João Sedoc

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Lining Zhang, Simon Mille, Yufang Hou et al. · deepmind, uw

To prevent the costly and inefficient use of resources on low-quality annotations, we want a method for creating a pool of dependable annotators who can effectively complete difficult tasks, such as evaluating automatic summarization. Thus, we investigate the recruitment of high-quality Amazon Mechanical Turk workers via a two-step pipeline. We show that we can successfully filter out subpar workers before they carry out the evaluations and obtain high-agreement annotations with similar constraints on resources. Although our workers demonstrate a strong consensus among themselves and CloudResearch workers, their alignment with expert judgments on a subset of the data is not as expected and needs further training in correctness. This paper still serves as a best practice for the recruitment of qualified annotators in other challenging annotation tasks.

Jeevesh Juneja, Rachit Bansal, Kyunghyun Cho et al. · cmu, harvard

It is widely accepted in the mode connectivity literature that when two neural networks are trained similarly on the same data, they are connected by a path through parameter space over which test set accuracy is maintained. Under some circumstances, including transfer learning from pretrained models, these paths are presumed to be linear. In contrast to existing results, we find that among text classifiers (trained on MNLI, QQP, and CoLA), some pairs of finetuned models have large barriers of increasing loss on the linear paths between them. On each task, we find distinct clusters of models which are linearly connected on the test loss surface, but are disconnected from models outside the cluster -- models that occupy separate basins on the surface. By measuring performance on specially-crafted diagnostic datasets, we find that these clusters correspond to different generalization strategies: one cluster behaves like a bag of words model under domain shift, while another cluster uses syntactic heuristics. Our work demonstrates how the geometry of the loss surface can guide models towards different heuristic functions.

Damilola Omitaomu, Shabnam Tafreshi, Tingting Liu et al.

Empathy is a cognitive and emotional reaction to an observed situation of others. Empathy has recently attracted interest because it has numerous applications in psychology and AI, but it is unclear how different forms of empathy (e.g., self-report vs counterpart other-report, concern vs. distress) interact with other affective phenomena or demographics like gender and age. To better understand this, we created the {\it Empathic Conversations} dataset of annotated negative, empathy-eliciting dialogues in which pairs of participants converse about news articles. People differ in their perception of the empathy of others. These differences are associated with certain characteristics such as personality and demographics. Hence, we collected detailed characterization of the participants' traits, their self-reported empathetic response to news articles, their conversational partner other-report, and turn-by-turn third-party assessments of the level of self-disclosure, emotion, and empathy expressed. This dataset is the first to present empathy in multiple forms along with personal distress, emotion, personality characteristics, and person-level demographic information. We present baseline models for predicting some of these features from conversations.

Shai Gretz, Assaf Toledo, Roni Friedman et al. · ibm-research

The COVID-19 pandemic has made a huge global impact and cost millions of lives. As COVID-19 vaccines were rolled out, they were quickly met with widespread hesitancy. To address the concerns of hesitant people, we launched VIRA, a public dialogue system aimed at addressing questions and concerns surrounding the COVID-19 vaccines. Here, we release VIRADialogs, a dataset of over 8k dialogues conducted by actual users with VIRA, providing a unique real-world conversational dataset. In light of rapid changes in users' intents, due to updates in guidelines or in response to new information, we highlight the important task of intent discovery in this use-case. We introduce a novel automatic evaluation framework for intent discovery, leveraging the existing intent classifier of VIRA. We use this framework to report baseline intent discovery results over VIRADialogs, that highlight the difficulty of this task.

Yukun Feng, Patrick Xia, Benjamin Van Durme et al.

Building pretrained language models is considered expensive and data-intensive, but must we increase dataset size to achieve better performance? We propose an alternative to larger training sets by automatically identifying smaller yet domain-representative subsets. We extend Cynical Data Selection, a statistical sentence scoring method that conditions on a representative target domain corpus. As an example, we treat the OntoNotes corpus as a target domain and pretrain a RoBERTa-like encoder from a cynically selected subset of the Pile. On both perplexity and across several downstream tasks in the target domain, it consistently outperforms random selection with 20x less data, 3x fewer training iterations, and 2x less estimated cloud compute cost, validating the recipe of automatic document selection for LM pretraining.

Mario Rodríguez-Cantelar, Chen Zhang, Chengguang Tang et al.

The advent and fast development of neural networks have revolutionized the research on dialogue systems and subsequently have triggered various challenges regarding their automatic evaluation. Automatic evaluation of open-domain dialogue systems as an open challenge has been the center of the attention of many researchers. Despite the consistent efforts to improve automatic metrics' correlations with human evaluation, there have been very few attempts to assess their robustness over multiple domains and dimensions. Also, their focus is mainly on the English language. All of these challenges prompt the development of automatic evaluation metrics that are reliable in various domains, dimensions, and languages. This track in the 11th Dialogue System Technology Challenge (DSTC11) is part of the ongoing effort to promote robust and multilingual automatic evaluation metrics. This article describes the datasets and baselines provided to participants and discusses the submission and result details of the two proposed subtasks.

Young Min Cho, Sunny Rai, Lyle Ungar et al.

