Explainable Slot Type Attentions to Improve
Joint Intent Detection and Slot Filling

Intent classifier is a single layer multi-class classifier that uses BERT as the language encoder. For a given utterance $u$, the embedding vector for [CLS] token is the context embedding $u^{c}\in\mathbb{R}^{d}$. Intent classifier outputs intent logits $g^{intent}$ as follows.

where, $W^{intent}\in\mathbb{R}^{d\times|I|}$ is the learnable weight matrix and $b^{intent}$ is the bias. $Y^{intent}$ is the one hot encoded intent vector and $p^{intent}=softmax(g^{intent})$. Intent loss $\mathcal{L}^{intent}$ for utterance $u$ is computed using cross entropy loss as follows.

We discuss details of our auxiliary network that helps the model learn explainable weights and generate general yet slot type specific features. First, after applying dropout to the encoded utterance $u^{e}$, we concatenate intent logits $g^{intent}$ with $u^{e}$. Note that $g^{intent}\in\mathbb{R}^{|I|}$, hence we expand the logits vector by one dimension, copy the values along the expanded dimension $l$ times (utterance length) to get the expanded intent logits tensor $g^{int\_e}$ ($\in\mathbb{R}^{l\times|I|}$). Concatenation of intent logits with token embeddings and then applying self attention provides direct fusion of intent features with the slot classification network to learn any correlations between intents and slots. We compute the new feature $u^{sa}$ ($\in\mathbb{R}^{l\times d}$) as follows. $LL$, $LN$, and $SA$ represent linear layer, layer normalization, and self attention functions, respectively. $\oplus$ is the concatenation.

$\theta^{LN}$, $\theta^{LL}$, and $\theta^{SA}$ are the sets of model parameters. Self attention function $SA(.)$ is as follows.

Query $Q_{x}$, key $K_{x}$, and value $V_{x}$ are calculated from input $x$ using three different linear projections, $LL_{q}(x)$, $LL_{k}(x)$, and $LL_{v}(x)$, respectively. $d_{x}$ is the dimension of the input $x$ (also $K_{x}$). In Equation 3, $LL$ projects back to dimension $d$ and $d_{x}$ = $d$ in $SA$. Then we add the original utterance embedding $u^{e}$ to $u^{sa}$ and perform a layer normalization to get $\hat{u^{e}}$ before the auxiliary network.

We project the utterance embedding $\hat{u^{e}}$ into multiple representations; exactly $|T|$ times. We want our model to explain slots at a fine-grained level and hence, we need weight distributions over the entire input for each token, for each slot type. We compute multiple self attentions, one per slot type, that provides slot type specific attention weights per token with respect to the utterance.

However, without proper supervision, these attention weights are not meaningful. Therefore, we make these multiple projections correspond to $|T|$ slot type binary classifiers. We apply the self attention defined in Equation 4, $|T|$ number of times, one for each slot type by projecting different query $Q_{type\_i}$, key $K_{type\_i}$, and value $V_{type\_i}$ tensors ($\in\mathbb{R}^{l\times d_{h}}$) from $\hat{u^{e}}$, each using three different linear projections. Self attended output $h_{type\_i}$ ($\in\mathbb{R}^{l\times d_{h}}$) and attention weights $\alpha_{type\_i}$ ($\in\mathbb{R}^{l\times l}$) for slot type $type\_i$ ($\in T$) are as follows. $d_{h}$ is the embedding dimension.

Slot type projections require additional optimization objectives so that the self attention weights are truly meaningful and not random noise. We train each slot type projection to predict binary output that states whether an input token in the utterance is true (1) or false (0) for that slot type. This training data is automatically generated from sequence labeling BIO ground truth as shown in the example in Figure 3. Binary format for a slot type except for O is generated by having 1s to slot tokens (non-O positions) and 0s to otherwise. For O, all non-slot tokens are marked as 1s and rest are 0s.

With the additional optimization objectives for each slot type, self attention weights can be used to visualize attention points of the input for each token with respect to each slot type as explanations. This is because: (i) per slot type, the embedding for the token being classified is computed attending to all the input tokens including itself and (ii) the same token embedding is used for the particular slot type classification (controlled supervision), and further, (iii) type classification output logits are fed to the final slot classification as additional features (explicit connection to final classification). Our binary classification output logits enable the model to learn general patterns/features specific to each slot type that also boost model performance. Binary classification for $i$th slot type $type\_i$ ($\in T$) is performed as follows. We initialize a set of weight tensors $W^{H}$ and bias $b^{H}$ where $|W^{H}|$ = $|b^{H}|$ = $|T|$.

where $h_{type\_i}\in\mathbb{R}^{l\times d_{h}}$, $W^{H}_{type\_i}\in\mathbb{R}^{d_{h}\times 1}$, $b^{H}_{type\_i}\in\mathbb{R}$, and slot type logits $g^{slot\_type}_{type\_i}\in\mathbb{R}^{l\times 1}$. Binary cross entropy loss $\mathcal{L}^{type}$ for type classification for utterance $u$ is as follows.

where, $Y^{t}$ is the one-hot encoded ground truth vector and $p^{t}$ is the predicted probabilities vector, and $N$ is the total number of data points. Note that we measure binary cross entropy loss collective for all the slot type classifiers. Therefore, if $|u|=l$, then $N=l\times|T|$.

