Tag Archives: Natural Language Processing

Unifying image-caption and image-classification datasets with prefix conditioning

Posted by Kuniaki Saito, Student Researcher, Cloud AI Team, and Kihyuk Sohn, Research Scientist, Perception Team

Pre-training visual language (VL) models on web-scale image-caption datasets has recently emerged as a powerful alternative to traditional pre-training on image classification data. Image-caption datasets are considered to be more “open-domain” because they contain broader scene types and vocabulary words, which result in models with strong performance in few- and zero-shot recognition tasks. However, images with fine-grained class descriptions can be rare, and the class distribution can be imbalanced since image-caption datasets do not go through manual curation. By contrast, large-scale classification datasets, such as ImageNet, are often curated and can thus provide fine-grained categories with a balanced label distribution. While it may sound promising, directly combining caption and classification datasets for pre-training is often unsuccessful as it can result in biased representations that do not generalize well to various downstream tasks.

In “Prefix Conditioning Unifies Language and Label Supervision”, presented at CVPR 2023, we demonstrate a pre-training strategy that uses both classification and caption datasets to provide complementary benefits. First, we show that naïvely unifying the datasets results in sub-optimal performance on downstream zero-shot recognition tasks as the model is affected by dataset bias: the coverage of image domains and vocabulary words is different in each dataset. We address this problem during training through prefix conditioning, a novel simple and effective method that uses prefix tokens to disentangle dataset biases from visual concepts. This approach allows the language encoder to learn from both datasets while also tailoring feature extraction to each dataset. Prefix conditioning is a generic method that can be easily integrated into existing VL pre-training objectives, such as Contrastive Language-Image Pre-training (CLIP) or Unified Contrastive Learning (UniCL).

High-level idea

We note that classification datasets tend to be biased in at least two ways: (1) the images mostly contain single objects from restricted domains, and (2) the vocabulary is limited and lacks the linguistic flexibility required for zero-shot learning. For example, the class embedding of “a photo of a dog” optimized for ImageNet usually results in a photo of one dog in the center of the image pulled from the ImageNet dataset, which does not generalize well to other datasets containing images of multiple dogs in different spatial locations or a dog with other subjects.

By contrast, caption datasets contain a wider variety of scene types and vocabularies. As shown below, if a model simply learns from two datasets, the language embedding can entangle the bias from the image classification and caption dataset, which can decrease the generalization in zero-shot classification. If we can disentangle the bias from two datasets, we can use language embeddings that are tailored for the caption dataset to improve generalization.

Top: Language embedding entangling the bias from image classification and caption dataset. Bottom: Language embeddings disentangles the bias from two datasets.

Prefix conditioning

Prefix conditioning is partially inspired by prompt tuning, which prepends learnable tokens to the input token sequences to instruct a pre-trained model backbone to learn task-specific knowledge that can be used to solve downstream tasks. The prefix conditioning approach differs from prompt tuning in two ways: (1) it is designed to unify image-caption and classification datasets by disentangling the dataset bias, and (2) it is applied to VL pre-training while the standard prompt tuning is used to fine-tune models. Prefix conditioning is an explicit way to specifically steer the behavior of model backbones based on the type of datasets provided by users. This is especially helpful in production when the number of different types of datasets is known ahead of time.

During training, prefix conditioning learns a text token (prefix token) for each dataset type, which absorbs the bias of the dataset and allows the remaining text tokens to focus on learning visual concepts. Specifically, it prepends prefix tokens for each dataset type to the input tokens that inform the language and visual encoder of the input data type (e.g., classification vs. caption). Prefix tokens are trained to learn the dataset-type-specific bias, which enables us to disentangle that bias in language representations and utilize the embedding learned on the image-caption dataset during test time, even without an input caption.

We utilize prefix conditioning for CLIP using a language and visual encoder. During test time, we employ the prefix used for the image-caption dataset since the dataset is supposed to cover broader scene types and vocabulary words, leading to better performance in zero-shot recognition.

Illustration of the Prefix Conditioning.

Experimental results

We apply prefix conditioning to two types of contrastive loss, CLIP and UniCL, and evaluate their performance on zero-shot recognition tasks compared to models trained with ImageNet21K (IN21K) and Conceptual 12M (CC12M). CLIP and UniCL models trained with two datasets using prefix conditioning show large improvements in zero-shot classification accuracy.

Zero-shot classification accuracy of models trained with only IN21K or CC12M compared to CLIP and UniCL models trained with both two datasets using prefix conditioning (“Ours”).

Study on test-time prefix

The table below describes the performance change by the prefix used during test time. We demonstrate that by using the same prefix used for the classification dataset (“Prompt”), the performance on the classification dataset (IN-1K) improves. When using the same prefix used for the image-caption dataset (“Caption”), the performance on other datasets (Zero-shot AVG) improves. This analysis illustrates that if the prefix is tailored for the image-caption dataset, it achieves better generalization of scene types and vocabulary words.

Analysis of the prefix used for test-time.

Study on robustness to image distribution shift

We study the shift in image distribution using ImageNet variants. We see that the “Caption” prefix performs better than “Prompt” in ImageNet-R (IN-R) and ImageNet-Sketch (IN-S), but underperforms in ImageNet-V2 (IN-V2). This indicates that the “Caption” prefix achieves generalization on domains far from the classification dataset. Therefore, the optimal prefix probably differs by how far the test domain is from the classification dataset.

Analysis on the robustness to image-level distribution shift. IN: ImageNet, IN-V2: ImageNet-V2, IN-R: Art, Cartoon style ImageNet, IN-S: ImageNet Sketch.

Conclusion and future work

We introduce prefix conditioning, a technique for unifying image caption and classification datasets for better zero-shot classification. We show that this approach leads to better zero-shot classification accuracy and that the prefix can control the bias in the language embedding. One limitation is that the prefix learned on the caption dataset is not necessarily optimal for the zero-shot classification. Identifying the optimal prefix for each test dataset is an interesting direction for future work.

Acknowledgements

This research was conducted by Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Thanks to Zizhao Zhang and Sergey Ioffe for their valuable feedback.

Source: Google AI Blog

Foundation models for reasoning on charts

Posted by Julian Eisenschlos, Research Software Engineer, Google Research

Visual language is the form of communication that relies on pictorial symbols outside of text to convey information. It is ubiquitous in our digital life in the form of iconography, infographics, tables, plots, and charts, extending to the real world in street signs, comic books, food labels, etc. For that reason, having computers better understand this type of media can help with scientific communication and discovery, accessibility, and data transparency.

While computer vision models have made tremendous progress using learning-based solutions since the advent of ImageNet, the focus has been on natural images, where all sorts of tasks, such as classification, visual question answering (VQA), captioning, detection and segmentation, have been defined, studied and in some cases advanced to reach human performance. However, visual language has not garnered a similar level of attention, possibly because of the lack of large-scale training sets in this space. But over the last few years, new academic datasets have been created with the goal of evaluating question answering systems on visual language images, like PlotQA, InfographicsVQA, and ChartQA.

Example from ChartQA. Answering the question requires reading the information and computing the sum and the difference.

Existing models built for these tasks relied on integrating optical character recognition (OCR) information and their coordinates into larger pipelines but the process is error prone, slow, and generalizes poorly. The prevalence of these methods was because existing end-to-end computer vision models based on convolutional neural networks (CNNs) or transformers pre-trained on natural images could not be easily adapted to visual language. But existing models are ill-prepared for the challenges in answering questions on charts, including reading the relative height of bars or the angle of slices in pie charts, understanding axis scales, correctly mapping pictograms with their legend values with colors, sizes and textures, and finally performing numerical operations with the extracted numbers.

In light of these challenges, we propose “MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering”. MatCha, which stands for math and charts, is a pixels-to-text foundation model (a pre-trained model with built-in inductive biases that can be fine-tuned for multiple applications) trained on two complementary tasks: (a) chart de-rendering and (b) math reasoning. In chart de-rendering, given a plot or chart, the image-to-text model is required to generate its underlying data table or the code used to render it. For math reasoning pre-training, we pick textual numerical reasoning datasets and render the input into images, which the image-to-text model needs to decode for answers. We also propose “DePlot: One-shot visual language reasoning by plot-to-table translation”, a model built on top of MatCha for one-shot reasoning on charts via translation to tables. With these methods we surpass the previous state of the art in ChartQA by more than 20% and match the best summarization systems that have 1000 times more parameters. Both papers will be presented at ACL2023.

