wikipedia articles) often generated acceptable paraphrases, but sentences from narrative genres such as fiction were more likely to sufer from a changed meaning. Nevertheless, results from all sources tended to be grammatical and understandable.
We believe our work falls in line with many of the design
guidelines for co-writing tools suggested by Calderwood
et al. at last year’s HAI-GEN workshop in their study of
novelists employing generative language models within
their writing [
One direction of improving this work is improving the process of searching for high-quality rewrites given ordering constraints. Techniques such as Diverse Beam Search [16] or the Gumbel Top-k trick [19] could help the search explore a wider range of possibilities that are likely to satisfy the ordering constraints. Improved techniques for tracking sequential constraints would enable more principled ways to control the search process [20, 21].
Improved language modeling could improve both the quality and capability of the outcomes. The use of a single pivot sentence sometimes leads to translationese bleeding into the output; this could be reduced by pivoting through several diferent target languages as proposed by Mallison [12]. We are also investigating decoder-only architectures such as GPT-3 and the approach of Guo et al. [22], possibly with an approach like prefix-tuning [ 23] to direct the output sequence. Finally, since the autoregressive setting is limiting for editing operations, we have begun to explore applying language models with flexible ordering such as XLNet.
express an idea already committed to, but by design does
not assist the writer in forming ideas in the first place.
Generative models have often been used to provide ideas
by generating target text [
Interventions that provide goal-oriented guidance to writers have shown benefits to the quality of the final result. For example, fourth- and sixth-grade students produced more efective essays when provided with a list of subgoals appropriate for argumentative writing, such as “You need to explain why those reasons are good reasons for your opinion” [24]. Structure-based planning, in which writers are given high-level goals to organize their outlining, may improve text quality [25]. IntroAssist [26] uses checklists paired with annotated examples, both generated by experts, to scafold writers in an uncommon but high-impact writing task.
Existing interventions are either specific to a certain kind of document or provide only shallow support to a range of documents. Language modeling presents an opportunity to scale these kinds of interventions in two ways: (1) to a wider range of document types and (2) to more targeted guidance within those documents.
In an exploratory study on encyclopedic writing, we compared the approach of giving guidance in the form of questions (“Questions”) with with two alternatives: no guidance, and exemplar sentences (“Snippets”) from high-quality related documents. Results of the study, summarized in Figure 3, suggested that while both types of guidance helped productivity, writers found Questions to be more relevant than Snippets and subjectively preferred them. 3.1.1. Task We designed a scenario for which writers would need to write isolated sentences in encyclopedia-style writing with optional guidance from a “bot.” The hypothetical premise was that Wikipedia needs to rewrite many articles because of licensing issues, so they designed two bots (corresponding to “Questions” and “Snippets” but identiifed to participants by number only) to make suggestions based on existing articles. 3.1.2. Participants of Presentation ( 2 = 48.99, p<.0001) and category relevance ( 2 = 7.35, p=.007), but no interaction between the two ( 2 = 3.75, p=.05). In this analysis, both Participant and Task were treated as random efects.
We recruited 30 participants from MTurk. Each selected a book, a film, and a travel destination of their choice, then wrote 10 sentences about each. For each sentence, participants were given a fixed set of 10 prompts in a ifxed order. Prompt presentation was counterbalanced between the three levels (Questions, Snippets, and a level in which no prompts were given). For each prompt, the writer was first asked whether the prompt gave them an idea about what to write for their article. If they answered Yes, they were then asked to write a sentence.
Participants were instructed not to worry about ordering or flow between sentences, and were instructed to invent plausible information if necessary.