We take the attention weights matrix $\alpha_{type\_i}$ for a slot type $type\_i$ and visualize attentions on the utterance for each token. See Figure 4 for examples.

Slot type logits generated from the previous step are new features that go into our slot classifier to improve accuracy. They are slot type specific and hence, can capture fine-grained slot patterns. We concatenate all the slot type logits per each input utterance token and then apply cross attention on it by having query $Q_{u^{e}}$ projected from the utterance $u^{e}$ and key $K_{g^{p}}$ and value $V_{g^{p}}$ projected from the concatenated and projected slot type logits $g^{p}$. The logits concatenation is performed at the token level where all the slot type logits for a token are concatenated to produce the tensor ${g}^{slot\_type\_c}(\in\mathbb{R}^{l\times|T|}$). Then the concatenated logits tensor is projected to get $g^{p}$ ($\in\mathbb{R}^{l\times d}$). We then apply cross attention between utterance embedding $u^{e}$ and $g^{p}$ to get slot type aware utterance representation $u^{cross}$ ($\in\mathbb{R}^{l\times d}$). We compute query, key, and value tensor projections ($\in\mathbb{R}^{l\times d}$) as follows. $\theta^{1}$, $\theta^{2}$, and $\theta^{3}$ are layer parameters for the three different linear projection layers.

$Q_{u^{e}}=LL(u^{e};\theta^{1}),K_{g^{p}}=LL(g^{p};\theta^{2}),V_{g^{p}}=LL(g^{p};\theta^{3})$

Cross attention between the utterance embeddings and slot type logits highlights slot type specific features from the utterance embeddings. This is added to utterance as follows to make the slot classifier input $u^{slot}$ ($\in\mathbb{R}^{l\times d}$).

Finally, the slot logits tensor $g^{slot}$ ($\in\mathbb{R}^{l\times|S|}$) is computed where $W^{slot}\in\mathbb{R}^{d\times|S|}$.

The slot loss $\mathcal{L}^{slot}$ for utterance $u$ is computed using the cross entropy loss. $Y_{i,s}^{slot}$ is the one hot encoded ground truth slot label and $p_{i,s}^{slot}$ is the softmax probability of slot BIO label $s$ for $i^{th}$ token.

We optimize our network by minimizing the joint loss of the three sub-networks. Overall loss $\mathcal{L}$ for utterance $u$ is defined as in Equation 13. $\alpha$, $\beta$, and $\gamma$ are hyperparameters that represent loss weights.

It is important to verify that the observations shown in Figure 4 are consistent. Our hypothesis is that positive slot type attention weights (Figure 4 (a) and (b)) have clear spikes or specific focus areas in attention compared to the negative slot type attention weights (Figure 4 (c)). We use entropy⁴⁴4https://en.wikipedia.org/wiki/Entropy_(information_theory) to capture this difference. We expect positive slot type attention weights to have lower entropy compared to the negative slot type attention weights, which tend to have a more uniform distribution, hence, higher entropy. We compute entropy for a list of attention weights [$x_{1}$, $x_{2}$,.., $x_{n}$] using the equation below, where, $Px_{i}$ = $x_{i}$/$\sum x_{i}$.

We compute the average entropy of the positive slot types and the negative slot types separately for each utterance and then average the entropy values. As shown in Table 4, the positive slot types achieve lower average entropy values compared to the negative slot types for both the datasets. The table also shows a breakdown of top $k\%$ attention weights (sorted in descending order and take the first $k\%$ items). We can see that when $k$ is small, like 5% or 10%, the entropy difference between positive and negative slot types is higher. The attention vectors are long, and hence several attention spikes for positive slot types may not seem prominent in the entropy computation when the entirety of the attention weights is considered. These results suggest that there are more attention spikes in positive slot types for utterances compared to negative slot types. Thus, we can use the distinguishable attention spikes to gain insights into slot filling decisions.

Our goal is to inspect whether the model attentions are consistent, trustworthy, and not arbitrary. We performed a small-scale human evaluation to check whether slot type attentions remain similar if the utterances are slightly modified. We randomly selected 20 utterances each, from ATIS and SNIPS. Three types of modifications were applied to each of these utterances: (i) modifying slot values only, (ii) utterance text only, and (iii) both utterance text and slot values. We obtained a total of 6 modifications for each utterance (2 per modification type). We make sure that utterance semantics (original slots types) and the meaning of the utterances do not change due to these modification. Figure 7 shows few examples of such modifications. We collected 120 modifications (20 x 6) for each dataset and 240 (120 x 2) in total for both datasets.

For each original utterance, a positive slot type was selected at random and the corresponding computed attention patterns were compared with the original utterance for each of the modifications. We used a pool of two judges and they were asked to rate the similarity of the attention patterns on a 3-point Likert scale (3 for exactly same, 2 for similarity with some overlap, 1 for no overlap). We obtained 344/360 for ATIS and 354/360 for SNIPS (the total possible score per dataset is 360=120x3). These scores indicate that the attentions for slot types were consistent across the modifications and they are not arbitrary or opaque. This shows that our slot type specific attention mechanism is able to successfully learn patterns that are specific to slot types ‘explicitly’. These patterns are also consistent, and hence, these slot type specific attention weights can provide a means for trustworthy explanations. See Appendix A for few more examples.