Mental health conversational agents (a.k.a. chatbots) are widely studied for their potential to offer accessible support to those experiencing mental health challenges. Previous surveys on the topic primarily consider papers published in either computer science or medicine, leading to a divide in understanding and hindering the sharing of beneficial knowledge between both domains. To bridge this gap, we conduct a comprehensive literature review using the PRISMA framework, reviewing 534 papers published in both computer science and medicine. Our systematic review reveals 136 key papers on building mental health-related conversational agents with diverse characteristics of modeling and experimental design techniques. We find that computer science papers focus on LLM techniques and evaluating response quality using automated metrics with little attention to the application while medical papers use rule-based conversational agents and outcome metrics to measure the health outcomes of participants. Based on our findings on transparency, ethics, and cultural heterogeneity in this review, we provide a few recommendations to help bridge the disciplinary divide and enable the cross-disciplinary development of mental health conversational agents.

Nikita Soni, H. Andrew Schwartz, João Sedoc et al.

As research in human-centered NLP advances, there is a growing recognition of the importance of incorporating human and social factors into NLP models. At the same time, our NLP systems have become heavily reliant on LLMs, most of which do not model authors. To build NLP systems that can truly understand human language, we must better integrate human contexts into LLMs. This brings to the fore a range of design considerations and challenges in terms of what human aspects to capture, how to represent them, and what modeling strategies to pursue. To address these, we advocate for three positions toward creating large human language models (LHLMs) using concepts from psychological and behavioral sciences: First, LM training should include the human context. Second, LHLMs should recognize that people are more than their group(s). Third, LHLMs should be able to account for the dynamic and temporally-dependent nature of the human context. We refer to relevant advances and present open challenges that need to be addressed and their possible solutions in realizing these goals.

Roni Friedman, João Sedoc, Shai Gretz et al.

Public trust in medical information is crucial for successful application of public health policies such as vaccine uptake. This is especially true when the information is offered remotely, by chatbots, which have become increasingly popular in recent years. Here, we explore the challenging task of human-bot turn-level trust classification. We rely on a recently released data of observationally-collected (rather than crowdsourced) dialogs with VIRA chatbot, a COVID-19 Vaccine Information Resource Assistant. These dialogs are centered around questions and concerns about COVID-19 vaccines, where trust is particularly acute. We annotated $3k$ VIRA system-user conversational turns for Low Institutional Trust or Low Agent Trust vs. Neutral or High Trust. We release the labeled dataset, VIRATrustData, the first of its kind to the best of our knowledge. We demonstrate how this task is non-trivial and compare several models that predict the different levels of trust.

Li S. Yifei, Lyle Ungar, João Sedoc

Pre-trained large language models (LLMs) reflect the inherent social biases of their training corpus. Many methods have been proposed to mitigate this issue, but they often fail to debias or they sacrifice model accuracy. We use conceptors--a soft projection method--to identify and remove the bias subspace in LLMs such as BERT and GPT. We propose two methods of applying conceptors (1) bias subspace projection by post-processing by the conceptor NOT operation; and (2) a new architecture, conceptor-intervened BERT (CI-BERT), which explicitly incorporates the conceptor projection into all layers during training. We find that conceptor post-processing achieves state-of-the-art (SoTA) debiasing results while maintaining LLMs' performance on the GLUE benchmark. Further, it is robust in various scenarios and can mitigate intersectional bias efficiently by its AND operation on the existing bias subspaces. Although CI-BERT's training takes all layers' bias into account and can beat its post-processing counterpart in bias mitigation, CI-BERT reduces the language model accuracy. We also show the importance of carefully constructing the bias subspace. The best results are obtained by removing outliers from the list of biased words, combining them (via the OR operation), and computing their embeddings using the sentences from a cleaner corpus.

João Sedoc, Baotong Zhang, Dean Foster

Reliably knowing when a language model is correct is almost as important as being correct. We introduce prover-verifier deliberation (PVD), an inference-time protocol grounded in interactive proof theory, as a mechanism for selective prediction: the protocol produces both an answer and a structured confidence verdict, allowing a system to report high-confidence answers while abstaining on uncertain cases. In each dialogue, a prover defends a candidate answer through checkable sub-claims while a verifier issues targeted challenges and returns \textsc{Accept}, \textsc{Challenge}, or \textsc{Reject}. Because frozen language models are imperfect provers and verifiers operating over a noisy channel, formal soundness and completeness guarantees do not transfer; instead, we characterize the protocol empirically through its coverage-precision behavior. Our main experiment uses Claude Sonnet 4.6 as prover and Claude Haiku 4.5 as verifier on GPQA Diamond. Questions accepted with no answer revision, which we call Accept + No Change (ANC), are reported as the high-confidence subset; we evaluate this subset by its precision and coverage. ANC separates reliable from unreliable answers, yielding a $\sim$30pp HC-Prec gap over the non-ANC complement. Robustness experiments with GPT and Gemini pairings show that high HC-Prec can transfer across model families, while verifier strictness and domain competence largely determine the size of the selection gap. On Humanity's Last Exam, weaker prover-verifier pairings can collapse or invert the ANC signal, illustrating a practical failure mode when the verifier operates outside its effective region. Comparisons with self-consistency, universal self-consistency, multi-agent debate, and Reflexion suggest that prover-verifier deliberation supplies a distinct argument-defensibility signal for selective prediction.