Since participants found prompting questions to be both relevant and useful in drafting tasks, we turn now to investigating approaches to generating questions. The desired system would take a partially written document and generate questions that are topically relevant and not yet answered. These questions could focus on eliciting new types of information (like the prompts of our design study) or on elaboration or clarification of alreadywritten material. The latter task may be well handled by 3.1.3. Stimuli learning to tag phrases with a small set of “wh”-questions. For this exploratory study, we chose 30 prompts non- However, the former task requires a more general apadaptively: For each sentence in a selection of Wikipedia proach to generating questions. Prior approaches and Featured Articles3 in each of the categories (book, film, or datasets for generating questions typically focus on spetravel destination), one of the authors attempted to iden- cific factual questions, often for reading comprehension tify a single clear question that it answered, which was assessment (e.g., [27]), which leads to questions that may typically straightforward for these encyclopedic texts. not generalize to as-yet-unwritten sentences. For exWe then picked the 10 sentences for which the identi- ample, SQuAD v2 [28] includes questions like “What is ifed questions seemed most relevant to similar articles. produced when the features of passive solar architecTable 1 gives an example of a prompt: the Snippets con- ture are customized to the environment?” or “Who is dition presented the original sentences; the Questions Beyoncé’s biggest musical influence?”, which (presumcondition instead showed the question we wrote based ably by design) have only a small number of possible on the snippets. Since the prompts may have been based correct answers—not useful for giving a writer new ideas. on a very diferent genre, the study can measure how Nevertheless, it may be feasible to generalize some of robust the interaction technique is to relevance failures these questions; for example, the second question could for a future adaptive prompt generation technique. be delexicalized to become “Who is the artist’s biggest musical influence?”. 3.1.4. Results Since any particular question may be applicable to a wide range of documents, we explored the feasibility of We found that prompts shown as Questions gave usable a hybrid (ML + crowdsourcing) algorithm to generate a ideas more often than prompts shown as Snippets, and collection of questions and identify which of those questhat writers expressed strong preference for Questions tions are relevant (and not yet answered) in a writing task. over Snippets presentation (Figure 3). Likelihood ratio We clustered sentences within a collection of documents tests in a binomial mixed model predicting number of (e.g., English Wikipedia articles about films), ensuring prompts marked as “relevant” found a significant efect that each cluster occurred in several diferent documents. We then picked several sentences close to the cluster centers and had a person (one of the authors) attempt to write a question that many of those sentences would answer; this process successfully produced a canonical question for many clusters. Therefore, if the system could identify
WikiVoyage. We omitted History sections of Wikivoyage articles and plot summaries in book and film articles since these are much more highly idiosyncratic.
Snippets easiest to understand 25 made it easiest to write 24 would most like to have 25 5 6 5 0 2 4 6 8 10 Number of prompts marked Relevant times picked a likely but not-yet-used cluster in a partially written document, it could ask the corresponding question to the writer. We found that a simplistic classification approach (Naive Bayes on cluster unigrams and bigrams) yielded a promising top-1 accuracy of 25% at predicting the cluster of a not-yet-seen sentence. So while much work remains to be able to reliably generate relevant questions, these results encourage us that is is feasible with current technology.
Alternatively, a language model like GPT-3 may have enough examples of interviews in its dataset (e.g., podcast transcriptions) to be able to be primed to generate sensible questions; we have applied for access to GPT-3 to evaluate this potential.
value that all meaning in the resulting writing should originate with the human author. Our systems eschew generation of novel text that a writer could appropriate directly. Instead, we present novel ideas as questions, not answers—making writing more like a conversation. Refinements are presented as a visual language for manipulating existing text. Generation is constrained to follow the semantics of text that the author provides.
Like many writing assistance technologies, the systems proposed here are dual-use: they can help writers clarify their ideas and make them intelligible to specific audiences, or they can be used to disguise plagiarism or for “article spinning” by content farms. However, existing countermeasures would readily detect paraphrases generated by our systems.
These proposed interactions barely scratch the surface of how high-capacity generative models of language can help writers. For example, our interactions support just a few of the many challenges that writers face when drafting and editing, and none of the challenges that writers face when revising (molding a document to achieve a desired goal) or other tasks such as giving or receiving feedback from others. However, language models can be of great help in these and many more tasks if we continue to think creatively about what we might ask them to generate for us.
Interactive Systems group at Harvard provided valuable formative feedback on part of this work and supported the MTurk experiments. This work was funded in part by a Calvin Research Fellowship and a Jansma Family Research Fellowships in the Sciences. We are grateful to the contributors to the Huggingface Transformers project, especially the Helsinki NLP group, for making easy-to-use APIs for pre-trained models. [4] R. Grundkiewicz, C. Bryant, M. Felice, A crash tion for Computational Linguistics, Valencia, Spain, course in automatic grammatical error correction, 2017, pp. 881–893. URL: https://www.aclweb.org/ in: Proceedings of the 28th International Confer- anthology/E17-1083. ence on Computational Linguistics: Tutorial Ab- [13] C. Federmann, O. Elachqar, C. Quirk, Multilingual stracts, International Committee for Computational whispers: Generating paraphrases with translation, Linguistics, Barcelona, Spain (Online), 2020, pp. in: Proceedings of the 5th Workshop on Noisy User33–38. URL: https://www.aclweb.org/anthology/ generated Text (W-NUT 2019), Association for Com2020.coling-tutorials.6. doi:10.18653/v1/2020. putational Linguistics, Hong Kong, China, 2019, pp. coling-tutorials.6. 17–26. URL: https://www.aclweb.org/anthology/ [5] K. I. Gero, L. B. Chilton, How a stylistic, machine- D19-5503. doi:10.18653/v1/D19-5503. generated thesaurus impacts a writer’s process, [14] M. Junczys-Dowmunt, R. Grundkiewicz, T. Dwoin: Proceedings of the 2019 on Creativity and jak, H. Hoang, K. Heafield, T. Neckermann, F. Seide, Cognition, C&C ’19, Association for Computing U. Germann, A. Fikri Aji, N. Bogoychev, A. F. T. Machinery, New York, NY, USA, 2019, pp. 597– Martins, A. Birch, Marian: Fast neural machine 603. URL: https://doi.org/10.1145/3325480.3326573. translation in C++, in: Proceedings of ACL 2018, doi:10.1145/3325480.3326573. System Demonstrations, Association for Computa[6] R. Louie, A. Coenen, C. Z. Huang, M. Terry, C. J. tional Linguistics, Melbourne, Australia, 2018, pp.
