Source-linked AI summary
Neural Responding Machine for Short-Text Conversation
Lifeng Shang, Zhengdong Lu, Hang Li
TL;DR
Short-Text Conversation requires generating suitable responses even when responses are not semantically equivalent to posts, unlike translation targets. The paper proposes Neural Responding Machine, an encoder-decoder neural generator with static and dynamically changing post representations. Its empirical studies find that NRM, especially the hybrid encoding scheme, outperforms retrieval-based and SMT-based methods.
Problem
Short-Text Conversation is difficult because responses are not necessarily semantically equivalent to posts, making translation-based generation unsuitable.
Method
NRM uses a neural encoder-decoder to represent a post and decode responses, with attention-based local encoding that can dynamically focus on input words.
Results
NRM, especially the hybrid encoding scheme, outperforms state-of-the-art retrieval-based and SMT-based methods in empirical studies.
Takeaways & Limitations
NRM can generate multiple responses with great variety to a given post.
Takeaways & Limitations
Retrieval-based methods use pre-existing responses that are difficult to customize, while SMT is unsuitable when posts and responses are not semantically equivalent.
Abstract
from arXiv · showhide
We propose Neural Responding Machine (NRM), a neural network-based response generator for Short-Text Conversation. NRM takes the general encoder-decoder framework: it formalizes the generation of response as a decoding process based on the latent representation of the input text, while both encoding and decoding are realized with recurrent neural networks (RNN). The NRM is trained with a large amount of one-round conversation data collected from a microblogging service. Empirical study shows that NRM can generate grammatically correct and content-wise appropriate responses to over 75% of the input text, outperforming state-of-the-arts in the same setting, including retrieval-based and SMT-based models.
1 Introduction
Short-Text Conversation is difficult because responses need not translate or match posts semantically. NRM addresses this with neural encoder-decoder response generation trained on conversation data.
- Motivation: Open-domain conversation is difficult because existing rule-based and learning-based methods require manual rules or limited task-specific data.These requirements make extensible open-domain conversation systems difficult to develop.
- Task setting: Short-Text Conversation uses one-round post-response pairs collected from rapidly growing microblogging services.The former short text is the input post, and the latter is the response.
- Prior methods: Retrieval-based methods select pre-existing responses using matching features, limiting customization and sometimes failing to distinguish positive from negative responses.Feature engineering alone may not adequately handle mismatched named entities or other semantic differences.
- Prior methods: SMT-based generation is intrinsically unsuitable because responses are not semantically equivalent to posts as they are in translation.A single post can receive responses with completely different content.
- NRM approach: NRM uses an encoder-decoder framework that summarizes a post into a representation and decodes that representation into a response.The model estimates response likelihood given a post and can allow the representation to change dynamically during generation.
- Contribution: The authors propose NRM as a neural response generator and report performance better than traditional retrieval-based and translation-based methods when trained with reasonable data.The supplied introduction states this as an empirical contribution rather than specifying a numerical result.
2 The Dataset for STC
The study builds an STC dataset from Sina Weibo, where short post-response pairs provide a large conversational resource. The corpus is cleaned and contains multiple responses per post.
- Dataset: The models are trained on roughly 4.4 million Weibo conversation pairs.The dataset and its machine-translated English version were planned for release.
- Dataset: Weibo is a Twitter-like Chinese microblogging service whose 140-character limits make post-response pairs suitable surrogates for short-text conversation.Posts are public or shared with followers, and other users comment with responses.
- Dataset statistics: Table 1 reports dataset statistics and identifies labeled data for retrieval ranking and fine-tuning data for SMT.The caption describes these auxiliary subsets by their evaluation use.
- Dataset construction: The corpus is constructed by crawling hundreds of millions of pairs, removing trivial responses and advertisements, and retaining the first 30 responses for topic consistency.Each post has 20 different responses on average.
3 Neural Responding Machines for STC
NRM uses recurrent encoder-decoder networks to generate responses from post representations, with attention allowing the context to change during decoding. The paper compares global, local, and hybrid encoding schemes for STC.