Andrew Smart, Ben Hutchinson, Lameck Mbangula Amugongo et al.

Large Language Models (LLMs) have rapidly increased in size and apparent capabilities in the last three years, but their training data is largely English text. There is growing interest in multilingual LLMs, and various efforts are striving for models to accommodate languages of communities outside of the Global North, which include many languages that have been historically underrepresented in digital realms. These languages have been coined as "low resource languages" or "long-tail languages", and LLMs performance on these languages is generally poor. While expanding the use of LLMs to more languages may bring many potential benefits, such as assisting cross-community communication and language preservation, great care must be taken to ensure that data collection on these languages is not extractive and that it does not reproduce exploitative practices of the past. Collecting data from languages spoken by previously colonized people, indigenous people, and non-Western languages raises many complex sociopolitical and ethical questions, e.g., around consent, cultural safety, and data sovereignty. Furthermore, linguistic complexity and cultural nuances are often lost in LLMs. This position paper builds on recent scholarship, and our own work, and outlines several relevant social, cultural, and ethical considerations and potential ways to mitigate them through qualitative research, community partnerships, and participatory design approaches. We provide twelve recommendations for consideration when collecting language data on underrepresented language communities outside of the Global North.

Charles Welch, Allison Lahnala, Vasudha Varadarajan et al.

We present a low-compute non-generative system for implementing interview-style conversational agents which can be used to facilitate qualitative data collection through controlled interactions and quantitative analysis. Use cases include applications to tracking attitude formation or behavior change, where control or standardization over the conversational flow is desired. We show how our system can be easily adjusted through an online administrative panel to create new interviews, making the tool accessible without coding. Two case studies are presented as example applications, one regarding the Expressive Interviewing system for COVID-19 and the other a semi-structured interview to survey public opinion on emerging neurotechnology. Our code is open-source, allowing others to build off of our work and develop extensions for additional functionality.

Ilias Triantafyllopoulos, Young-Min Cho, Ren Tao et al.

Activation-based steering provides control of LLM behavior at inference time, but the dominant paradigm reduces each concept to a single direction whose geometry is left largely unexamined. Rather than selecting a single steering direction, we use conceptors: soft projection matrices estimated from activations pooled across both poles of a bipolar concept, which preserve the concept's full multidimensional subspace. A geometric analysis shows the bipolar subspace strictly subsumes the single-vector baseline. We further show that the conceptor quota provides a parameter-free layer-selection diagnostic, predicting concept separability with Pearson correlations up to r=0.96 across three instruction-tuned models and three semantic dimensions. Beyond selection, conceptors admit a closed-form Boolean algebra (AND, OR, NOT): we evaluate conceptor compositionality on thematically related sub-concepts. Across a systematic five-axis design-space evaluation, conceptors match or outperform additive baselines at layers where concept subspaces are multi-dimensional while producing substantially fewer degenerate outputs. Conceptor steering is a geometrically principled, compositional, and practically safer alternative to single-direction steering from a limited number of contrastive pairs.

Tingting Liu, Salvatore Giorgi, Ankit Aich et al.

As AI chatbots increasingly incorporate empathy, understanding user-centered perceptions of chatbot empathy and its impact on conversation quality remains essential yet under-explored. This study examines how chatbot identity and perceived empathy influence users' overall conversation experience. Analyzing 155 conversations from two datasets, we found that while GPT-based chatbots were rated significantly higher in conversational quality, they were consistently perceived as less empathetic than human conversational partners. Empathy ratings from GPT-4o annotations aligned with user ratings, reinforcing the perception of lower empathy in chatbots compared to humans. Our findings underscore the critical role of perceived empathy in shaping conversation quality, revealing that achieving high-quality human-AI interactions requires more than simply embedding empathetic language; it necessitates addressing the nuanced ways users interpret and experience empathy in conversations with chatbots.

Marcel Nawrath, Agnieszka Nowak, Tristan Ratz et al. · amazon-science

At the heart of the Pyramid evaluation method for text summarization lie human written summary content units (SCUs). These SCUs are concise sentences that decompose a summary into small facts. Such SCUs can be used to judge the quality of a candidate summary, possibly partially automated via natural language inference (NLI) systems. Interestingly, with the aim to fully automate the Pyramid evaluation, Zhang and Bansal (2021) show that SCUs can be approximated by automatically generated semantic role triplets (STUs). However, several questions currently lack answers, in particular: i) Are there other ways of approximating SCUs that can offer advantages? ii) Under which conditions are SCUs (or their approximations) offering the most value? In this work, we examine two novel strategies to approximate SCUs: generating SCU approximations from AMR meaning representations (SMUs) and from large language models (SGUs), respectively. We find that while STUs and SMUs are competitive, the best approximation quality is achieved by SGUs. We also show through a simple sentence-decomposition baseline (SSUs) that SCUs (and their approximations) offer the most value when ranking short summaries, but may not help as much when ranking systems or longer summaries.