Chart de-rendering

Plots and charts are usually generated by an underlying data table and a piece of code. The code defines the overall layout of the figure (e.g., type, direction, color/shape scheme) and the underlying data table establishes the actual numbers and their groupings. Both the data and code are sent to a compiler/rendering engine to create the final image. To understand a chart, one needs to discover the visual patterns in the image and effectively parse and group them to extract the key information. Reversing the plot rendering process demands all such capabilities and can thus serve as an ideal pre-training task.

A chart created from a table in the Airbus A380 Wikipedia page using random plotting options. The pre-training task for MatCha consists of recovering the source table or the source code from the image.

In practice, it is challenging to simultaneously obtain charts, their underlying data tables, and their rendering code. To collect sufficient pre-training data, we independently accumulate [chart, code] and [chart, table] pairs. For [chart, code], we crawl all GitHub IPython notebooks with appropriate licenses and extract blocks with figures. A figure and the code block right before it are saved as a [chart, code] pair. For [chart, table] pairs, we explored two sources. For the first source, synthetic data, we manually write code to convert web-crawled Wikipedia tables from the TaPas codebase to charts. We sampled from and combined several plotting options depending on the column types. In addition, we also add [chart, table] pairs generated in PlotQA to diversify the pre-training corpus. The second source is web-crawled [chart, table] pairs. We directly use the [chart, table] pairs crawled in the ChartQA training set, containing around 20k pairs in total from four websites: Statista, Pew, Our World in Data, and OECD.

Math reasoning

We incorporate numerical reasoning knowledge into MatCha by learning math reasoning skills from textual math datasets. We use two existing textual math reasoning datasets, MATH and DROP for pre-training. MATH is synthetically created, containing two million training examples per module (type) of questions. DROP is a reading-comprehension–style QA dataset where the input is a paragraph context and a question.

To solve questions in DROP, the model needs to read the paragraph, extract relevant numbers and perform numerical computation. We found both datasets to be complementary. MATH contains a large number of questions across different categories, which helps us identify math operations needed to explicitly inject into the model. DROP’s reading-comprehension format resembles the typical QA format wherein models simultaneously perform information extraction and reasoning. In practice, we render inputs of both datasets into images. The model is trained to decode the answer.

To improve the math reasoning skills of MatCha we incorporate examples from MATH and DROP into the pre-training objective, by rendering the input text as images.

End-to-end results

We use a Pix2Struct model backbone, which is an image-to-text transformer tailored for website understanding, and pre-train it with the two tasks described above. We demonstrate the strengths of MatCha by fine-tuning it on several visual language tasks — tasks involving charts and plots for question answering and summarization where no access to the underlying table is possible. MatCha surpasses previous models’ performance by a large margin and also outperforms the previous state of the art, which assumes access to underlying tables.

In the figure below, we first evaluate two baseline models that incorporate information from an OCR pipeline, which until recently was the standard approach for working with charts. The first is based on T5, the second on VisionTaPas. We also compare against PaLI-17B, which is a large (~1000 times larger than the other models) image plus text-to-text transformer trained on a diverse set of tasks but with limited capabilities for reading text and other forms of visual language. Finally, we report the Pix2Struct and MatCha model results.

Experimental results on two chart QA benchmarks ChartQA & PlotQA (using relaxed accuracy) and a chart summarization benchmark chart-to-text (using BLEU4). Matcha surpasses the state of the art by a large margin on QA, compared to larger models, and matches these larger models on summarization.

For QA datasets, we use the official relaxed accuracy metric that allows for small relative errors in numerical outputs. For chart-to-text summarization, we report BLEU scores. MatCha achieves noticeably improved results compared to baselines for question answering, and comparable results to PaLI in summarization, where large size and extensive long text/captioning generation pre-training are advantageous for this kind of long-form text generation.

Derendering plus large language model chains

While extremely performant for their number of parameters, particularly on extractive tasks, we observed that fine-tuned MatCha models could still struggle with end-to-end complex reasoning (e.g., mathematical operations involving large numbers or multiple steps). Thus, we also propose a two-step method to tackle this: 1) a model reads a chart, then outputs the underlying table, 2) a large language model (LLM) reads this output and then tries to answer the question solely based on the textual input.

For the first model, we fine-tuned MatCha solely on the chart-to-table task, increasing the output sequence length to guarantee it could recover all or most of the information in the chart. DePlot is the resulting model. In the second stage, any LLM (such as FlanPaLM or Codex) can be used for the task, and we can rely on the standard methods to increase performance on LLMs, for example chain-of-thought and self-consistency. We also experimented with program-of-thoughts where the model produces executable Python code to offload complex computations.

An illustration of the DePlot+LLM method. This is a real example using FlanPaLM and Codex. The blue boxes are input to the LLM and the red boxes contain the answer generated by the LLMs. We highlight some of the key reasoning steps in each answer.

As shown in the example above, the DePlot model in combination with LLMs outperforms fine-tuned models by a significant margin, especially so in the human-sourced portion of ChartQA, where the questions are more natural but demand more difficult reasoning. Furthermore, DePlot+LLM can do so without access to any training data.

We have released the new models and code at our GitHub repo, where you can try it out yourself in colab. Checkout the papers for MatCha and DePlot for more details on the experimental results. We hope that our results can benefit the research community and make the information in charts and plots more accessible to everyone.

Acknowledgements

This work was carried out by Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen and Yasemin Altun from our Language Team as part of Fangyu's internship project. Nigel Collier from Cambridge also was a collaborator. We would like to thank Joshua Howland, Alex Polozov, Shrestha Basu Mallick, Massimo Nicosia and William Cohen for their valuable comments and suggestions.

Source: Google AI Blog

Foundation models for reasoning on charts

Posted by Julian Eisenschlos, Research Software Engineer, Google Research

Example from ChartQA. Answering the question requires reading the information and computing the sum and the difference.

Chart de-rendering

Math reasoning

To improve the math reasoning skills of MatCha we incorporate examples from MATH and DROP into the pre-training objective, by rendering the input text as images.

End-to-end results

Derendering plus large language model chains

Acknowledgements

Source: Google AI Blog

Resolving code review comments with ML

Posted by Alexander Frömmgen, Staff Software Engineer, and Lera Kharatyan, Senior Software Engineer, Core Systems & Experiences

Code-change reviews are a critical part of the software development process at scale, taking a significant amount of the code authors’ and the code reviewers’ time. As part of this process, the reviewer inspects the proposed code and asks the author for code changes through comments written in natural language. At Google, we see millions of reviewer comments per year, and authors require an average of ~60 minutes active shepherding time between sending changes for review and finally submitting the change. In our measurements, the required active work time that the code author must do to address reviewer comments grows almost linearly with the number of comments. However, with machine learning (ML), we have an opportunity to automate and streamline the code review process, e.g., by proposing code changes based on a comment’s text.

Today, we describe applying recent advances of large sequence models in a real-world setting to automatically resolve code review comments in the day-to-day development workflow at Google (publication forthcoming). As of today, code-change authors at Google address a substantial amount of reviewer comments by applying an ML-suggested edit. We expect that to reduce time spent on code reviews by hundreds of thousands of hours annually at Google scale. Unsolicited, very positive feedback highlights that the impact of ML-suggested code edits increases Googlers' productivity and allows them to focus on more creative and complex tasks.

Predicting the code edit

We started by training a model that predicts code edits needed to address reviewer comments. The model is pre-trained on various coding tasks and related developer activities (e.g., renaming a variable, repairing a broken build, editing a file). It’s then fine-tuned for this specific task with reviewed code changes, the reviewer comments, and the edits the author performed to address those comments.

An example of an ML-suggested edit of refactorings that are spread within the code.

Google uses a monorepo, a single repository for all of its software artifacts, which allows our training dataset to include all unrestricted code used to build Google's most recent software, as well as previous versions.

To improve the model quality, we iterated on the training dataset. For example, we compared the model performance for datasets with a single reviewer comment per file to datasets with multiple comments per file, and experimented with classifiers to clean up the training data based on a small, curated dataset to choose the model with the best offline precision and recall metrics.

Serving infrastructure and user experience

We designed and implemented the feature on top of the trained model, focusing on the overall user experience and developer efficiency. As part of this, we explored different user experience (UX) alternatives through a series of user studies. We then refined the feature based on insights from an internal beta (i.e., a test of the feature in development) including user feedback (e.g., a “Was this helpful?” button next to the suggested edit).

