How to evaluate the coding abilities of Large Language Models (LLMs) remains an open question. We find that existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of LLMs.To address the knowledge gap, we propose a new benchmark named DevEval, which has three advances. (1) DevEval aligns with real-world repositories in multiple dimensions, e.g., code and dependency distributions. (2) DevEval is annotated by 13 developers and contains comprehensive annotations (e.g., requirements, original repositories, reference code, and reference dependencies). (3) DevEval comprises 1,825 testing samples from 115 repositories, covering 10 popular domains (e.g., Internet, Database). Based on DevEval, we propose repository-level code generation and evaluate 8 popular LLMs on DevEval (e.g., gpt-4, gpt-3.5, StarCoder 2, DeepSeek Coder, CodeLLaMa). Our experiments reveal these LLMs’ coding abilities in real-world code repositories. For example, the highest Pass@1 of gpt-4 only is 53.04% in our experiments. We also analyze LLMs’ failed cases and summarize their shortcomings. We hope DevEval can facilitate the development of LLMs in real code repositories. DevEval, prompts, and LLMs’ predictions have been released.

Humans learn social skills through both imitation and social interaction. This social learning process is largely understudied by existing research on building language agents. Motivated by this gap, we propose an interactive learning method, SOTOPIA-π, that improves the social intelligence of language agents. This method leverages behavior cloning and self-reinforcement based training on filtered social interaction data according to large language model (LLM) rating. We show that our training method allows a 7B LLM to reach the social goal completion ability of an expert model (GPT-4-based agent) without the loss of more generic abilities, such as the ability to answer knowledge-based questions. We also demonstrate that this training paradigm uncovers some weaknesses in standard evaluation and safety training paradigms that (1) LLM-based evaluation of social intelligence overestimates the abilities of the language agents trained specifically for social interaction, and that (2) despite not training for better safety or question answering (QA) ability, our methods improve the safety of language agents and maintain general QA ability on the MMLU benchmark.

Warning: This paper contains content that may be offensive or upsetting. Understanding the harms and offensiveness of statements requires reasoning about the social and situational context in which statements are made. For example, the utterance “your English is very good” may implicitly signal an insult when uttered by a white man to a non-white colleague, but uttered by an ESL teacher to their student would be interpreted as a genuine compliment. Such contextual factors have been largely ignored by previous approaches to toxic language detection. We introduce COBRA frames, the first context-aware formalism for explaining the intents, reactions, and harms of offensive or biased statements grounded in their social and situational context. We create COBRACORPUS, a dataset of 33k potentially offensive statements paired with machine-generated contexts and free-text explanations of offensiveness, implied biases, speaker intents, and listener reactions. To study the contextual dynamics of offensiveness, we train models to generate COBRA explanations, with and without access to the context. We find that explanations by context-agnostic models are significantly worse than by context-aware ones, especially in situations where the context inverts the statement’s offensiveness (29% accuracy drop). Our work highlights the importance and feasibility of contextualized NLP by modeling social factors.

Large language models (LLMs) struggle on processing complicated observations in interactive decision making. To alleviate this issue, we propose a simple hierarchical prompting approach. Diverging from previous prompting approaches that always put the full observation (a web page) to the prompt, we propose to first construct an action-aware observation which is more condensed and relevant with a dedicated Summarizer prompt. The Actor prompt then predicts the next action based on the summarized history. While our method has broad applicability, we particularly demonstrate its efficacy in the complex domain of web navigation where a full observation often contains redundant and irrelevant information. Our approach outperforms the previous state-of-the-art prompting mechanism with the same LLM by 6.2% on task success rate, demonstrating its potential on interactive decision making tasks with long observation traces.

Embodied dialogue instruction following requires an agent to complete a complex sequence of tasks from a natural language exchange. The recent introduction of benchmarks raises the question of how best to train and evaluate models for this multi-turn, multi-agent, long-horizon task. This paper contributes to that conversation, by arguing that imitation learning (IL) and related low-level metrics are actually misleading and do not align with the goals of embodied dialogue research and may hinder progress.We provide empirical comparisons of metrics, analysis of three models, and make suggestions for how the field might best progress. First, we observe that models trained with IL take spurious actions during evaluation. Second, we find that existing models fail to ground query utterances, which are essential for task completion. Third, we argue evaluation should focus on higher-level semantic goals. We will release code to additionally filter the data and benchmark models for improved evaluation.