Rubing Li, João Sedoc, Arun Sundararajan

When encountering increasingly frequent performance improvements or cost reductions from a new large language model (LLM), developers of applications leveraging LLMs must decide whether to take advantage of these improvements or stay with older tried-and-tested models. Low perceived switching frictions can lead to choices that do not consider more subtle behavior changes that the transition may induce. Our experiments use a popular game-theoretic behavioral economics model of trust to show stark differences in the trusting behavior of OpenAI's and DeepSeek's models. We highlight a collapse in the economic trust behavior of the o1-mini and o3-mini models as they reconcile profit-maximizing and risk-seeking with future returns from trust, and contrast it with DeepSeek's more sophisticated and profitable trusting behavior that stems from an ability to incorporate deeper concepts like forward planning and theory-of-mind. As LLMs form the basis for high-stakes commercial systems, our results highlight the perils of relying on LLM performance benchmarks that are too narrowly defined and suggest that careful analysis of their hidden fault lines should be part of any organization's AI strategy.

John Mendonça, Lining Zhang, Rahul Mallidi et al.

The rapid advancement of Large Language Models (LLMs) has intensified the need for robust dialogue system evaluation, yet comprehensive assessment remains challenging. Traditional metrics often prove insufficient, and safety considerations are frequently narrowly defined or culturally biased. The DSTC12 Track 1, "Dialog System Evaluation: Dimensionality, Language, Culture and Safety," is part of the ongoing effort to address these critical gaps. The track comprised two subtasks: (1) Dialogue-level, Multi-dimensional Automatic Evaluation Metrics, and (2) Multilingual and Multicultural Safety Detection. For Task 1, focused on 10 dialogue dimensions, a Llama-3-8B baseline achieved the highest average Spearman's correlation (0.1681), indicating substantial room for improvement. In Task 2, while participating teams significantly outperformed a Llama-Guard-3-1B baseline on the multilingual safety subset (top ROC-AUC 0.9648), the baseline proved superior on the cultural subset (0.5126 ROC-AUC), highlighting critical needs in culturally-aware safety. This paper describes the datasets and baselines provided to participants, as well as submission evaluation results for each of the two proposed subtasks.

Ilias Triantafyllopoulos, Renyi Qu, Salvatore Giorgi et al.

Ensuring safety and in-domain responses for Retrieval-Augmented Generation (RAG) systems is paramount in safety-critical applications, yet remains a significant challenge. To address this, we evaluate four methodologies for Out-Of-Domain (OOD) query detection: GPT-4o, regression-based, Principal Component Analysis (PCA)-based, and Neural Collapse (NC), to ensure the RAG system only responds to queries confined to the system's knowledge base. Specifically, our evaluation explores two novel dimensionality reduction and feature separation strategies: \textit{PCA}, where top components are selected using explained variance or OOD separability, and an adaptation of \textit{Neural Collapse Feature Separation}. We validate our approach on standard datasets (StackExchange and MSMARCO) and real-world applications (Substance Use and COVID-19), including tests against LLM-simulated and actual attacks on a COVID-19 vaccine chatbot. Through human and LLM-based evaluations of response correctness and relevance, we confirm that an external OOD detector is crucial for maintaining response relevance.

Vasant Dhar, João Sedoc

Long-term investing was previously seen as requiring human judgment. With the advent of generative artificial intelligence (AI) systems, automated systematic long-term investing is now feasible. In this paper, we present DBOT, a system whose goal is to reason about valuation like Aswath Damodaran, who is a unique expert in the investment arena in terms of having published thousands of valuations on companies in addition to his numerous writings on the topic, which provide ready training data for an AI system. DBOT can value any publicly traded company. DBOT can also be back-tested, making its behavior and performance amenable to scientific inquiry. We compare DBOT to its analytic parent, Damodaran, and highlight the research challenges involved in raising its current capability to that of Damodaran's. Finally, we examine the implications of DBOT-like AI agents for the financial industry, especially how they will impact the role of human analysts in valuation.

Xiang Cheng, Raveesh Mayya, João Sedoc

Unstructured text data annotation is foundational to management research and Large Language Models (LLMs) promise a cost-effective and scalable alternative to human annotation. The validity of insights drawn from LLM annotated data critically depends on minimizing the discrepancy between LLM assigned labels and the unobserved ground truth, as well as ensuring long-term reproducibility of results. We address the gap in the literature on LLM annotation by decomposing measurement error in LLM-based text annotation into four distinct sources: (1) guideline-induced error from inconsistent annotation criteria, (2) baseline-induced error from unreliable human reference standards, (3) prompt-induced error from suboptimal meta-instruction formatting, and (4) model-induced error from architectural differences across LLMs. We develop the SILICON methodology to systematically reduce measurement error from LLM annotation in all four sources above. Empirical validation across seven management research cases shows iteratively refined guidelines substantially increases the LLM-human agreement compared to one-shot guidelines; expert-generated baselines exhibit higher inter-annotator agreement as well as are less prone to producing misleading LLM-human agreement estimates compared to crowdsourced baselines; placing content in the system prompt reduces prompt-induced error; and model performance varies substantially across tasks. To further reduce error, we introduce a cost-effective multi-LLM labeling method, where only low-confidence items receive additional labels from alternative models. Finally, in addressing closed source model retirement cycles, we introduce an intuitive regression-based methodology to establish robust reproducibility protocols. Our evidence indicates that reducing each error source is necessary, and that SILICON supports reproducible, rigorous annotation in management research.