The final model was calibrated for a target precision of 50%. That is, we tuned the model and the suggestions filtering, so that 50% of suggested edits on our evaluation dataset are correct. In general, increasing the target precision reduces the number of shown suggested edits, and decreasing the target precision leads to more incorrect suggested edits. Incorrect suggested edits take the developers time and reduce the developers’ trust in the feature. We found that a target precision of 50% provides a good balance.

At a high level, for every new reviewer comment, we generate the model input in the same format that is used for training, query the model, and generate the suggested code edit. If the model is confident in the prediction and a few additional heuristics are satisfied, we send the suggested edit to downstream systems. The downstream systems, i.e., the code review frontend and the integrated development environment (IDE), expose the suggested edits to the user and log user interactions, such as preview and apply events. A dedicated pipeline collects these logs and generates aggregate insights, e.g., the overall acceptance rates as reported in this blog post.

Architecture of the ML-suggested edits infrastructure. We process code and infrastructure from multiple services, get the model predictions and surface the predictions in the code review tool and IDE.

The developer interacts with the ML-suggested edits in the code review tool and the IDE. Based on insights from the user studies, the integration into the code review tool is most suitable for a streamlined review experience. The IDE integration provides additional functionality and supports 3-way merging of the ML-suggested edits (left in the figure below) in case of conflicting local changes on top of the reviewed code state (right) into the merge result (center).

3-way-merge UX in IDE.

Results

Offline evaluations indicate that the model addresses 52% of comments with a target precision of 50%. The online metrics of the beta and the full internal launch confirm these offline metrics, i.e., we see model suggestions above our target model confidence for around 50% of all relevant reviewer comments. 40% to 50% of all previewed suggested edits are applied by code authors.

We used the “not helpful” feedback during the beta to identify recurring failure patterns of the model. We implemented serving-time heuristics to filter these and, thus, reduce the number of shown incorrect predictions. With these changes, we traded quantity for quality and observed an increased real-world acceptance rate.

Code review tool UX. The suggestion is shown as part of the comment and can be previewed, applied and rated as helpful or not helpful.

Our beta launch showed a discoverability challenge: code authors only previewed ~20% of all generated suggested edits. We modified the UX and introduced a prominent “Show ML-edit” button (see the figure above) next to the reviewer comment, leading to an overall preview rate of ~40% at launch. We additionally found that suggested edits in the code review tool are often not applicable due to conflicting changes that the author did during the review process. We addressed this with a button in the code review tool that opens the IDE in a merge view for the suggested edit. We now observe that more than 70% of these are applied in the code review tool and fewer than 30% are applied in the IDE. All these changes allowed us to increase the overall fraction of reviewer comments that are addressed with an ML-suggested edit by a factor of 2 from beta to the full internal launch. At Google scale, these results help automate the resolution of hundreds of thousands of comments each year.

Suggestions filtering funnel.

We see ML-suggested edits addressing a wide range of reviewer comments in production. This includes simple localized refactorings and refactorings that are spread within the code, as shown in the examples throughout the blog post above. The feature addresses longer and less formally-worded comments that require code generation, refactorings and imports.

Example of a suggestion for a longer and less formally worded comment that requires code generation, refactorings and imports.

The model can also respond to complex comments and produce extensive code edits (shown below). The generated test case follows the existing unit test pattern, while changing the details as described in the comment. Additionally, the edit suggests a comprehensive name for the test reflecting the test semantics.

Example of the model's ability to respond to complex comments and produce extensive code edits.

Conclusion and future work

In this post, we introduced an ML-assistance feature to reduce the time spent on code review related changes. At the moment, a substantial amount of all actionable code review comments on supported languages are addressed with applied ML-suggested edits at Google. A 12-week A/B experiment across all Google developers will further measure the impact of the feature on the overall developer productivity.

We are working on improvements throughout the whole stack. This includes increasing the quality and recall of the model and building a more streamlined experience for the developer with improved discoverability throughout the review process. As part of this, we are investigating the option of showing suggested edits to the reviewer while they draft comments and expanding the feature into the IDE to enable code-change authors to get suggested code edits for natural-language commands.

Acknowledgements

This is the work of many people in Google Core Systems & Experiences team, Google Research, and DeepMind. We'd like to specifically thank Peter Choy for bringing the collaboration together, and all of our team members for their key contributions and useful advice, including Marcus Revaj, Gabriela Surita, Maxim Tabachnyk, Jacob Austin, Nimesh Ghelani, Dan Zheng, Peter Josling, Mariana Stariolo, Chris Gorgolewski, Sascha Varkevisser, Katja Grünwedel, Alberto Elizondo, Tobias Welp, Paige Bailey, Pierre-Antoine Manzagol, Pascal Lamblin, Chenjie Gu, Petros Maniatis, Henryk Michalewski, Sara Wiltberger, Ambar Murillo, Satish Chandra, Madhura Dudhgaonkar, Niranjan Tulpule, Zoubin Ghahramani, Juanjo Carin, Danny Tarlow, Kevin Villela, Stoyan Nikolov, David Tattersall, Boris Bokowski, Kathy Nix, Mehdi Ghissassi, Luis C. Cobo, Yujia Li, David Choi, Kristóf Molnár, Vahid Meimand, Amit Patel, Brett Wiltshire, Laurent Le Brun, Mingpan Guo, Hermann Loose, Jonas Mattes, Savinee Dancs.

Source: Google AI Blog

Using reinforcement learning for dynamic planning in open-ended conversations

Posted by Deborah Cohen, Staff Research Scientist, and Craig Boutilier, Principal Scientist, Google Research

As virtual assistants become ubiquitous, users increasingly interact with them to learn about new topics or obtain recommendations and expect them to deliver capabilities beyond narrow dialogues of one or two turns. Dynamic planning, namely the capability to look ahead and replan based on the flow of the conversation, is an essential ingredient for the making of engaging conversations with the deeper, open-ended interactions that users expect.

While large language models (LLMs) are now beating state-of-the-art approaches in many natural language processing benchmarks, they are typically trained to output the next best response, rather than planning ahead, which is required for multi-turn interactions. However, in the past few years, reinforcement learning (RL) has delivered incredible results addressing specific problems that involve dynamic planning, such as winning games and protein folding.

Today, we are sharing our recent advances in dynamic planning for human-to-assistant conversations, in which we enable an assistant to plan a multi-turn conversation towards a goal and adapt that plan in real-time by adopting an RL-based approach. Here we look at how to improve long interactions by applying RL to compose answers based on information extracted from reputable sources, rather than relying on content generated by a language model. We expect that future versions of this work could combine LLMs and RL in multi-turn dialogues. The deployment of RL “in the wild” in a large-scale dialogue system proved a formidable challenge due to the modeling complexity, tremendously large state and action spaces, and significant subtlety in designing reward functions.

What is dynamic planning?

Many types of conversations, from gathering information to offering recommendations, require a flexible approach and the ability to modify the original plan for the conversation based on its flow. This ability to shift gears in the middle of a conversation is known as dynamic planning, as opposed to static planning, which refers to a more fixed approach. In the conversation below, for example, the goal is to engage the user by sharing interesting facts about cool animals. To begin, the assistant steers the conversation to sharks via a sound quiz. Given the user's lack of interest in sharks, the assistant then develops an updated plan and pivots the conversation to sea lions, lions, and then cheetahs.

The assistant dynamically modifies its original plan to talk about sharks and shares facts about other animals.

Dynamic composition

To cope with the challenge of conversational exploration, we separate the generation of assistant responses into two parts: 1) content generation, which extracts relevant information from reputable sources, and 2) flexible composition of such content into assistant responses. We refer to this two-part approach as dynamic composition. Unlike LLM methods, this approach gives the assistant the ability to fully control the source, correctness, and quality of the content that it may offer. At the same time, it can achieve flexibility via a learned dialogue manager that selects and combines the most appropriate content.

In an earlier paper, “Dynamic Composition for Conversational Domain Exploration”, we describe a novel approach which consists of: (1) a collection of content providers, which offer candidates from different sources, such as news snippets, knowledge graph facts, and questions; (2) a dialogue manager; and (3) a sentence fusion module. Each assistant response is incrementally constructed by the dialogue manager, which selects candidates proposed by the content providers. The selected sequence of utterances is then fused into a cohesive response.