Cai, Novice-AI Music Co-Creation via AI-Steering 116–121. URL: http://www.aclweb.org/anthology/ Tools for Deep Generative Models, Association P18-4020. for Computing Machinery, New York, NY, USA, [15] J. Tiedemann, Parallel data, tools and interfaces in 2020, p. 1–13. URL: https://doi.org/10.1145/3313831. OPUS, in: N. Calzolari, K. Choukri, T. Declerck, 3376739. M. U. Dogan, B. Maegaard, J. Mariani, J. Odijk, [7] A. Fan, D. Grangier, M. Auli, Controllable abstrac- S. Piperidis (Eds.), Proceedings of the Eight Internative summarization, in: Proceedings of the 2nd tional Conference on Language Resources and EvalWorkshop on Neural Machine Translation and Gen- uation (LREC’12), European Language Resources eration, 2018, pp. 45–54. URL: https://arxiv.org/abs/ Association (ELRA), Istanbul, Turkey, 2012. 1711.05217. [16] A. Vijayakumar, M. Cogswell, R. Selvaraju, Q. Sun, [8] K. Gero, C. Kedzie, J. Reeve, L. Chilton, Low level S. Lee, D. Crandall, D. Batra, Diverse beam search linguistic controls for style transfer and content for improved description of complex scenes, in: preservation, in: Proceedings of the 12th Inter- Thirty-Second AAAI Conference on Artificial Intelnational Conference on Natural Language Gener- ligence, 2018. URL: https://www.aaai.org/ocs/index. ation, Association for Computational Linguistics, php/AAAI/AAAI18/paper/view/17329/16334. Tokyo, Japan, 2019, pp. 208–218. URL: https://www. [17] M. Honnibal, I. Montani, S. Van Landeghem, aclweb.org/anthology/W19-8628. doi:10.18653/ A. Boyd, spaCy: Industrial-strength Natural Lanv1/W19-8628. guage Processing in Python, 2020. URL: https: [9] D. Bau, H. Strobelt, W. Peebles, J. Wulf, B. Zhou, //doi.org/10.5281/zenodo.1212303. doi:10.5281/ J. Zhu, A. Torralba, Semantic photo manipulation zenodo.1212303. with a generative image prior, ACM Transactions [18] R. Bhagat, E. Hovy, Squibs: What is a on Graphics (Proceedings of ACM SIGGRAPH) 38 paraphrase?, volume 39, 2013, pp. 463–472. (2019). URL: https://www.aclweb.org/anthology/J13-3001. [10] S. Gehrmann, H. Strobelt, R. Krüger, H. Pfister, A. M. doi:10.1162/COLI_a_00166.
Rush, Visual interaction with deep learning mod- [19] W. Kool, H. van Hoof, M. Welling, Stochastic beams els through collaborative semantic inference, IEEE and where to find them: The gumbel-top-k trick Transactions on Visualization and Computer Graph- for sampling sequences without replacement, 2019. ics 26 (2020) 884–894. doi:10.1109/TVCG.2019. arXiv:1903.06059.
2934595. [20] J. E. Hu, H. Khayrallah, R. Culkin, P. Xia, T. Chen, [11] T. Kane, The Oxford Essential Guide to Writ- M. Post, B. Van Durme, Improved lexically coning, Berkley reference, Berkley Books, 2000, strained decoding for translation and monolingual p. 238. URL: https://books.google.com/books?id= rewriting, in: Proceedings of the 2019 ConfertTgjAQAAIAAJ. ence of the North American Chapter of the As[12] J. Mallinson, R. Sennrich, M. Lapata, Paraphras- sociation for Computational Linguistics: Human ing revisited with neural machine translation, in: Language Technologies, Volume 1 (Long and Short Proceedings of the 15th Conference of the Eu- Papers), Association for Computational Linguisropean Chapter of the Association for Computa- tics, Minneapolis, Minnesota, 2019, pp. 839–850. tional Linguistics: Volume 1, Long Papers, Associa- URL: https://www.aclweb.org/anthology/N19-1090.