Salvatore Giorgi, Tingting Liu, Ankit Aich et al.

Large language models (LLMs) are increasingly being used in human-centered social scientific tasks, such as data annotation, synthetic data creation, and engaging in dialog. However, these tasks are highly subjective and dependent on human factors, such as one's environment, attitudes, beliefs, and lived experiences. Thus, it may be the case that employing LLMs (which do not have such human factors) in these tasks results in a lack of variation in data, failing to reflect the diversity of human experiences. In this paper, we examine the role of prompting LLMs with human-like personas and asking the models to answer as if they were a specific human. This is done explicitly, with exact demographics, political beliefs, and lived experiences, or implicitly via names prevalent in specific populations. The LLM personas are then evaluated via (1) subjective annotation task (e.g., detecting toxicity) and (2) a belief generation task, where both tasks are known to vary across human factors. We examine the impact of explicit vs. implicit personas and investigate which human factors LLMs recognize and respond to. Results show that explicit LLM personas show mixed results when reproducing known human biases, but generally fail to demonstrate implicit biases. We conclude that LLMs may capture the statistical patterns of how people speak, but are generally unable to model the complex interactions and subtleties of human perceptions, potentially limiting their effectiveness in social science applications.

Aadesh Salecha, Molly E. Ireland, Shashanka Subrahmanya et al.

As Large Language Models (LLMs) become widely used to model and simulate human behavior, understanding their biases becomes critical. We developed an experimental framework using Big Five personality surveys and uncovered a previously undetected social desirability bias in a wide range of LLMs. By systematically varying the number of questions LLMs were exposed to, we demonstrate their ability to infer when they are being evaluated. When personality evaluation is inferred, LLMs skew their scores towards the desirable ends of trait dimensions (i.e., increased extraversion, decreased neuroticism, etc). This bias exists in all tested models, including GPT-4/3.5, Claude 3, Llama 3, and PaLM-2. Bias levels appear to increase in more recent models, with GPT-4's survey responses changing by 1.20 (human) standard deviations and Llama 3's by 0.98 standard deviations-very large effects. This bias is robust to randomization of question order and paraphrasing. Reverse-coding all the questions decreases bias levels but does not eliminate them, suggesting that this effect cannot be attributed to acquiescence bias. Our findings reveal an emergent social desirability bias and suggest constraints on profiling LLMs with psychometric tests and on using LLMs as proxies for human participants.

Huda Khayrallah, Zuhaib Akhtar, Edward Cohen et al.

We release MMSMR, a Massively Multi-System MultiReference dataset to enable future work on metrics and evaluation for dialog. Automatic metrics for dialogue evaluation should be robust proxies for human judgments; however, the verification of robustness is currently far from satisfactory. To quantify the robustness correlation and understand what is necessary in a test set, we create and release an 8-reference dialog dataset by extending single-reference evaluation sets and introduce this new language learning conversation dataset. We then train 1750 systems and evaluate them on our novel test set and the DailyDialog dataset. We release the novel test set, and model hyper parameters, inference outputs, and metric scores for each system on a variety of datasets.

Prasanna Parasurama, João Sedoc

We investigate whether it is feasible to remove gendered information from resumes to mitigate potential bias in algorithmic resume screening. Using a corpus of 709k resumes from IT firms, we first train a series of models to classify the self-reported gender of the applicant, thereby measuring the extent and nature of gendered information encoded in resumes. We then conduct a series of gender obfuscation experiments, where we iteratively remove gendered information from resumes. Finally, we train a resume screening algorithm and investigate the trade-off between gender obfuscation and screening algorithm performance. Results show: (1) There is a significant amount of gendered information in resumes. (2) Lexicon-based gender obfuscation method (i.e. removing tokens that are predictive of gender) can reduce the amount of gendered information to a large extent. However, after a certain point, the performance of the resume screening algorithm starts suffering. (3) General-purpose gender debiasing methods for NLP models such as removing gender subspace from embeddings are not effective in obfuscating gender.

Qi He, João Sedoc, Jordan Rodu

Transformer networks are the de facto standard architecture in natural language processing. To date, there are no theoretical analyses of the Transformer's ability to capture tree structures. We focus on the ability of Transformer networks to learn tree structures that are important for tree transduction problems. We first analyze the theoretical capability of the standard Transformer architecture to learn tree structures given enumeration of all possible tree backbones, which we define as trees without labels. We then prove that two linear layers with ReLU activation function can recover any tree backbone from any two nonzero, linearly independent starting backbones. This implies that a Transformer can learn tree structures well in theory. We conduct experiments with synthetic data and find that the standard Transformer achieves similar accuracy compared to a Transformer where tree position information is explicitly encoded, albeit with slower convergence. This confirms empirically that Transformers can learn tree structures.

Chen Zhang, João Sedoc, Luis Fernando D'Haro et al.