- General Framework: NRM encodes the input sequence into hidden representations and uses a context generator plus decoder RNN to produce each response word.The context is linearly transformed by matrix L before stimulating the generating RNN.
- General Framework: Attention determines which parts of the encoded post receive emphasis during response generation and may remain fixed or change dynamically.In dynamic settings, attention can depend on previously generated words, the input sequence, or their latent representations.
- Decoder: The decoder is an RNN language model conditioned on context, with its hidden state computed from the previous response word, previous state, and current context.The implementation uses a nonlinear recurrent function; the paper uses GRU units for sequence modeling.
- Encoding Schemes: The global encoder uses the final hidden state h_T as the sentence representation, but this summary may lose important post details.The paper notes that representing an entire post in one vector can be difficult, particularly when the hidden-state dimension is limited.
- Encoding Schemes: The local encoder forms context as an attention-weighted combination of hidden states, adaptively focusing on important input words according to generated response words.The weighting factors model alignment between input positions and output positions.
- Encoding Schemes: The hybrid encoder combines global and local representations so global information can provide context for local matching.NRM-hyb concatenates encoded hidden states and combines global and local information during response generation; it is initialized from separately trained NRM-loc and NRM-glo models before fine-tuning.
4 Experiments
The experiments compare three NRM variants with retrieval-based and SMT-based competitors using large-scale Weibo conversation data and established competitor implementations.
- Compared methods: Three NRM settings—NRM-glo, NRM-loc, and NRM-hyb—are evaluated against retrieval-based and SMT-based methods.The evaluation compares the three NRM configurations described in Section 3 with both competitor categories.
- Compared methods: The retrieval-based competitor ranks candidate responses from a 4.4 million-pair repository using 14 matching features.The features range from cosine similarity to deep matching models and are combined through a linear model.
- Compared methods: The retrieval model is trained with ranking SVM on 6,017 manually labeled post-response pairs from 225 posts.A baseline retriever supplies about 30 responses per post, and only the top response is evaluated against NRM.
- Compared methods: The SMT competitor uses post-response pairs as parallel data for training the open-source Moses phrase-based translation model.An additional 3,000 post-response pairs are used to tune the system.
5 Results and Analysis
Human evaluation finds that NRM variants generally produce fluent and relevant responses, with the hybrid model achieving the strongest statistically significant performance and diverse outputs.
- Human evaluation: 74.4% of SMT-generated responses were labeled unsuitable, mainly because of fluency and relevance errors.The result supports the authors’ concern that STC post-response pairs are not a straightforward parallel corpus for SMT.
- Human evaluation: More than 60% of responses from all three NRM variants were labeled suitable or neutral.The authors interpret this as evidence that most generated responses were fluent and semantically relevant to the post.
- Model comparison: NRM-hyb outperformed NRM-loc and NRM-glo, indicating that global post representation complements dynamically generated local context.The hybrid model combines the two representation strategies evaluated in the NRM variants.
- Statistical significance: NRM-hyb significantly outperformed all other methods under the Friedman test, with p < 0.05.Retrieval-based and NRM-glo performance was comparable, while SMT was significantly worse than retrieval-based and NRM-hyb methods.
- Case study: NRM variants generated distinct suitable responses, whereas retrieval sometimes returned responses with mismatched entity names.The case study attributes the diversity partly to architectural and random initialization differences and observes that NRMs tend to avoid specific details.
- Case study: Five beam-search responses for one post were fluent, relevant, and substantially different from one another.The example uses beam size 500 and keeps the highest-likelihood response for each first word.
6 Conclusions and Future Work
The paper presents NRM as an encoder-decoder response generator for STC, reports stronger performance for its hybrid scheme, and identifies controllable user intention or sentiment as future work.
- Conclusions: The paper proposes Neural Responding Machine, an encoder-decoder neural network system for generating responses to posts.The paper reports empirical support for NRMs, especially the hybrid encoding scheme, against retrieval-based and SMT-based methods.
- Future work: NRM can generate multiple responses with substantial variety for a given post.Future work would add user intention or sentiment as an external decoder signal for responses with specific goals.