Most previous work on grammar induction focuses on learning phrasal or dependency structure purely from text. However, because the signal provided by text alone is limited, recently introduced visually grounded syntax models make use of multimodal information leading to improved performance in constituency grammar induction. However, as compared to dependency grammars, constituency grammars do not provide a straightforward way to incorporate visual information without enforcing language-specific heuristics. In this paper, we propose an unsupervised grammar induction model that leverages word concreteness and a structural vision-based heuristic to jointly learn constituency-structure and dependency-structure grammars. Our experiments find that concreteness is a strong indicator for learning dependency grammars, improving the direct attachment score (DAS) by over 50% as compared to state-of-the-art models trained on pure text. Next, we propose an extension of our model that leverages both word concreteness and visual semantic role labels in constituency and dependency parsing. Our experiments show that the proposed extension outperforms the current state-of-the-art visually grounded models in constituency parsing even with a smaller grammar size.

In this paper we demonstrate that context free grammar (CFG) based methods for grammar induction benefit from modeling lexical dependencies. This contrasts to the most popular current methods for grammar induction, which focus on discovering either constituents or dependencies. Previous approaches to marry these two disparate syntactic formalisms (e.g., lexicalized PCFGs) have been plagued by sparsity, making them unsuitable for unsupervised grammar induction. However, in this work, we present novel neural models of lexicalized PCFGs that allow us to overcome sparsity problems and effectively induce both constituents and dependencies within a single model. Experiments demonstrate that this unified framework results in stronger results on both representations than achieved when modeling either formalism alone.1

In this paper, we propose a novel graph neural network with generated parameters (GP-GNNs). The parameters in the propagation module, i.e. the transition matrices used in message passing procedure, are produced by a generator taking natural language sentences as inputs. We verify GP-GNNs in relation extraction from text, both on bag- and instance-settings. Experimental results on a human-annotated dataset and two distantly supervised datasets show that multi-hop reasoning mechanism yields significant improvements. We also perform a qualitative analysis to demonstrate that our model could discover more accurate relations by multi-hop relational reasoning.

We introduce a conceptually simple and effective method to quantify the similarity between relations in knowledge bases. Specifically, our approach is based on the divergence between the conditional probability distributions over entity pairs. In this paper, these distributions are parameterized by a very simple neural network. Although computing the exact similarity is in-tractable, we provide a sampling-based method to get a good approximation. We empirically show the outputs of our approach significantly correlate with human judgments. By applying our method to various tasks, we also find that (1) our approach could effectively detect redundant relations extracted by open information extraction (Open IE) models, that (2) even the most competitive models for relational classification still make mistakes among very similar relations, and that (3) our approach could be incorporated into negative sampling and softmax classification to alleviate these mistakes.

We introduce neural finite state transducers (NFSTs), a family of string transduction models defining joint and conditional probability distributions over pairs of strings. The probability of a string pair is obtained by marginalizing over all its accepting paths in a finite state transducer. In contrast to ordinary weighted FSTs, however, each path is scored using an arbitrary function such as a recurrent neural network, which breaks the usual conditional independence assumption (Markov property). NFSTs are more powerful than previous finite-state models with neural features (Rastogi et al., 2016.) We present training and inference algorithms for locally and globally normalized variants of NFSTs. In experiments on different transduction tasks, they compete favorably against seq2seq models while offering interpretable paths that correspond to hard monotonic alignments.

Learning to hash via generative model has become a powerful paradigm for fast similarity search in documents retrieval. To get binary representation (i.e., hash codes), the discrete distribution prior (i.e., Bernoulli Distribution) is applied to train the variational autoencoder (VAE). However, the discrete stochastic layer is usually incompatible with the backpropagation in the training stage, and thus causes a gradient flow problem because of non-differentiable operators. The reparameterization trick of sampling from a discrete distribution usually inc non-differentiable operators. In this paper, we propose a method, Doc2hash, that solves the gradient flow problem of the discrete stochastic layer by using continuous relaxation on priors, and trains the generative model in an end-to-end manner to generate hash codes. In qualitative and quantitative experiments, we show the proposed model outperforms other state-of-art methods.