Dynamic planning using RL

At the core of the assistant response composition loop is a dialogue manager trained using off-policy RL, namely an algorithm that evaluates and improves a policy that is different from the policy used by the agent (in our case, the latter is based on a supervised model). Applying RL to dialogue management presents several challenges, including a large state space (as the state represents the conversation state, which needs to account for the whole conversation history) and an effectively unbounded action space (that may include all existing words or sentences in natural language).

We address these challenges using a novel RL construction. First, we leverage powerful supervised models — specifically, recurrent neural networks (RNNs) and transformers — to provide a succinct and effective dialogue state representation. These state encoders are fed with the dialogue history, composed of a sequence of user and assistant turns, and output a representation of the dialogue state in the form of a latent vector.

Second, we use the fact that a relatively small set of reasonable candidate utterances or actions can be generated by content providers at each conversation turn, and limit the action space to these. Whereas the action space is typically fixed in RL settings, because all states share the same action space, ours is a non-standard space in which the candidate actions may differ with each state, since content providers generate different actions depending on the dialogue context. This puts us in the realm of stochastic action sets, a framework that formalizes cases where the set of actions available in each state is governed by an exogenous stochastic process, which we address using Stochastic Action Q-Learning, a variant of the Q-learning approach. Q-learning is a popular off-policy RL algorithm, which does not require a model of the environment to evaluate and improve the policy. We trained our model on a corpus of crowd-compute–rated conversations obtained using a supervised dialogue manager.

Given the current dialogue history and a new user query, content providers generate candidates from which the assistant selects one. This process runs in a loop, and at the end the selected utterances are fused into a cohesive response.

Reinforcement learning model evaluation

We compared our RL dialogue manager with a launched supervised transformer model in an experiment using Google Assistant, which conversed with users about animals. A conversation starts when a user triggers the experience by asking an animal-related query (e.g., “How does a lion sound?”). The experiment was conducted using an A/B testing protocol, in which a small percentage of Assistant users were randomly sampled to interact with our RL-based assistant while other users interacted with the standard assistant.

We found that the RL dialogue manager conducts longer, more engaging conversations. It increases conversation length by 30% while improving user engagement metrics. We see an increase of 8% in cooperative responses to the assistant’s questions — e.g., “Tell me about lions,” in response to “Which animal do you want to hear about next?” Although there is also a large increase in nominally “non-cooperative” responses (e.g., “No,” as a reply to a question proposing additional content, such as “Do you want to hear more?”), this is expected as the RL agent takes more risks by asking pivoting questions. While a user may not be interested in the conversational direction proposed by the assistant (e.g., pivoting to another animal), the user will often continue to engage in a dialogue about animals.

From the non-cooperative user response in the 3rd turn (“No.”) and the query “Make a dog sound,” in the 5th turn, the assistant recognizes that the user is mostly interested in animal sounds and modifies its plan, providing sounds and sound quizzes.

In addition, some user queries contain explicit positive (e.g., “Thank you, Google,” or “I’m happy.”) or negative (e.g., “Shut up,” or “Stop.”) feedback. While an order of magnitude fewer than other queries, they offer a direct measure of user (dis)satisfaction. The RL model increases explicit positive feedback by 32% and reduces negative feedback by 18%.

Learned dynamic planning characteristics and strategies

We observe several characteristics of the (unseen) RL plan to improve user engagement while conducting longer conversations. First, the RL-based assistant ends 20% more turns in questions, prompting the user to choose additional content. It also better harnesses content diversity, including facts, sounds, quizzes, yes/no questions, open questions, etc. On average, the RL assistant uses 26% more distinct content providers per conversation than the supervised model.

Two observed RL planning strategies are related to the existence of sub-dialogues with different characteristics. Sub-dialogues about animal sounds are poorer in content and exhibit entity pivoting at every turn (i.e., after playing the sound of a given animal, we can either suggest the sound of a different animal or quiz the user about other animal sounds). In contrast, sub-dialogues involving animal facts typically contain richer content and have greater conversation depth. We observe that RL favors the richer experience of the latter, selecting 31% more fact-related content. Lastly, when restricting analysis to fact-related dialogues, the RL assistant exhibits 60% more focus-pivoting turns, that is, conversational turns that change the focus of the dialogue.

Below, we show two example conversations, one conducted by the supervised model (left) and the second by the RL model (right), in which the first three user turns are identical. With a supervised dialogue manager, after the user declined to hear about “today’s animal”, the assistant pivots back to animal sounds to maximize the immediate user satisfaction. While the conversation conducted by the RL model begins identically, it exhibits a different planning strategy to optimize the overall user engagement, introducing more diverse content, such as fun facts.

In the left conversation, conducted by the supervised model, the assistant maximizes the immediate user satisfaction. The right conversation, conducted by the RL model, shows different planning strategies to optimize the overall user engagement.

Future research and challenges

In the past few years, LLMs trained for language understanding and generation have demonstrated impressive results across multiple tasks, including dialogue. We are now exploring the use of an RL framework to empower LLMs with the capability of dynamic planning so that they can dynamically plan ahead and delight users with a more engaging experience.

Acknowledgements

The work described is co-authored by: Moonkyung Ryu, Yinlam Chow, Orgad Keller, Ido Greenberg, Avinatan Hassidim, Michael Fink, Yossi Matias, Idan Szpektor and Gal Elidan. We would like to thank: Roee Aharoni, Moran Ambar, John Anderson, Ido Cohn, Mohammad Ghavamzadeh, Lotem Golany, Ziv Hodak, Adva Levin, Fernando Pereira, Shimi Salant, Shachar Shimoni, Ronit Slyper, Ariel Stolovich, Hagai Taitelbaum, Noam Velan, Avital Zipori and the CrowdCompute team led by Ashwin Kakarla. We thank Sophie Allweis for her feedback on this blogpost and Tom Small for the visualization.

Source: Google AI Blog

Larger language models do in-context learning differently

Posted by Jerry Wei, Student Researcher, and Denny Zhou, Principal Scientist, Google Research

There have recently been tremendous advances in language models, partly because they can perform tasks with strong performance via in-context learning (ICL), a process whereby models are prompted with a few examples of input-label pairs before performing the task on an unseen evaluation example. In general, models’ success at in-context learning is enabled by:

Their use of semantic prior knowledge from pre-training to predict labels while following the format of in-context examples (e.g., seeing examples of movie reviews with “positive sentiment” and “negative sentiment” as labels and performing sentiment analysis using prior knowledge).
Learning the input-label mappings in context from the presented examples (e.g., finding a pattern that positive reviews should be mapped to one label, and negative reviews should be mapped to a different label).

In “Larger language models do in-context learning differently”, we aim to learn about how these two factors (semantic priors and input-label mappings) interact with each other in ICL settings, especially with respect to the scale of the language model that’s used. We investigate two settings to study these two factors — ICL with flipped labels (flipped-label ICL) and ICL with semantically-unrelated labels (SUL-ICL). In flipped-label ICL, labels of in-context examples are flipped so that semantic priors and input-label mappings disagree with each other. In SUL-ICL, labels of in-context examples are replaced with words that are semantically unrelated to the task presented in-context. We found that overriding prior knowledge is an emergent ability of model scale, as is the ability to learn in-context with semantically-unrelated labels. We also found that instruction tuning strengthens the use of prior knowledge more than it increases the capacity to learn input-label mappings.

An overview of flipped-label ICL and semantically-unrelated label ICL (SUL-ICL), compared with regular ICL, for a sentiment analysis task. Flipped-label ICL uses flipped labels, forcing the model to override semantic priors in order to follow the in-context examples. SUL-ICL uses labels that are not semantically related to the task, which means that models must learn input-label mappings in order to perform the task because they can no longer rely on the semantics of natural language labels.

Experiment design

For a diverse dataset mixture, we experiment on seven natural language processing (NLP) tasks that have been widely used: sentiment analysis, subjective/objective classification, question classification, duplicated-question recognition, entailment recognition, financial sentiment analysis, and hate speech detection. We test five language model families, PaLM, Flan-PaLM, GPT-3, InstructGPT, and Codex.

Flipped labels

In this experiment, labels of in-context examples are flipped, meaning that prior knowledge and input-label mappings disagree (e.g., sentences containing positive sentiment labeled as “negative sentiment”), thereby allowing us to study whether models can override their priors. In this setting, models that are able to override prior knowledge and learn input-label mappings in-context should experience a decrease in performance (since ground-truth evaluation labels are not flipped).