The development of Open-Domain Dialogue Systems (ODS)is a trending topic due to the large number of research challenges, large societal and business impact, and advances in the underlying technology. However, the development of these kinds of systems requires two important characteristics:1) automatic evaluation mechanisms that show high correlations with human judgements across multiple dialogue evaluation aspects (with explainable features for providing constructive and explicit feedback on the quality of generative models' responses for quick development and deployment)and 2) mechanisms that can help to control chatbot responses,while avoiding toxicity and employing intelligent ways to handle toxic user comments and keeping interaction flow and engagement. This track at the 10th Dialogue System Technology Challenge (DSTC10) is part of the ongoing effort to promote scalable and toxic-free ODS. This paper describes the datasets and baselines provided to participants, as well as submission evaluation results for each of the two proposed subtasks.

Seolhwa Lee, Kisu Yang, Chanjun Park et al.

Abstractive dialogue summarization is a challenging task for several reasons. First, most of the important pieces of information in a conversation are scattered across utterances through multi-party interactions with different textual styles. Second, dialogues are often informal structures, wherein different individuals express personal perspectives, unlike text summarization, tasks that usually target formal documents such as news articles. To address these issues, we focused on the association between utterances from individual speakers and unique syntactic structures. Speakers have unique textual styles that can contain linguistic information, such as voiceprint. Therefore, we constructed a syntax-aware model by leveraging linguistic information (i.e., POS tagging), which alleviates the above issues by inherently distinguishing sentences uttered from individual speakers. We employed multi-task learning of both syntax-aware information and dialogue summarization. To the best of our knowledge, our approach is the first method to apply multi-task learning to the dialogue summarization task. Experiments on a SAMSum corpus (a large-scale dialogue summarization corpus) demonstrated that our method improved upon the vanilla model. We further analyze the costs and benefits of our approach relative to baseline models.

Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal et al.

We introduce GEM, a living benchmark for natural language Generation (NLG), its Evaluation, and Metrics. Measuring progress in NLG relies on a constantly evolving ecosystem of automated metrics, datasets, and human evaluation standards. Due to this moving target, new models often still evaluate on divergent anglo-centric corpora with well-established, but flawed, metrics. This disconnect makes it challenging to identify the limitations of current models and opportunities for progress. Addressing this limitation, GEM provides an environment in which models can easily be applied to a wide set of tasks and in which evaluation strategies can be tested. Regular updates to the benchmark will help NLG research become more multilingual and evolve the challenge alongside models. This paper serves as the description of the data for which we are organizing a shared task at our ACL 2021 Workshop and to which we invite the entire NLG community to participate.

Huda Khayrallah, João Sedoc

Non-task-oriented dialog models suffer from poor quality and non-diverse responses. To overcome limited conversational data, we apply Simulated Multiple Reference Training (SMRT; Khayrallah et al., 2020), and use a paraphraser to simulate multiple responses per training prompt. We find SMRT improves over a strong Transformer baseline as measured by human and automatic quality scores and lexical diversity. We also find SMRT is comparable to pretraining in human evaluation quality, and outperforms pretraining on automatic quality and lexical diversity, without requiring related-domain dialog data.

Huda Khayrallah, João Sedoc

We consider the intrinsic evaluation of neural generative dialog models through the lens of Grice's Maxims of Conversation (1975). Based on the maxim of Quantity (be informative), we propose Relative Utterance Quantity (RUQ) to diagnose the `I don't know' problem, in which a dialog system produces generic responses. The linguistically motivated RUQ diagnostic compares the model score of a generic response to that of the reference response. We find that for reasonable baseline models, `I don't know' is preferred over the reference the majority of the time, but this can be reduced to less than 5% with hyperparameter tuning. RUQ allows for the direct analysis of the `I don't know' problem, which has been addressed but not analyzed by prior work.

Seolhwa Lee, Heuiseok Lim, João Sedoc

There is a multitude of novel generative models for open-domain conversational systems; however, there is no systematic evaluation of different systems. Systematic comparisons require consistency in experimental design, evaluation sets, conversational systems and their outputs, and statistical analysis. We lay out a protocol for the evaluation of conversational models using head-to-head pairwise comparison. We analyze ten recent models that claim state-of-the-art performance using a paired head-to-head performance (win-loss-tie) on five evaluation datasets. Our findings show that DialoGPT and Blender are superior systems using Bradley-Terry model and TrueSkill ranking methods. These findings demonstrate the feasibility of our protocol to evaluate conversational agents and evaluation sets. Finally, we make all code and evaluations publicly available for researchers to compare their model to other state-of-the-art dialog models.

Alexandra DeLucia, Aaron Mueller, Xiang Lisa Li et al.

Narrative generation is an open-ended NLP task in which a model generates a story given a prompt. The task is similar to neural response generation for chatbots; however, innovations in response generation are often not applied to narrative generation, despite the similarity between these tasks. We aim to bridge this gap by applying and evaluating advances in decoding methods for neural response generation to neural narrative generation. In particular, we employ GPT-2 and perform ablations across nucleus sampling thresholds and diverse decoding hyperparameters -- specifically, maximum mutual information -- analyzing results over multiple criteria with automatic and human evaluation. We find that (1) nucleus sampling is generally best with thresholds between 0.7 and 0.9; (2) a maximum mutual information objective can improve the quality of generated stories; and (3) established automatic metrics do not correlate well with human judgments of narrative quality on any qualitative metric.