We present FewRel 2.0, a more challenging task to investigate two aspects of few-shot relation classification models: (1) Can they adapt to a new domain with only a handful of instances? (2) Can they detect none-of-the-above (NOTA) relations? To construct FewRel 2.0, we build upon the FewRel dataset by adding a new test set in a quite different domain, and a NOTA relation choice. With the new dataset and extensive experimental analysis, we found (1) that the state-of-the-art few-shot relation classification models struggle on these two aspects, and (2) that the commonly-used techniques for domain adaptation and NOTA detection still cannot handle the two challenges well. Our research calls for more attention and further efforts to these two real-world issues. All details and resources about the dataset and baselines are released at https://github.com/thunlp/fewrel.

Sememes are minimum semantic units of concepts in human languages, such that each word sense is composed of one or multiple sememes. Words are usually manually annotated with their sememes by linguists, and form linguistic common-sense knowledge bases widely used in various NLP tasks. Recently, the lexical sememe prediction task has been introduced. It consists of automatically recommending sememes for words, which is expected to improve annotation efficiency and consistency. However, existing methods of lexical sememe prediction typically rely on the external context of words to represent the meaning, which usually fails to deal with low-frequency and out-of-vocabulary words. To address this issue for Chinese, we propose a novel framework to take advantage of both internal character information and external context information of words. We experiment on HowNet, a Chinese sememe knowledge base, and demonstrate that our framework outperforms state-of-the-art baselines by a large margin, and maintains a robust performance even for low-frequency words.

Sememes are defined as the minimum semantic units of human languages. As important knowledge sources, sememe-based linguistic knowledge bases have been widely used in many NLP tasks. However, most languages still do not have sememe-based linguistic knowledge bases. Thus we present a task of cross-lingual lexical sememe prediction, aiming to automatically predict sememes for words in other languages. We propose a novel framework to model correlations between sememes and multi-lingual words in low-dimensional semantic space for sememe prediction. Experimental results on real-world datasets show that our proposed model achieves consistent and significant improvements as compared to baseline methods in cross-lingual sememe prediction. The codes and data of this paper are available at https://github.com/thunlp/CL-SP.

Entity typing aims to classify semantic types of an entity mention in a specific context. Most existing models obtain training data using distant supervision, and inevitably suffer from the problem of noisy labels. To address this issue, we propose entity typing with language model enhancement. It utilizes a language model to measure the compatibility between context sentences and labels, and thereby automatically focuses more on context-dependent labels. Experiments on benchmark datasets demonstrate that our method is capable of enhancing the entity typing model with information from the language model, and significantly outperforms the state-of-the-art baseline. Code and data for this paper can be found from https://github.com/thunlp/LME.

Most language modeling methods rely on large-scale data to statistically learn the sequential patterns of words. In this paper, we argue that words are atomic language units but not necessarily atomic semantic units. Inspired by HowNet, we use sememes, the minimum semantic units in human languages, to represent the implicit semantics behind words for language modeling, named Sememe-Driven Language Model (SDLM). More specifically, to predict the next word, SDLM first estimates the sememe distribution given textual context. Afterwards, it regards each sememe as a distinct semantic expert, and these experts jointly identify the most probable senses and the corresponding word. In this way, SDLM enables language models to work beyond word-level manipulation to fine-grained sememe-level semantics, and offers us more powerful tools to fine-tune language models and improve the interpretability as well as the robustness of language models. Experiments on language modeling and the downstream application of headline generation demonstrate the significant effectiveness of SDLM.

We present a Few-Shot Relation Classification Dataset (dataset), consisting of 70, 000 sentences on 100 relations derived from Wikipedia and annotated by crowdworkers. The relation of each sentence is first recognized by distant supervision methods, and then filtered by crowdworkers. We adapt the most recent state-of-the-art few-shot learning methods for relation classification and conduct thorough evaluation of these methods. Empirical results show that even the most competitive few-shot learning models struggle on this task, especially as compared with humans. We also show that a range of different reasoning skills are needed to solve our task. These results indicate that few-shot relation classification remains an open problem and still requires further research. Our detailed analysis points multiple directions for future research.