The ability to override semantic priors when presented with flipped in-context example labels emerges with model scale. Smaller models cannot flip predictions to follow flipped labels (performance only decreases slightly), while larger models can do so (performance decreases to well below 50%).

We found that when no labels are flipped, larger models have better performance than smaller models (as expected). But when we flip more and more labels, the performance of small models stays relatively flat, but large models experience large performance drops to well-below random guessing (e.g., 90% → 22.5% for code-davinci-002).

These results indicate that large models can override prior knowledge from pre-training when contradicting input-label mappings are presented in-context. Small models can’t do this, making this ability an emergent phenomena of model scale.

Semantically-unrelated labels

In this experiment, we replace labels with semantically-irrelevant ones (e.g., for sentiment analysis, we use “foo/bar” instead of “negative/positive”), which means that the model can only perform ICL by learning from input-label mappings. If a model mostly relies on prior knowledge for ICL, then its performance should decrease after this change since it will no longer be able to use semantic meanings of labels to make predictions. A model that can learn input–label mappings in-context, on the other hand, would be able to learn these semantically-unrelated mappings and should not experience a major drop in performance.

Small models rely more on semantic priors than large models do, as indicated by the greater decrease in performance for small models than for large models when using semantically-unrelated labels (i.e., targets) instead of natural language labels. For each plot, models are shown in order of increasing model size (e.g., for GPT-3 models, a is smaller than b, which is smaller than c).

Indeed, we see that using semantically-unrelated labels results in a greater performance drop for small models. This suggests that smaller models primarily rely on their semantic priors for ICL rather than learning from the presented input-label mappings. Large models, on the other hand, have the ability to learn input-label mappings in-context when the semantic nature of labels is removed.

We also find that including more in-context examples (i.e., exemplars) results in a greater performance improvement for large models than it does for small models, indicating that large models are better at learning from in-context examples than small models are.

In the SUL-ICL setup, larger models benefit more from additional examples than smaller models do.

Instruction tuning

Instruction tuning is a popular technique for improving model performance, which involves tuning models on various NLP tasks that are phrased as instructions (e.g., “Question: What is the sentiment of the following sentence, ‘This movie is great.’ Answer: Positive”). Since the process uses natural language labels, however, an open question is whether it improves the ability to learn input-label mappings or whether it strengthens the ability to recognize and apply semantic prior knowledge. Both of these would lead to an improvement in performance on standard ICL tasks, so it’s unclear which of these occur.

We study this question by running the same two setups as before, only this time we focus on comparing standard language models (specifically, PaLM) with their instruction-tuned variants (Flan-PaLM).

First, we find that Flan-PaLM is better than PaLM when we use semantically-unrelated labels. This effect is very prominent in small models, as Flan-PaLM-8B outperforms PaLM-8B by 9.6% and almost catches up to PaLM-62B. This trend suggests that instruction tuning strengthens the ability to learn input-label mappings, which isn’t particularly surprising.

Instruction-tuned language models are better at learning input–label mappings than pre-training–only language models are.

More interestingly, we saw that Flan-PaLM is actually worse than PaLM at following flipped labels, meaning that the instruction tuned models were unable to override their prior knowledge (Flan-PaLM models don’t reach below random guessing with 100% flipped labels, but PaLM models without instruction tuning can reach 31% accuracy in the same setting). These results indicate that instruction tuning must increase the extent to which models rely on semantic priors when they’re available.

Instruction-tuned models are worse than pre-training–only models at learning to override semantic priors when presented with flipped labels in-context.

Combined with the previous result, we conclude that although instruction tuning improves the ability to learn input-label mappings, it strengthens the usage of semantic prior knowledge more.

Conclusion

We examined the extent to which language models learn in-context by utilizing prior knowledge learned during pre-training versus input-label mappings presented in-context.

We first showed that large language models can learn to override prior knowledge when presented with enough flipped labels, and that this ability emerges with model scale. We then found that successfully doing ICL using semantically-unrelated labels is another emergent ability of model scale. Finally, we analyzed instruction-tuned language models and saw that instruction tuning improves the capacity to learn input-label mappings but also strengthens the use of semantic prior knowledge even more.

Future work

These results underscore how the ICL behavior of language models can change depending on their scale, and that larger language models have an emergent ability to map inputs to many types of labels, a form of reasoning in which input-label mappings can potentially be learned for arbitrary symbols. Future research could help provide insights on why these phenomena occur with respect to model scale.

Acknowledgements

This work was conducted by Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou, and Tengyu Ma. We would like to thank Sewon Min and our fellow collaborators at Google Research for their advice and helpful discussions.

Source: Google AI Blog

A vision-language approach for foundational UI understanding

Posted by Yang Li, Research Scientist, and Gang Li, Software Engineer, Google Research

The computational understanding of user interfaces (UI) is a key step towards achieving intelligent UI behaviors. Previously, we investigated various UI modeling tasks, including widget captioning, screen summarization, and command grounding, that address diverse interaction scenarios such as automation and accessibility. We also demonstrated how machine learning can help user experience practitioners improve UI quality by diagnosing tappability confusion and providing insights for improving UI design. These works along with those developed by others in the field have showcased how deep neural networks can potentially transform end user experiences and the interaction design practice.

With these successes in addressing individual UI tasks, a natural question is whether we can obtain foundational understandings of UIs that can benefit specific UI tasks. As our first attempt to answer this question, we developed a multi-task model to address a range of UI tasks simultaneously. Although the work made some progress, a few challenges remain. Previous UI models heavily rely on UI view hierarchies — i.e., the structure or metadata of a mobile UI screen like the Document Object Model for a webpage — that allow a model to directly acquire detailed information of UI objects on the screen (e.g., their types, text content and positions). This metadata has given previous models advantages over their vision-only counterparts. However, view hierarchies are not always accessible, and are often corrupted with missing object descriptions or misaligned structure information. As a result, despite the short-term gains from using view hierarchies, it may ultimately hamper the model performance and applicability. In addition, previous models had to deal with heterogeneous information across datasets and UI tasks, which often resulted in complex model architectures that were difficult to scale or generalize across tasks.

In “Spotlight: Mobile UI Understanding using Vision-Language Models with a Focus”, accepted for publication at ICLR 2023, we present a vision-only approach that aims to achieve general UI understanding completely from raw pixels. We introduce a unified approach to represent diverse UI tasks, the information for which can be universally represented by two core modalities: vision and language. The vision modality captures what a person would see from a UI screen, and the language modality can be natural language or any token sequences related to the task. We demonstrate that Spotlight substantially improves accuracy on a range of UI tasks, including widget captioning, screen summarization, command grounding and tappability prediction.

Spotlight Model

The Spotlight model input includes a tuple of three items: the screenshot, the region of interest on the screen, and the text description of the task. The output is a text description or response about the region of interest. This simple input and output representation of the model is expressive to capture various UI tasks and allows scalable model architectures. This model design allows a spectrum of learning strategies and setups, from task-specific fine-tuning, to multi-task learning and to few-shot learning. The Spotlight model, as illustrated in the above figure, leverages existing architecture building blocks such as ViT and T5 that are pre-trained in the high-resourced, general vision-language domain, which allows us to build on top of the success of these general domain models.

Because UI tasks are often concerned with a specific object or area on the screen, which requires a model to be able to focus on the object or area of interest, we introduce a Focus Region Extractor to a vision-language model that enables the model to concentrate on the region in light of the screen context.

In particular, we design a Region Summarizer that acquires a latent representation of a screen region based on ViT encodings by using attention queries generated from the bounding box of the region (see paper for more details). Specifically, each coordinate (a scalar value, i.e., the left, top, right or bottom) of the bounding box, denoted as a yellow box on the screenshot, is first embedded via a multilayer perceptron (MLP) as a collection of dense vectors, and then fed to a Transformer model along their coordinate-type embedding. The dense vectors and their corresponding coordinate-type embeddings are color coded to indicate their affiliation with each coordinate value. Coordinate queries then attend to screen encodings output by ViT via cross attention, and the final attention output of the Transformer is used as the region representation for the downstream decoding by T5.

A target region on the screen is summarized by using its bounding box to query into screen encodings from ViT via attentional mechanisms.

Results

We pre-train the Spotlight model using two unlabeled datasets (an internal dataset based on C4 corpus and an internal mobile dataset) with 2.5 million mobile UI screens and 80 million web pages. We then separately fine-tune the pre-trained model for each of the four downstream tasks (captioning, summarization, grounding, and tappability). For widget captioning and screen summarization tasks, we report CIDEr scores, which measure how similar a model text description is to a set of references created by human raters. For command grounding, we report accuracy that measures the percentage of times the model successfully locates a target object in response to a user command. For tappability prediction, we report F1 scores that measure the model’s ability to tell tappable objects from untappable ones.