Nathaniel Weir, João Sedoc, Benjamin Van Durme

We present COD3S, a novel method for generating semantically diverse sentences using neural sequence-to-sequence (seq2seq) models. Conditioned on an input, seq2seq models typically produce semantically and syntactically homogeneous sets of sentences and thus perform poorly on one-to-many sequence generation tasks. Our two-stage approach improves output diversity by conditioning generation on locality-sensitive hash (LSH)-based semantic sentence codes whose Hamming distances highly correlate with human judgments of semantic textual similarity. Though it is generally applicable, we apply COD3S to causal generation, the task of predicting a proposition's plausible causes or effects. We demonstrate through automatic and human evaluation that responses produced using our method exhibit improved diversity without degrading task performance.

Patrick Xia, João Sedoc, Benjamin Van Durme

We investigate modeling coreference resolution under a fixed memory constraint by extending an incremental clustering algorithm to utilize contextualized encoders and neural components. Given a new sentence, our end-to-end algorithm proposes and scores each mention span against explicit entity representations created from the earlier document context (if any). These spans are then used to update the entity's representations before being forgotten; we only retain a fixed set of salient entities throughout the document. In this work, we successfully convert a high-performing model (Joshi et al., 2020), asymptotically reducing its memory usage to constant space with only a 0.3% relative loss in F1 on OntoNotes 5.0.

João Sedoc, Sven Buechel, Yehonathan Nachmany et al.

Despite the excellent performance of black box approaches to modeling sentiment and emotion, lexica (sets of informative words and associated weights) that characterize different emotions are indispensable to the NLP community because they allow for interpretable and robust predictions. Emotion analysis of text is increasing in popularity in NLP; however, manually creating lexica for psychological constructs such as empathy has proven difficult. This paper automatically creates empathy word ratings from document-level ratings. The underlying problem of learning word ratings from higher-level supervision has to date only been addressed in an ad hoc fashion and has not used deep learning methods. We systematically compare a number of approaches to learning word ratings from higher-level supervision against a Mixed-Level Feed Forward Network (MLFFN), which we find performs best, and use the MLFFN to create the first-ever empathy lexicon. We then use Signed Spectral Clustering to gain insights into the resulting words. The empathy and distress lexica are publicly available at: http://www.wwbp.org/lexica.html.

Daphne Ippolito, Reno Kriz, Maria Kustikova et al.

While conditional language models have greatly improved in their ability to output high-quality natural language, many NLP applications benefit from being able to generate a diverse set of candidate sequences. Diverse decoding strategies aim to, within a given-sized candidate list, cover as much of the space of high-quality outputs as possible, leading to improvements for tasks that re-rank and combine candidate outputs. Standard decoding methods, such as beam search, optimize for generating high likelihood sequences rather than diverse ones, though recent work has focused on increasing diversity in these methods. In this work, we perform an extensive survey of decoding-time strategies for generating diverse outputs from conditional language models. We also show how diversity can be improved without sacrificing quality by over-sampling additional candidates, then filtering to the desired number.

Saket Karve, Lyle Ungar, João Sedoc

Bias in word embeddings such as Word2Vec has been widely investigated, and many efforts made to remove such bias. We show how to use conceptors debiasing to post-process both traditional and contextualized word embeddings. Our conceptor debiasing can simultaneously remove racial and gender biases and, unlike standard debiasing methods, can make effect use of heterogeneous lists of biased words. We show that conceptor debiasing diminishes racial and gender bias of word representations as measured using the Word Embedding Association Test (WEAT) of Caliskan et al. (2017).

Tianlin Liu, Lyle Ungar, João Sedoc

Distributed representations of sentences have become ubiquitous in natural language processing tasks. In this paper, we consider a continual learning scenario for sentence representations: Given a sequence of corpora, we aim to optimize the sentence encoder with respect to the new corpus while maintaining its accuracy on the old corpora. To address this problem, we propose to initialize sentence encoders with the help of corpus-independent features, and then sequentially update sentence encoders using Boolean operations of conceptor matrices to learn corpus-dependent features. We evaluate our approach on semantic textual similarity tasks and show that our proposed sentence encoder can continually learn features from new corpora while retaining its competence on previously encountered corpora.

Reno Kriz, João Sedoc, Marianna Apidianaki et al.

Sentence simplification is the task of rewriting texts so they are easier to understand. Recent research has applied sequence-to-sequence (Seq2Seq) models to this task, focusing largely on training-time improvements via reinforcement learning and memory augmentation. One of the main problems with applying generic Seq2Seq models for simplification is that these models tend to copy directly from the original sentence, resulting in outputs that are relatively long and complex. We aim to alleviate this issue through the use of two main techniques. First, we incorporate content word complexities, as predicted with a leveled word complexity model, into our loss function during training. Second, we generate a large set of diverse candidate simplifications at test time, and rerank these to promote fluency, adequacy, and simplicity. Here, we measure simplicity through a novel sentence complexity model. These extensions allow our models to perform competitively with state-of-the-art systems while generating simpler sentences. We report standard automatic and human evaluation metrics.