In this experiment, we compare Spotlight with several benchmark models. Widget Caption uses view hierarchy and the image of each UI object to generate a text description for the object. Similarly, Screen2Words uses view hierarchy and the screenshot as well as auxiliary features (e.g., app description) to generate a summary for the screen. In the same vein, VUT combines screenshots and view hierarchies for performing multiple tasks. Finally, the original Tappability model leverages object metadata from view hierarchy and the screenshot to predict object tappability. Taperception, a follow-up model of Tappability, uses a vision-only tappability prediction approach. We examine two Spotlight model variants with respect to the size of its ViT building block, including B/16 and L/16. Spotlight drastically exceeded the state-of-the-art across four UI modeling tasks.

	Model	Captioning	Summarization	Grounding	Tappability
Baselines	Widget Caption	97	-	-	-
	Screen2Words	-	61.3	-	-
	VUT	99.3	65.6	82.1	-
	Taperception	-	-	-	85.5
	Tappability	-	-	-	87.9
Spotlight	B/16	136.6	103.5	95.7	86.9
Spotlight	L/16	141.8	106.7	95.8	88.4

We then pursue a more challenging setup where we ask the model to learn multiple tasks simultaneously because a multi-task model can substantially reduce model footprint. As shown in the table below, the experiments showed that our model still performs competitively.

Model	Captioning	Summarization	Grounding	Tappability
VUT multi-task	99.3	65.1	80.8	-
Spotlight B/16	140	102.7	90.8	89.4
Spotlight L/16	141.3	99.2	94.2	89.5

To understand how the Region Summarizer enables Spotlight to focus on a target region and relevant areas on the screen, we analyze the attention weights (which indicate where the model attention is on the screenshot) for both widget captioning and screen summarization tasks. In the figure below, for the widget captioning task, the model predicts “select Chelsea team” for the checkbox on the left side, highlighted with a red bounding box. We can see from its attention heatmap (which illustrates the distribution of attention weights) on the right that the model learns to attend to not only the target region of the check box, but also the text “Chelsea" on the far left to generate the caption. For the screen summarization example, the model predicts “page displaying the tutorial of a learning app” given the screenshot on the left. In this example, the target region is the entire screen, and the model learns to attend to important parts on the screen for summarization.

For the widget captioning task, the attention heatmap shows the model attending to the checkbox, i.e., the target object, and the text label on its left when generating a caption for the object. The red bounding box in the figure is for illustration purposes.

For the screen summarization task that the target region encloses the entire screen, the attention heatmap shows the model attending to various locations on the screen that contribute to generating the summary.

Conclusion

We demonstrate that Spotlight outperforms previous methods that use both screenshots and view hierarchies as the input, and establishes state-of-the-art results on multiple representative UI tasks. These tasks range from accessibility, automation to interaction design and evaluation. Our vision-only approach for mobile UI understanding alleviates the need to use view hierarchy, allows the architecture to easily scale and benefits from the success of large vision-language models pre-trained for the general domain. Compared to recent large vision-language model efforts such as Flamingo and PaLI, Spotlight is relatively small and our experiments show the trend that larger models yield better performance. Spotlight can be easily applied to more UI tasks and potentially advance the fronts of many interaction and user experience tasks.

Acknowledgment

We thank Mandar Joshi and Tao Li for their help in processing the web pre-training dataset, and Chin-Yi Cheng and Forrest Huang for their feedback for proofreading the paper. Thanks to Tom Small for his help in creating animated figures in this post.

Source: Google AI Blog

Deciphering Clinical Abbreviations with Privacy Protecting ML

Posted by Posted by Alvin Rajkomar, Research Scientist, and Eric Loreaux, Software Engineer, Google Research

Today many people have digital access to their medical records, including their doctor’s clinical notes. However, clinical notes are hard to understand because of the specialized language that clinicians use, which contains unfamiliar shorthand and abbreviations. In fact, there are thousands of such abbreviations, many of which are specific to certain medical specialities and locales or can mean multiple things in different contexts. For example, a doctor might write in their clinical notes, “pt referred to pt for lbp“, which is meant to convey the statement: “Patient referred to physical therapy for low back pain.” Coming up with this translation is tough for laypeople and computers because some abbreviations are uncommon in everyday language (e.g., “lbp” means “low back pain”), and even familiar abbreviations, such as “pt” for “patient”, can have alternate meanings, such as “physical therapy.” To disambiguate between multiple meanings, the surrounding context must be considered. It’s no easy task to decipher all the meanings, and prior research suggests that expanding the shorthand and abbreviations can help patients better understand their health, diagnoses, and treatments.

In “Deciphering clinical abbreviations with a privacy protecting machine learning system”, published in Nature Communications, we report our findings on a general method that deciphers clinical abbreviations in a way that is both state-of-the-art and is on-par with board certified physicians in this task. We built the model using only public data on the web that wasn't associated with any patient (i.e., no potentially sensitive data) and evaluated performance on real, de-identified notes from inpatient and outpatient clinicians from different health systems. To enable the model to generalize from web-data to notes, we created a way to algorithmically re-write large amounts of internet text to look as if it were written by a doctor (called web-scale reverse substitution), and we developed a novel inference method, (called elicitive inference).

The model input is a string that may or may not contain medical abbreviations. We trained a model to output a corresponding string in which all abbreviations are simultaneously detected and expanded. If the input string does not contain an abbreviation, the model will output the original string. By Rajkomar et al used under CC BY 4.0/ Cropped from original.

Rewriting Text to Include Medical Abbreviations

Building a system to translate doctors’ notes would usually start with a large, representative dataset of clinical text where all abbreviations are labeled with their meanings. But no such dataset for general use by researchers exists. We therefore sought to develop an automated way to create such a dataset but without the use of any actual patient notes, which might include sensitive data. We also wanted to ensure that models trained on this data would still work well on real clinical notes from multiple hospital sites and types of care, such as both outpatient and inpatient.

To do this, we referenced a dictionary of thousands of clinical abbreviations and their expansions, and found sentences on the web that contained uses of the expansions from this dictionary. We then “rewrote” those sentences by abbreviating each expansion, resulting in web data that looked like it was written by a doctor. For instance, if a website contained the phrase “patients with atrial fibrillation can have chest pain,” we would rewrite this sentence to “pts with af can have cp.” We then used the abbreviated text as input to the model, with the original text serving as the label. This approach provided us with large amounts of data to train our model to perform abbreviation expansion.

The idea of “reverse substituting” the long-forms for their abbreviations was introduced in prior research, but our distributed algorithm allows us to extend the technique to large, web-sized datasets. Our algorithm, called web-scale reverse substitution (WSRS), is designed to ensure that rare terms occur more frequently and common terms are down-sampled across the public web to derive a more balanced dataset. With this data in-hand, we trained a series of large transformer-based language models to expand the web text.

We generate text to train our model on the decoding task by extracting phrases from public web pages that have corresponding medical abbreviations (shaded boxes on the left) and then substituting in the appropriate abbreviations (shaded dots, right). Since some words are found much more frequently than others ("patient" more than "posterior tibialis", both of which can be abbreviated “pt”), we downsampled common expansions to derive a more balanced dataset across the thousands of abbreviations. By Rajkomar et al used under CC BY 4.0.

Adapting Protein Alignment Algorithms to Unstructured Clinical Text

Evaluation of these models on the particular task of abbreviation expansion is difficult. Because they produce unstructured text as output, we had to figure out which abbreviations in the input correspond to which expansion in the output. To achieve this, we created a modified version of the Needleman Wunsch algorithm, which was originally designed for divergent sequence alignment in molecular biology, to align the model input and output and extract the corresponding abbreviation-expansion pairs. Using this alignment technique, we were able to evaluate the model’s capacity to detect and expand abbreviations accurately. We evaluated Text-to-Text Transfer Transformer (T5) models of various sizes (ranging from 60 million to over 60 billion parameters) and found that larger models performed translation better than smaller models, with the biggest model achieving the best performance.

Creating New Model Inference Techniques to Coax the Model

However, we did find something unexpected. When we evaluated the performance on multiple external test sets from real clinical notes, we found the models would leave some abbreviations unexpanded, and for larger models, the problem of incomplete expansion was even worse. This is mainly due to the fact that while we substitute expansions on the web for their abbreviations, we have no way of handling the abbreviations that are already present. This means that the abbreviations appear in both the original and rewritten text used as respective labels and input, and the model learns not to expand them.

To address this, we developed a new inference-chaining technique in which the model output is fed again as input to coax the model to make further expansions as long as the model is confident in the expansion. In technical terms, our best-performing technique, which we call elicitive inference, involves examining the outputs from a beam search above a certain log-likelihood threshold. Using elicitive inference, we were able to achieve state-of-the-art capability of expanding abbreviations in multiple external test sets.

Real example of the model’s input (left) and output (right).

Comparative Performance

We also sought to understand how patients and doctors currently perform at deciphering clinical notes, and how our model compared. We found that lay people (people without specific medical training) demonstrated less than 30% comprehension of the abbreviations present in the sample medical texts. When we allowed them to use Google Search, their comprehension increased to nearly 75%, still leaving 1 out of 5 abbreviations indecipherable. Unsurprisingly, medical students and trained physicians performed much better at the task with an accuracy of 90%. We found that our largest model was capable of matching or exceeding experts, with an accuracy of 98%.

How does the model perform so well compared to physicians in this task? There are two important factors in the model’s high comparative performance. Part of the discrepancy is that there were some abbreviations that clinicians did not even attempt to expand (such as "cm" for centimeter), which partly lowered the measured performance. This might seem unimportant, but for non-english speakers, these abbreviations may not be familiar, and so it may be helpful to have them written out. In contrast, our model is designed to comprehensively expand abbreviations. In addition, clinicians are familiar with abbreviations they commonly see in their speciality, but other specialists use shorthand that are not understood by those outside their fields. Our model is trained on thousands of abbreviations across multiple specialities and therefore can decipher a breadth of terms.

Towards Improved Health Literacy

We think there are numerous avenues in which large language models (LLMs) can help advance the health literacy of patients by augmenting the information they see and read. Most LLMs are trained on data that does not look like clinical note data, and the unique distribution of this data makes it challenging to deploy these models in an out-of-the-box fashion. We have demonstrated how to overcome this limitation. Our model also serves to "normalize" clinical note data, facilitating additional capabilities of ML to make the text easier for patients of all educational and health-literacy levels to understand.

Acknowledgements

This work was carried out in collaboration with Yuchen Liu, Jonas Kemp, Benny Li, Ming-Jun Chen, Yi Zhang, Afroz Mohiddin, and Juraj Gottweis. We thank Lisa Williams, Yun Liu, Arelene Chung, and Andrew Dai for many useful conversations and discussions about this work.

Source: Google AI Blog

Accelerating Text Generation with Confident Adaptive Language Modeling (CALM)

Posted by Tal Schuster, Research Scientist, Google Research

Language models (LMs) are the driving force behind many recent breakthroughs in natural language processing. Models like T5, LaMDA, GPT-3, and PaLM have demonstrated impressive performance on various language tasks. While multiple factors can contribute to improving the performance of LMs, some recent studies suggest that scaling up the model’s size is crucial for revealing emergent capabilities. In other words, some instances can be solved by small models, while others seem to benefit from increased scale.

Despite recent efforts that enabled the efficient training of LMs over large amounts of data, trained models can still be slow and costly for practical use. When generating text at inference time, most autoregressive LMs output content similar to how we speak and write (word after word), predicting each new word based on the preceding words. This process cannot be parallelized since LMs need to complete the prediction of one word before starting to compute the next one. Moreover, predicting each word requires significant computation given the model’s billions of parameters.

In “Confident Adaptive Language Modeling”, presented at NeurIPS 2022, we introduce a new method for accelerating the text generation of LMs by improving efficiency at inference time. Our method, named CALM, is motivated by the intuition that some next word predictions are easier than others. When writing a sentence, some continuations are trivial, while others might require more effort. Current LMs devote the same amount of compute power for all predictions. Instead, CALM dynamically distributes the computational effort across generation timesteps. By selectively allocating more computational resources only to harder predictions, CALM generates text faster while preserving output quality.

Confident Adaptive Language Modeling

When possible, CALM skips some compute effort for certain predictions. To demonstrate this, we use the popular encoder-decoder T5 architecture. The encoder reads the input text (e.g., a news article to summarize) and converts the text to dense representations. Then, the decoder outputs the summary by predicting it word by word. Both the encoder and decoder include a long sequence of Transformer layers. Each layer includes attention and feedforward modules with many matrix multiplications. These layers gradually modify the hidden representation that is ultimately used for predicting the next word.

Instead of waiting for all decoder layers to complete, CALM attempts to predict the next word earlier, after some intermediate layer. To decide whether to commit to a certain prediction or to postpone the prediction to a later layer, we measure the model’s confidence in its intermediate prediction. The rest of the computation is skipped only when the model is confident enough that the prediction won’t change. For quantifying what is “confident enough”, we calibrate a threshold that statistically satisfies arbitrary quality guarantees over the full output sequence.

Text generation with a regular language model (top) and with CALM (bottom). CALM attempts to make early predictions. Once confident enough (darker blue tones), it skips ahead and saves time.

Language Models with Early Exits

Enabling this early exit strategy for LMs requires minimal modifications to the training and inference processes. During training, we encourage the model to produce meaningful representations in intermediate layers. Instead of predicting only using the top layer, our learning loss function is a weighted average over the predictions of all layers, assigning higher weight to top layers. Our experiments demonstrate that this significantly improves the intermediate layer predictions while preserving the full model’s performance. In one model variant, we also include a small early-exit classifier trained to classify if the local intermediate layer prediction is consistent with the top layer. We train this classifier in a second quick step where we freeze the rest of the model.

Once the model is trained, we need a method to allow early-exiting. First, we define a local confidence measure for capturing the model’s confidence in its intermediate prediction. We explore three confidence measures (described in the results section below): (1) softmax response, taking the maximum predicted probability out of the softmax distribution; (2) state propagation, the cosine distance between the current hidden representation and the one from the previous layer; and (3) early-exit classifier, the output of a classifier specifically trained for predicting local consistency. We find the softmax response to be statistically strong while being simple and fast to compute. The other two alternatives are lighter in floating point operations (FLOPS).

Another challenge is that the self-attention of each layer depends on hidden-states from previous words. If we exit early for some word predictions, these hidden-states might be missing. Instead, we attend back to the hidden state of the last computed layer.

Finally, we set up the local confidence threshold for exiting early. In the next section, we describe our controlled process for finding good threshold values. As a first step, we simplify this infinite search space by building on a useful observation: mistakes that are made at the beginning of the generation process are more detrimental since they can affect all of the following outputs. Therefore, we start with a higher (more conservative) threshold, and gradually reduce it with time. We use a negative exponent with user-defined temperature to control this decay rate. We find this allows better control over the performance-efficiency tradeoff (the obtained speedup per quality level).

Reliably Controlling the Quality of the Accelerated Model

Early exit decisions have to be local; they need to happen when predicting each word. In practice, however, the final output should be globally consistent or comparable to the original model. For example, if the original full model generated “the concert was wonderful and long”, one would accept CALM switching the order of the adjectives and outputting “the concert was long and wonderful”. However, at the local level, the word “wonderful” was replaced with “long”. Therefore, the two outputs are globally consistent, but include some local inconsistencies. We build on the Learn then Test (LTT) framework to connect local confidence-based decisions to globally consistent outputs.

In CALM, local per-timestep confidence thresholds for early exiting decisions are derived, via LTT calibration, from user-defined consistency constraints over the full output text. Red boxes indicate that CALM used most of the decoder’s layers for that specific prediction. Green boxes indicate that CALM saved time by using only a few Transformer layers. Full sentence shown in the last example of this post.

First, we define and formulate two types of consistency constraints from which to choose:

Textual consistency: We bound the expected textual distance between the outputs of CALM and the outputs of the full model. This doesn’t require any labeled data.
Risk consistency: We bound the expected increase in loss that we allow for CALM compared to the full model. This requires reference outputs against which to compare.

For each of these constraints, we can set the tolerance that we allow and calibrate the confidence threshold to allow early exits while reliably satisfying our defined constraint with an arbitrarily high probability.