Tianlin Liu, João Sedoc, Lyle Ungar

Distributed representations of words, better known as word embeddings, have become important building blocks for natural language processing tasks. Numerous studies are devoted to transferring the success of unsupervised word embeddings to sentence embeddings. In this paper, we introduce a simple representation of sentences in which a sentence embedding is represented as a weighted average of word vectors followed by a soft projection. We demonstrate the effectiveness of this proposed method on the clinical semantic textual similarity task of the BioCreative/OHNLP Challenge 2018.

Tianlin Liu, Lyle Ungar, João Sedoc

Word vectors are at the core of many natural language processing tasks. Recently, there has been interest in post-processing word vectors to enrich their semantic information. In this paper, we introduce a novel word vector post-processing technique based on matrix conceptors (Jaeger2014), a family of regularized identity maps. More concretely, we propose to use conceptors to suppress those latent features of word vectors having high variances. The proposed method is purely unsupervised: it does not rely on any corpus or external linguistic database. We evaluate the post-processed word vectors on a battery of intrinsic lexical evaluation tasks, showing that the proposed method consistently outperforms existing state-of-the-art alternatives. We also show that post-processed word vectors can be used for the downstream natural language processing task of dialogue state tracking, yielding improved results in different dialogue domains.

Sven Buechel, João Sedoc, H. Andrew Schwartz et al.

One of the major downsides of Deep Learning is its supposed need for vast amounts of training data. As such, these techniques appear ill-suited for NLP areas where annotated data is limited, such as less-resourced languages or emotion analysis, with its many nuanced and hard-to-acquire annotation formats. We conduct a questionnaire study indicating that indeed the vast majority of researchers in emotion analysis deems neural models inferior to traditional machine learning when training data is limited. In stark contrast to those survey results, we provide empirical evidence for English, Polish, and Portuguese that commonly used neural architectures can be trained on surprisingly few observations, outperforming $n$-gram based ridge regression on only 100 data points. Our analysis suggests that high-quality, pre-trained word embeddings are a main factor for achieving those results.

Sven Buechel, Anneke Buffone, Barry Slaff et al.

Computational detection and understanding of empathy is an important factor in advancing human-computer interaction. Yet to date, text-based empathy prediction has the following major limitations: It underestimates the psychological complexity of the phenomenon, adheres to a weak notion of ground truth where empathic states are ascribed by third parties, and lacks a shared corpus. In contrast, this contribution presents the first publicly available gold standard for empathy prediction. It is constructed using a novel annotation methodology which reliably captures empathy assessments by the writer of a statement using multi-item scales. This is also the first computational work distinguishing between multiple forms of empathy, empathic concern, and personal distress, as recognized throughout psychology. Finally, we present experimental results for three different predictive models, of which a CNN performs the best.

Sajal Choudhary, Prerna Srivastava, Lyle Ungar et al.

We investigate the task of building a domain aware chat system which generates intelligent responses in a conversation comprising of different domains. The domain, in this case, is the topic or theme of the conversation. To achieve this, we present DOM-Seq2Seq, a domain aware neural network model based on the novel technique of using domain-targeted sequence-to-sequence models (Sutskever et al., 2014) and a domain classifier. The model captures features from current utterance and domains of the previous utterances to facilitate the formation of relevant responses. We evaluate our model on automatic metrics and compare our performance with the Seq2Seq model.

Grishma Jena, Mansi Vashisht, Abheek Basu et al.

Human interactions and human-computer interactions are strongly influenced by style as well as content. Adding a persona to a chatbot makes it more human-like and contributes to a better and more engaging user experience. In this work, we propose a design for a chatbot that captures the "style" of Star Trek by incorporating references from the show along with peculiar tones of the fictional characters therein. Our Enterprise to Computer bot (E2Cbot) treats Star Trek dialog style and general dialog style differently, using two recurrent neural network Encoder-Decoder models. The Star Trek dialog style uses sequence to sequence (SEQ2SEQ) models (Sutskever et al., 2014; Bahdanau et al., 2014) trained on Star Trek dialogs. The general dialog style uses Word Graph to shift the response of the SEQ2SEQ model into the Star Trek domain. We evaluate the bot both in terms of perplexity and word overlap with Star Trek vocabulary and subjectively using human evaluators.

João Sedoc, Jean Gallier, Lyle Ungar et al.

Vector space representations of words capture many aspects of word similarity, but such methods tend to make vector spaces in which antonyms (as well as synonyms) are close to each other. We present a new signed spectral normalized graph cut algorithm, signed clustering, that overlays existing thesauri upon distributionally derived vector representations of words, so that antonym relationships between word pairs are represented by negative weights. Our signed clustering algorithm produces clusters of words which simultaneously capture distributional and synonym relations. We evaluate these clusters against the SimLex-999 dataset (Hill et al.,2014) of human judgments of word pair similarities, and also show the benefit of using our clusters to predict the sentiment of a given text.