CALM Saves Inference Time

We run experiments on three popular generation datasets: CNN/DM for summarization, WMT for machine translation, and SQuAD for question answering. We evaluate each of the three confidence measures (softmax response, state propagation and early-exit classifier) using an 8-layer encoder-decoder model. To evaluate global sequence-level performance, we use the standard Rouge-L, BLEU, and Token-F1 scores that measure distances against human-written references. We show that one can maintain full model performance while using only a third or half of the layers on average. CALM achieves this by dynamically distributing the compute effort across the prediction timesteps.

As an approximate upper bound, we also compute the predictions using a local oracle confidence measure, which enables exiting at the first layer that leads to the same prediction as the top one. On all three tasks, the oracle measure can preserve full model performance when using only 1.5 decoder layers on average. In contrast to CALM, a static baseline uses the same number of layers for all predictions, requiring 3 to 7 layers (depending on the dataset) to preserve its performance. This demonstrates why the dynamic allocation of compute effort is important. Only a small fraction of the predictions require most of the model’s complexity, while for others much less should suffice.

Performance per task against the average number of decoder layers used.

Finally, we also find that CALM enables practical speedups. When benchmarking on TPUs, we saved almost half of the compute time while maintaining the quality of the outputs.

Example of a generated news summary. The top cell presents the reference human-written summary. Below is the prediction of the full model (8 layers) followed by two different CALM output examples. The first CALM output is 2.9x faster and the second output is 3.6x faster than the full model, benchmarked on TPUs.

Conclusion

CALM allows faster text generation with LMs, without reducing the quality of the output text. This is achieved by dynamically modifying the amount of compute per generation timestep, allowing the model to exit the computational sequence early when confident enough.

As language models continue to grow in size, studying how to efficiently use them becomes crucial. CALM is orthogonal and can be combined with many efficiency related efforts, including model quantization, distillation, sparsity, effective partitioning, and distributed control flows.

Acknowledgements

It was an honor and privilege to work on this with Adam Fisch, Ionel Gog, Seungyeon Kim, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. We also thank Anselm Levskaya, Hyung Won Chung, Tao Wang, Paul Barham, Michael Isard, Orhan Firat, Carlos Riquelme, Aditya Menon, Zhifeng Chen, Sanjiv Kumar, and Jeff Dean for helpful discussions and feedback. Finally, we thank Tom Small for preparing the animation in this blog post.

Source: Google AI Blog

Conversation Summaries in Google Chat

Posted by Mohammad Saleh, Software Engineer, Google Research, Brain Team, and Yinan Wang, Software Engineer, Google Workspace

Information overload is a significant challenge for many organizations and individuals today. It can be overwhelming to keep up with incoming chat messages and documents that arrive at our inbox everyday. This has been exacerbated by the increase in virtual work and remains a challenge as many teams transition to a hybrid work environment with a mix of those working both virtually and in an office. One solution that can address information overload is summarization — for example, to help users improve their productivity and better manage so much information, we recently introduced auto-generated summaries in Google Docs.

Today, we are excited to introduce conversation summaries in Google Chat for messages in Spaces. When these summaries are available, a card with automatically generated summaries is shown as users enter Spaces with unread messages. The card includes a list of summaries for the different topics discussed in Spaces. This feature is enabled by our state-of-the-art abstractive summarization model, Pegasus, which generates useful and concise summaries for chat conversations, and is currently available to selected premium Google Workspace business customers.

Conversation summaries provide a helpful digest of conversations in Spaces, allowing users to quickly catch-up on unread messages and navigate to the most relevant threads.

Conversation Summarization Modeling

The goal of text summarization is to provide helpful and concise summaries for different types of text, such as documents, articles, or spoken conversations. A good summary covers the key points succinctly, and is fluent and grammatically correct. One approach to summarization is to extract key parts from the text and concatenate them together into a summary (i.e., extractive summarization). Another approach is to use natural language generation (NLG) techniques to summarize using novel words and phrases not necessarily present in the original text. This is referred to as abstractive summarization and is considered closer to how a person would generally summarize text. A main challenge with abstractive summarization, however, is that it sometimes struggles to generate accurate and grammatically correct summaries, especially in real world applications.

ForumSum Dataset

The majority of abstractive summarization datasets and research focuses on single-speaker text documents, like news and scientific articles, mainly due to the abundance of human-written summaries for such documents. On the other hand, datasets of human-written summaries for other types of text, like chat or multi-speaker conversations, are very limited.

To address this we created ForumSum, a diverse and high-quality conversation summarization dataset with human-written summaries. The conversations in the dataset are collected from a wide variety of public internet forums, and are cleaned up and filtered to ensure high quality and safe content (more details in the paper).

An example from the ForumSum dataset.

Each utterance in the conversation starts on a new line, contains an author name and a message text that is separated with a colon. Human annotators are then given detailed instructions to write a 1-3 sentence summary of the conversation. These instructions went through multiple iterations to ensure annotators wrote high quality summaries. We have collected summaries for over six thousand conversations, with an average of more than 6 speakers and 10 utterances per conversation. ForumSum provides quality training data for the conversation summarization problem: it has a variety of topics, number of speakers, and number of utterances commonly encountered in a chat application.

Conversation Summarization Model Design

As we have written previously, the Transformer is a popular model architecture for sequence-to-sequence tasks, like abstractive summarization, where the inputs are the document words and the outputs are the summary words. Pegasus combined transformers with self-supervised pre-training customized for abstractive summarization, making it a great model choice for conversation summarization. First, we fine-tune Pegasus on the ForumSum dataset where the input is the conversation words and the output is the summary words. Second, we use knowledge distillation to distill the Pegasus model into a hybrid architecture of a transformer encoder and a recurrent neural network (RNN) decoder. The resulting model has lower latency and memory footprint while maintaining similar quality as the Pegasus model.

Quality and User Experience

A good summary captures the essence of the conversation while being fluent and grammatically correct. Based on human evaluation and user feedback, we learned that the summarization model generates useful and accurate summaries most of the time. But occasionally the model generates low quality summaries. After looking into issues reported by users, we found that there are two main types of low quality summaries. The first one is misattribution, when the model confuses which person or entity said or performed a certain action. The second one is misrepresentation, when the model’s generated summary misrepresents or contradicts the chat conversation.

To address low quality summaries and improve the user experience, we have made progress in several areas:

Improving ForumSum: While ForumSum provides a good representation of chat conversations, we noticed certain patterns and language styles in Google Chat conversations that differ from ForumSum, e.g., how users mention other users and the use of abbreviations and special symbols. After exploring examples reported by users, we concluded that these out-of-distribution language patterns contributed to low quality summaries. To address this, we first performed data formatting and clean-ups to reduce mismatches between chat and ForumSum conversations whenever possible. Second, we added more training data to ForumSum to better represent these style mismatches. Collectively, these changes resulted in reduction of low quality summaries.
Controlled triggering: To make sure summaries bring the most value to our users, we first need to make sure that the chat conversation is worthy of summarization. For example, we found that there is less value in generating a summary when the user is actively engaged in a conversation and does not have many unread messages, or when the conversation is too short.
Detecting low quality summaries: While the two methods above limited low quality and low value summaries, we still developed methods to detect and abstain from showing such summaries to the user when they are generated. These are a set of heuristics and models to measure the overall quality of summaries and whether they suffer from misattribution or misrepresentation issues.

Finally, while the hybrid model provided significant performance improvements, the latency to generate summaries was still noticeable to users when they opened Spaces with unread messages. To address this issue, we instead generate and update summaries whenever there is a new message sent, edited or deleted. Then summaries are cached ephemerally to ensure they surface smoothly when users open Spaces with unread messages.

Conclusion and Future Work

We are excited to apply state-of-the-art abstractive summarization models to help our Workspace users improve their productivity in Spaces. While this is great progress, we believe there are many opportunities to further improve the experience and the overall quality of summaries. Future directions we are exploring include better modeling and summarizing entangled conversations that include multiple topics, and developing metrics that better measure the factual consistency between chat conversations and summaries.

Acknowledgements

The authors would like to thank the many people across Google that contributed to this work: Ahmed Chowdhury, Alejandro Elizondo, Anmol Tukrel, Benjamin Lee, Chao Wang, Chris Carroll, Don Kim, Jackie Tsay, Jennifer Chou, Jesse Sliter, John Sipple, Kate Montgomery, Maalika Manoharan, Mahdis Mahdieh, Mia Chen, Misha Khalman, Peter Liu, Robert Diersing, Sarah Read, Winnie Yeung, Yao Zhao, and Yonghui Wu.